ARTIFICIAL INTELLIGENCE WITH GREAT POWER COMES GREAT RESPONSIBILITY JOINT HEARING BEFORE THE SUBCOMMITTEE ON RESEARCH AND TECHNOLOGY SUBCOMMITTEE ON ENERGY COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY HOUSE OF REPRESENTATIVES ONE HUNDRED FIFTEENTH CONGRESS SECOND SESSION JUNE 26 2018 Serial No 115–67 Printed for the use of the Committee on Science Space and Technology Available via the World Wide Web http science house gov U S GOVERNMENT PUBLISHING OFFICE 30–877PDF WASHINGTON 2018 COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY HON LAMAR S SMITH Texas Chair FRANK D LUCAS Oklahoma EDDIE BERNICE JOHNSON Texas DANA ROHRABACHER California ZOE LOFGREN California MO BROOKS Alabama DANIEL LIPINSKI Illinois RANDY HULTGREN Illinois SUZANNE BONAMICI Oregon BILL POSEY Florida AMI BERA California THOMAS MASSIE Kentucky ELIZABETH H ESTY Connecticut RANDY K WEBER Texas MARC A VEASEY Texas STEPHEN KNIGHT California DONALD S BEYER JR Virginia BRIAN BABIN Texas JACKY ROSEN Nevada BARBARA COMSTOCK Virginia CONOR LAMB Pennsylvania BARRY LOUDERMILK Georgia JERRY MCNERNEY California ED PERLMUTTER Colorado RALPH LEE ABRAHAM Louisiana PAUL TONKO New York GARY PALMER Alabama BILL FOSTER Illinois DANIEL WEBSTER Florida MARK TAKANO California ANDY BIGGS Arizona COLLEEN HANABUSA Hawaii ROGER W MARSHALL Kansas CHARLIE CRIST Florida NEAL P DUNN Florida CLAY HIGGINS Louisiana RALPH NORMAN South Carolina DEBBIE LESKO Arizona SUBCOMMITTEE ON RESEARCH AND TECHNOLOGY HON BARBARA COMSTOCK Virginia Chair FRANK D LUCAS Oklahoma DANIEL LIPINSKI Illinois RANDY HULTGREN Illinois ELIZABETH H ESTY Connecticut STEPHEN KNIGHT California JACKY ROSEN Nevada BARRY LOUDERMILK Georgia SUZANNE BONAMICI Oregon DANIEL WEBSTER Florida AMI BERA California ROGER W MARSHALL Kansas DONALD S BEYER JR Virginia DEBBIE LESKO Arizona EDDIE BERNICE JOHNSON Texas LAMAR S SMITH Texas SUBCOMMITTEE ON ENERGY HON RANDY K WEBER Texas Chair DANA ROHRABACHER California MARC A VEASEY Texas Ranking Member FRANK D LUCAS Oklahoma ZOE LOFGREN California MO BROOKS Alabama DANIEL LIPINSKI Illinois RANDY HULTGREN Illinois JACKY ROSEN Nevada THOMAS MASSIE Kentucky JERRY MCNERNEY California PAUL TONKO New York STEPHEN KNIGHT California BILL FOSTER Illinois GARY PALMER Alabama MARK TAKANO California DANIEL WEBSTER Florida EDDIE BERNICE JOHNSON Texas NEAL P DUNN Florida RALPH NORMAN South Carolina LAMAR S SMITH Texas II CONTENTS June 26 2018 Page Witness List Hearing Charter 2 3 Opening Statements Statement by Representative Barbara Comstock Chairwoman Subcommittee on Research and Technology Committee on Science Space and Technology U S House of Representatives Written Statement Statement by Representative Daniel Lipinski Ranking Member Subcommittee on Research and Technology Committee on Science Space and Technology U S House of Representatives Written Statement Statement by Representative Lamar Smith Chairman Committee on Science Space and Technology U S House of Representatives Written Statement Statement by Representative Marc A Veasey Ranking Member Subcommittee on Energy Committee on Science Space and Technology U S House of Representatives Written Statement Statement by Representative Randy K Weber Chairman Subcommittee on Energy Committee on Science Space and Technology U S House of Representatives Written Statement Written statement by Representative Eddie Bernice Johnson Ranking Member Committee on Science Space and Technology U S House of Representatives 4 6 8 10 12 13 14 15 16 18 21 Witnesses Dr Tim Persons Chief Scientist U S Government Accountability Office Oral Statement Written Statement Mr Greg Brockman Co-Founder and Chief Technology Officer OpenAI Oral Statement Written Statement Dr Fei-Fei Li Chairperson of the Board and Co-Founder AI4ALL Oral Statement Written Statement Discussion 22 25 40 42 50 52 59 Appendix I Answers to Post-Hearing Questions Dr Jaime Carbonell Director Language Technologies Institute and Allen Newell Professor School of Computer Science Carnegie Mellon University Dr Tim Persons Chief Scientist U S Government Accountability Office Mr Greg Brockman Co-Founder and Chief Technology Officer OpenAI Dr Fei-Fei Li Chairperson of the Board and Co-Founder AI4ALL III 82 89 97 105 IV Page Appendix II Additional Material for the Record Dr Jaime Carbonell Director Language Technologies Institute and Allen Newell Professor School of Computer Science Carnegie Mellon University written statement Document submitted by Representative Bill Foster Subcommittee on Research and Technology Committee on Science Space and Technology U S House of Representatives Document submitted by Representative Neal P Dunn Subcommittee on Energy Committee on Science Space and Technology U S House of Representatives 112 123 150 ARTIFICIAL INTELLIGENCE WITH GREAT POWER COMES GREAT RESPONSIBILITY TUESDAY JUNE 26 2018 SUBCOMMITTEE HOUSE OF REPRESENTATIVES RESEARCH AND TECHNOLOGY AND SUBCOMMITTEE ON ENERGY ON SCIENCE SPACE AND TECHNOLOGY Washington D C ON COMMITTEE The Subcommittees met pursuant to call at 10 37 a m in Room 2318 of the Rayburn House Office Building Hon Barbara Comstock Chairwoman of the Subcommittee on Research and Technology presiding 1 2 EDDIE 8ERN CE JOHNSON Texas RANKING MEM8ER I AMAA S SMITH Texas CHAIRMAN ongrrss of the ilnited tares iiousc of Rcprcscntatiucs COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY 2321 RAYBURN HOUSE OFFICE BUILDING WASHINGTON DC 20515-6301 202 225-6371 www s iencehousl' gov Artificial Intelligence - With Great Power Comes Great Responsibility Tuesday June 26 2018 0 30a m 2318 Rayburn House Office Building Witnesses Dr Jaime Carbonell Director Language Technologies Institute and Allen Newell Professor School of Computer Science Carnegie Mellon University Dr Tim Persons Chief Scientist U S Government Accountability Office Mr Greg Brockman Co-Founder and Chief Technology Officer OpenAI Dr Fei-Fei Li Chairperson of the Board and Co-Founder AI4ALL 3 U S HOUSE OF REPRESENTATIVES COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY HEARING CHARTER June 26 2018 TO Members Subconnnittee on Research and Technology Subconnnittee on Energy FROM Majority Staff Connnittee on Science Space and Technology SUBJECT Joint Subcommittee Hearing Artificial Intelligence- With Great Power Comes Great Responsibility The Subconnnittees on Research and Technology and Energy will hold a hearing titled Artificial Intelligence- With Great Power Comes Great Responsibility on Tuesday June 26 2018 at 10 30 a m in Room 2318 of the Rayburn House Office Building Hearing Purpose The purpose of the hearing is to understand the state of artificial intelligence technology and the difference between narrow and general intelligence The hearing will also examine the types of research being conducted to advance artificial general intelligence technology and explore its game-changing potential and implications Witness List • • • • Dr Jaime Carbonell Director Language Technologies Institute and Allen Newell Professor School of Computer Science Carnegie Mellon University Dr Tim Persons Chief Scientist U S Government Accountability Office Mr Greg Brockman Co-Founder and Chief Technology Officer OpenAI Dr Fei-Fei Li Chairperson of the Board and Co-Founder AI4ALL Staff Contact For questions related to the hearing please contact Raj Bharwani of the Majority Staff at 202-225-6371 4 Chairwoman COMSTOCK The Committee on Space Science and Technology will come to order Without objection the Chair is authorized to declare recesses of the Committee at any time Good morning and welcome to today’s hearing entitled ‘‘Artificial Intelligence—With Great Power Comes Great Responsibility ’’ I now recognize myself for five minutes for an opening statement First I would like to note that one of our witnesses Dr Jaime Carbonell from Carnegie Mellon University is unable to be here today due to a medical emergency We wish him well and a speedy recovery and without objection we’ll ensure his written testimony is made part of the hearing record The prepared statement of Mr Carbonell appears in Appendix II Chairwoman COMSTOCK One of the reasons I’ve been looking forward to today’s hearing is to get a better sense from our witnesses about the nuances of the term artificial intelligence and implications for our society in a future where AI is ubiquitous Of course one might say AI is already pervasive Since the term was first coined in the 1950s we have made huge advances in the field of artificial narrow intelligence which has been applied to many familiar everyday items such as the technology underlying Siri and Alexa Called ANI for short such systems are designed to conduct specific and usually limited tasks For example a machine that excels at playing poker wouldn’t be able to parallel park a car Conversely AGI or artificial general intelligence refers to intelligent behavior across a range of cognitive tasks If you enjoy science fiction movies this definition may conjure up scenes from any number of classics such as Blade Runner The Matrix or The Terminator For many individuals the term AGI invokes images of robots or machines with human intelligence As it turns out we are decades away from realizing such AGI systems Nevertheless discussions about AGI and a future in which AGI is commonplace lead to some interesting questions worthy of analysis For example Elon Musk has been quoted as saying that AI quote ‘‘is a fundamental risk to the existence of human civilization’’ and poses ‘‘vastly more risk’’ than North Korea Does that mean that AGI may evolve to a point one day when we will lose control over machines of our own creation As farfetched as that sounds minds and scientists are certainly discussing such questions For the short term however my constituents are concerned about less existential issues that usually accompany new and evolving technologies topics such as cybersecurity protecting our privacy and impacts to our nation’s economy and to jobs I am an original cosponsor of a bill introduced earlier this year titled the AI JOBS Act of 2018 to help our workplace prepare for the ways AI will shape the economy of the future I will also introduce legislation today to reauthorize the National Institute of Standards and Technology which includes language directing NIST to support development of artificial intelligence and data science 5 There is immense potential for AGI to help humans and to help our economy and all of the issues we’re dealing with today but that potential is also accompanied by some of the concerns that we will discuss today I look forward to what our panel has to share with us about the bright as well as the challenging sides of the future with AGI The prepared statement of Chairwoman Comstock follows 6 s For Immediate Release June 26 2018 HNOLOGY Media Contact Heather Vaughan Bridget Dunn 202 225-6371 Statement by Chairwoman Barbara Comstock R-Va Artificial Intelligence- With Great Power Comes Great Responsibility Chairwoman Comstock I would like to say first that one of our witnesses Dr Jaime Carbonell from Carnegie Mellon University is unable to be here today due to a medical emergency We wish him well and a speedy recovery and without objection will ensure his written testimony is made part of the hearing record One of the reasons I have been looking forward to today's hearing is to get a better sense from our witnesses about the nuances of the term artificial intelligence AI and implications for our society in a future where AI is ubiquitous Of course one might say AI is already pervasive Since the term was first coined in the 1950s we have made huge advances in the field of artificial narrow intelligence which has been applied to many familiar every-day items such as the technology underlying Siri and Alexa Called ANI for short such systems are designed to conduct specific and usually limited tasks For example a machine that excels at playing poker wouldn't be able to parallel park my car Conversely AGI or artificial general intelligence refers to intelligent behavior across a range of cognitive tasks If you enjoy science fiction movies this definition may conjure up scenes from any number of classics such as Blade Runner The Matrix or The Terminator For many individuals the term AGI invokes images of robots or machines with human intelligence As it turns out we are decades away from realizing such AGI systems Nevertheless discussions about AGI and a future in which AGI is commonplace lead to some interesting questions worthy of analysis For example Elan Musk has been quoted as saying that AI 'is a fundamental risk to the existence of human civilization' and poses 'vastly more risk' than North Korea Does that mean that AGI may evolve to a point one day when we will lose control over machines of our own creation As far-fetched as that sounds great minds are certainly discussing such questions For the short term however my constituents are concerned about less existential issues that usually accompany new or evolving technologies-topics such as cybersecurity privacy and impacts to our nation's economy and American jobs 7 I am an original cosponsor of a bill introduced earlier this year titled the AI JOBS Act of 2018 to help our workforce prepare for the ways AI will shape the economy of the future I will also introduce legislation today to reauthorize the National Institute of Standards and Technology NIST which includes language directing NIST to support development of artificial intelligence and data science There is immense potential for AGI to help humans But that potential is also accompanied by some of the concerns I just referenced I look forward to what our panel has to share with us about the bright and the dark side of a future with AGI ### 8 Chairwoman COMSTOCK I now recognize the Ranking Member of the Research and Technology Subcommittee the gentleman from Illinois Mr Lipinski for his opening statement Mr LIPINSKI Thank you Chairwoman Comstock and thank you to Chairman Weber for holding this hearing to understand the current state of artificial intelligence technology Because of the rapid development of computational power the capacity of AI to perform new and more complicated tasks is quickly advancing Depending on who you ask AI is the stuff of dreams or nightmares I believe it is definitely the former and I strongly fear that it could also be the latter The science fiction fantasy worlds depicted on Hollywood’s big and small screens alike capture imaginations about what the world might be like if humans and highly intelligent robots shared the Earth Today’s hearing is an opportunity to begin to understand the real issues in AI and to begin to move forward with informed science-based policymaking This is a hearing that we may remember years from now hopefully as a bright beginning of a new era Current AI technologies touch a broad scope of industries and sectors including manufacturing transportation energy health care and many others As we will hear from the witnesses today artificial intelligence can be classified as artificial general intelligence or artificial narrow intelligence From my understanding it is applications of the latter such as machine learning that are underlying technologies that support some of the services and devices widely used by Americans today These include virtual assistants such as Siri and Alexa translation services such as Google Translate and autonomous vehicle technologies As the capabilities of AI improve it will undoubtedly become a more essential part of our lives and our economy While technology developers and industry look forward to making great strides in AI I want to make sure my colleagues and I in Congress are asking the tough questions and carefully considering the most crucial roles that the federal government may have in shaping the future of AI Federal investments in AI research are long-standing and we must consider the appropriate balance and scope of federal involvement as we begin to better understand the various roles AI will play in our society We are not starting from scratch in thinking about the appropriate role of the federal government in this arena In 2016 the White House issued the National Artificial Intelligence Research and Development Strategic Plan that outlines seven priorities for federally funded AI research These included making long-term investments in AI developing effective methods for human AI collaboration and addressing the ethical legal and societal implications of AI additional issues to address our safety and security public data sets standards and workforce needs Earlier this year the Government Accountability Office issued a technology assessment report led by one of our witnesses Dr Persons titled ‘‘Artificial Intelligence Emerging Opportunities Challenges and Implications ’’ While noting significant potential for AI to improve many industries including finance transportation and cybersecurity the report also noted areas where research is still needed including how to optimally regulate AI how to ensure the 9 availability and use of high-quality data understanding AI’s effects on employment and education and the development of computational ethics to guide the decisions made by software These are all critical issues but more and more I hear concern and widely varying predictions about AI’s impact on jobs AI has the potential to make some job functions safer and more efficient but it also may replace others We need to ask what are the longterm projections for the job market as AI grows In this context we also need to ask how well do our AI capabilities compare to those of other countries What education skills and retraining will the workforce of the future need These are very important questions as we think about ensuring a skilled workforce of the future that will help solidify U S leadership in AI as other countries vie for dominance in the field If AI threatens some careers it likely creates many others We need to consider what Congress should do to shape this impact to make sure Americans are ready for it and make sure the benefits of AI are distributed widely One other obvious issue of major concern when it comes to AI is ethics There are many places where this becomes relevant Currently we need to grapple with issues regarding the data that are being used to educate machines Biased data will lead to biased results from seemingly objective machines A little further down the line are many difficult questions being raised in science fiction about a world of humans and intelligent robots These are questions we will likely be called on to deal with in Congress and we need to be ready I want to thank all of our witnesses for being here today and I look forward to your testimony I’ll yield back The prepared statement of Mr Lipinski follows 10 OPENING STATEMENT Ranking Member Daniel W Lipinski D-IL ofthe Subcommittee on Research and Technology House Committee on Science Space and Technology Subcommittee on Research and Technology Subcommittee on Energy Artificial Intelligence With Great Power Comes Great Responsibility June 26 2018 Thank you Chairwoman Comstock and Chairman Weber for holding this hearing to understand the current state of artificial intelligence technology Because of the rapid development of computational power the capacity of AI to perform new and more complicated tasks is quickly advancing Depending on who you ask AI is the stuff dreams or nightmares I believe it is definitely the former and I strongly fear that it could also be the latter The science fiction fantasy worlds depicted on Hollywood's big and small screens alike capture our imaginations about what the world might be like if humans and highly intelligent robots shared the Earth Today's hearing is an opportunity to begin to understand the real issues in AI and to begin to move forward with informed science-based policymaking This is a hearing that we may remember years from now hopefully as the bright beginning of a new era Current AI technologies touch a broad scope of industries and sectors including manufacturing transportation energy healthcare and many others As we will hear from the witnesses today artificial intelligence can be classified as rutificial general intelligence or artificial narrow intelligence From my understanding it is applications of the latter such as machine learning that are the underlying technologies that support some of the services and devices widely used by Americans today These include virtual assistants such as Siri and Alexa translation services such as Google Translate and autonomous vehicle technologies As the capabilities of AI improve it will undoubtedly become a more essential part of our lives and our economy While technology developers and industry look forward to making great strides in AI I want to make sure my colleagues and I here in Congress are asking the tough questions and carefully considering the most crucial roles that the federal goverrunent may have in shaping the future of AI Federal investments in AI research are longstanding and we must consider the appropriate balance and scope of federal involvement as we begin to better understand the various roles AI will play in our society We are not starting from scratch in thinking about the appropriate role of the federal government in this arena In 2016 the White House issued the National Artificial Intelligence Research and Development Strategic Plan that outlined seven priorities for federally-funded AI research These included making long-term investments in AI developing effective methods for human-AI collaboration and addressing the ethical legal and societal implications of AI Additional issues to address are safety and security public datasets standards and workforce needs Earlier this year the Government Accountability Office issued a technology assessment report led by one of our witnesses Dr Persons titled Artificial Intelligence Emerging Opportunities Challenges and Implications While noting significant potential for AI to improve many 11 industries including finance transportation and cybersecurity the report also noted areas where research is still needed including how to optimally regulate AI how to ensure the availability and use of high quality data understanding AI's effects on employment and education and the development of computational ethics to guide the decisions made by software These are all critical issues but more and more I hear concern and widely varying predictionsabout AI's impact on jobs AI has the potential to make some job functions safer and more efficient but may replace others We need to ask what are the long-term projections for the job market as AI grows In this context we need to ask how well do our AI capabilities compare to those of other countries What education skills and retraining will the workforce of the future need These are very important questions as we think about ensuring a skilled workforce for the future that will help solidify U S leadership in AI as other countries vie for dominance in this field If AI threatens some careers it likely creates many others We need to consider what Congress should do to shape this impact make sure Americans are ready for it and make sure the benefits of AI are distributed widely One other obvious issue of major concern when it comes to AI is ethics There are many places where this becomes relevant Currently we need to grapple with issues regarding the data that are being used to educate machines Biased data will lead to biased results from seemingly objective machines A little further down the line are many difficult questions being raised in science fiction about a world of humans and intelligent robots These are questions we will likely be called on to deal with in Congress and we need to be ready I want to thank all of our witnesses for being here today and I look forward to hearing their testimony I yield back 12 Chairwoman COMSTOCK Thank you Mr Lipinski And I now recognize the Chairman of the Energy Subcommittee the gentleman from Texas Mr Weber for his opening statement Mr WEBER Madam Chair can I defer to the Chairman of the full Committee for his statement Chairwoman COMSTOCK Yes you may Mr WEBER Thank you Chairman SMITH Thank you Madam Chair Thank you Mr Chairman I didn’t know you were going to do that Madam Chair often unknown to us advances in artificial intelligence or AI touch many aspects of our lives In the area of cybersecurity AI reduces our reaction times to security threats In the field of agriculture AI monitors soil moisture and targets crop watering And in the transportation lane AI steers self-driving cars and manages intelligent traffic systems Multiple technical disciplines including quantum computing science converge to form AI Tomorrow the Science Committee will mark up the National Quantum Initiative Act which establishes a federal program to accelerate quantum research and development This is a bipartisan bill that Ranking Member Eddie Bernice Johnson and I and others will introduce today My hope is that every member of the committee will sponsor it or at least a majority Transforming our current quantum research into real-world applications will create scientific and technological discoveries especially in the field of artificial intelligence These discoveries will stimulate economic growth and improve our global competitiveness important considerations in light of China’s advances in artificial intelligence and quantum computing By some accounts China is investing $7 billion in AI through 2030 and $10 billion in quantum research The European Union has also issued a preliminary plan outlining a $24 billion public-private investment in AI between 2018 and 2020 And Russian President Putin has noted that quote ‘‘The leader in AI will rule the world ’’ end quote No doubt that’s appealing to him Yet the Department of Defense’s unclassified investment in AI was only $600 million in 2016 while federal spending on quantum totals only about $250 million a year The Committee will mark up a second piece of legislation to reauthorize the National Institute of Standards and Technology The bill directs NIST to continue supporting the development of artificial intelligence and data science including the development of machine learning and other artificial intelligence applications It is simply vital to our nation’s future that we accelerate our quantum computing and artificial intelligence efforts Thank you Madam Chair and I yield back The prepared statement of Chairman Smith follows 13 HNOLOGY Media Contacts Heather Vaughan Bridget Dunn 202 225-6371 For Immediate Release June 26 2018 Statement by Chairman Lamar Smith R-Texas Artificial Intelligence With Great Power Comes Great Responsibility Chairman Smith Often unknown to us advances in artificial intelligence-orAl-touch many aspects of our lives In the area of cybersecurity threats In the field of agriculture AI monitors soil the transportation lane AI steers self-driving cars Multiple technical disciplines including quantum AI reduces our reaction times to security moisture and targets crop watering And in and manages intelligent traffic systems computing science converge to form AI Tomorrow the Science Committee will mark up the National Quantum Initiative Act which establishes a federal program to accelerate quantum research and development This is a bipartisan bill that Ranking Member Eddie Bernice Johnson and I and others will introduce today My hope is that every member of the committee will sponsor it Transforming our current quantum research into real world applications will create scientific and technological discoveries especially in the field of artificial intelligence These discoveries will stimulate economic growth and improve our global competitivenessimportant considerations in light of China's advances in artificial intelligence and quantum computing By some accounts China is investing $7 billion in AI through 2030 and $10 billion in quantum research The European Union also has issued a preliminary plan outlining a $24 billion public-private investment in AI between 2018 and 2020 And Russian President Vladimir Putin has noted that the leader in AI will 'rule the world ' No doubt that's appealing to him Yet the Department of Defense's unclassified investment in AI was only $600 million in 2016 while federal spending on quantum totals about $250 million a year The committee will mark-up a second piece of legislation to reauthorize the National Institute of Standards and Technology NIST The bill directs NIST to continue supporting the development of artificial intelligence and data science including the development of machine learning and other artificial intelligence applications It is vital to our nation's future that we accelerate our quantum computing and AI efforts ### 14 Chairwoman COMSTOCK Thank you And I now recognize the Ranking Member of the Energy Subcommittee the gentleman from Texas Mr Veasey for an opening statement Mr VEASEY I want to thank you Chairwoman Comstock and Chairman Weber for holding this hearing today and thank you for all of the witnesses for providing expertise on this topic I’m looking forward to hearing what everyone has to say today America of course is a country of innovation and in the digital world of today more and more industries are relying on advanced technologies and connectivity to overcome new challenges Artificial intelligence and big data are impacting every facet of production and commerce AI has the ability to mimic cognitive functions such as problem-solving and learning making it a critical resource as we encounter never-before-seen problems Those in the energy sector have already seen improvements in productivity and efficiency and can expect to see even more advancement in the coming years AI can be used to process and analyze data in previously unexplored ways Technology such as sensor-equipped aircraft engines locomotive gas and wind turbines are now able to track production efficiency and wear and tear on vital machinery AI could also significantly improve our ability to detect failures before they occur and prevent disasters saving money time and lives And through the use of analytics sensors and operational data AI can be used to manage maintain and optimize systems ranging from energy storage components to power plants to the electric grid As digital technologies revolutionize the energy sector we must ensure safe and responsible execution of these processes AI systems can learn and adapt through continuous modeling of interaction and data feedback Production must be put in place to guarantee the integrity of these mechanisms as they evaluate mass quantities of machine and user data With Americans’ right to privacy under threat security of these connected systems is of the utmost importance Nevertheless I’m excited to learn about the valuable benefits that AI may be able to provide for our economy and our well-being alike With a Gartner research study reporting that AI will generate 2 3 million jobs by 2020 that’s a lot of jobs The growth AI will bring not only to the energy sector but to health care transportation education and so many others will help ensure the prosperity of our nation I look forward to seeing what light our witnesses can shed on these topics and what we can do in Congress to help enable the development and deployment of these promising technologies Madam Chairwoman I yield back the balance of my time The prepared statement of Mr Veasey follows 15 OPENING STATEMENT Ranking Member Marc Veasey D-TX of the Subcommittee on Energy House Committee on Science Space and Technology Subcommittee on Research and Technology Subcommittee on Energy Artificial Intelligence- With Great Power Comes Great Responsibility June 26 2018 Thank you Chairwoman Comstock and Chairman Weber for holding this hearing today and thank you to all our witnesses for providing their expertise on this topic I am looking forward to hearing what you all have to say America is a country of innovation and in the digital world of today more and more industries are relying on advanced technologies and connectivity to overcome new challenges Artificial Intelligence and Big Data are impacting every facet of production and commerce AI has the ability to mimic cognitive functions such as problem solving and learning making it a critical resource as we encounter never-before-seen problems Those in the energy sector have already seen improvements in productivity and efficiency and can expect to see even more advancements in the coming years AI can be used to process and analyze data in previously unexplored ways Technologies such as sensor-equipped aircraft engines locomotives gas turbines and wind turbines are now able to track production efficiency and the wear and tear on vital machinery With that technology we can expect significant reductions in fuel consumption as well as carbon emissions AI could also significantly improve our ability to detect failures before they occur and prevent disasters saving money time and lives And through the use of analytics sensors and operational data AI can be used to manage maintain and optimize systems ranging from energy storage components to power plants to the electric grid As digital technologies revolutionize the energy sector we must ensure safe and responsible execution of these processes AI systems learn and adapt through continuous modelling of interaction data and feedback Precautions must be put in place to guarantee the integrity of these mechanisms as they evaluate mass quantities of machine and user data With Americans right to privacy under threat security of these connected systems is of the utmost importance Nevertheless I am excited to learn more about valuable benefits that AI may be able to provide for our economy and wellbeing alike With a Gartner research study reporting that AI will generate 2 3 million jobs by 2020 the growth AI will bring not only to the energy sector but to healthcare transportation education and more will help ensure the prosperity of our nation I look forward to seeing what light our witnesses can shed on these topics and what we in Congress can do to enable the development and deployment of these promising technologies Thank you and I yield back the remainder of my time 16 Chairwoman COMSTOCK Thank you And I now recognize Mr Weber for his opening statement Mr WEBER Thank you Madam Chair Today we will hear from a panel of experts on next-generation artificial intelligence AI as we’ve all heard it described And while some have raised concerns about the negative consequences of AI this technology has the potential to solve fundamental science problems and improve everyday life In fact it’s likely that everyone in this room benefits from artificial intelligence For example users of voice assistants online purchase prediction fraud detection that the gentleman from Texas mentioned and music recommendation services are already utilizing aspects of this technology in their day-to-day life In the past few years the use of AI technology has rapidly expanded due to the increase in the volume of data worldwide and to the proliferation of advanced computing hardware that allows for the powerful parallel processing of this data The field of AI has broadened to include other advanced computing disciplines such as machine learning We’ve heard about neural networks deep learning computer vision and natural language processing just to name a few These learning techniques are key to the development of AI technologies and can be used to explore complex relationships and produce previously unseen results on unprecedented timescales The Department of Energy DOE is the nation’s largest federal supporter of basic research in the physical science with expertise in big-data science high-performance computing advanced algorithms and data analytics and is uniquely positioned to enable fundamental research in AI and machine learning DOE’s Office of Science Advanced Scientific Computing Research program or ASCR as we call it program develops next-generation supercomputing systems that can achieve the computational power needed for this type of critical research This includes the Department’s newest and most powerful supercomputer called Summit which just yesterday just yesterday was ranked as the fastest computing system in the entire world AI also has broad applications in the DOE mission space In materials science AI helps researchers speed the experimental process and discover new compounds faster than ever before In high-energy physics AI finds patterns in atomic and particle collisions previously unseen by scientists In fusion energy research AI modeling predicts plasma behavior that will assist in building tokamak reactors making the best of our investments in space Even in fossil fuel energy production AI systems will optimize efficiency and predict needed maintenance at power-generating facilities AI technology has the potential to improve computational science methods for any big-data problem any big-data problem And with the next generation of supercomputers the exascale computing systems that DOE is expected to field by 2021 American researchers utilizing AI technology will be able to track even bigger challenges We cannot afford to fall behind in this compelling area of research and big investments in AI by China and Europe already threaten U S dominance in this field With the immense potential 17 for AI technology to answer fundamental scientific challenges it’s quite clear we should prioritize this research We should maintain I will add American competitive edge and American exceptionalism This will help us to do that I want to thank our accomplished panel of witnesses for their testimony today and I look forward to hearing what role Congress can play and should play in advancing this critical area of discovery science And Madam Chair I yield back The prepared statement of Mr Weber follows 18 HNOLOGY For Immediate Release June 26 2018 Media Contacts Heather Vaughan Bridget Dunn 202 225-6371 Statement by Chairman Randy Weber R-Texas Artificial Intelligence- With Great Power Comes Great Responsibility Chairman Weber Today we will hear from a panel of experts on next generation artificial intelligence-or AI-technologies While some have raised concerns about the negative consequences of AI this technology has the potential to solve fundamental science problems and improve everyday life In fact it is likely that everyone in this room benefits from artificial intelligence For example users of voice assistants online purchase prediction fraud detection and music recommendation services are already utilizing aspects of this technology in their day to day life In the past few years the use of AI technology has rapidly expanded due to the increase in the volume of data worldwide and to the proliferation of advanced computing hardware that allows for the powerful parallel processing of this data The field of AI has broadened to include other advanced computing disciplines such as machine learning neural networks deep learning computer vision and natural language processing to name a few These learning techniques are key to the development of AI technologies and can be used to explore complex relationships and produce previously unseen results on unprecedented timescales The Department of Energy DOE is the nation's largest federal supporter of basic research in the physical sciences with expertise in big data science high performance computing advanced algorithms and data analytics and is uniquely positioned to enable fundamental research in AI and machine learning DOE's Office of Science Advanced Scientific Computing Research program or ASCR program develops next generation supercomputing systems that can achieve the computational power needed for this critical research This includes the Department's newest and most powerful supercomputer called Summit which just yesterday was ranked as the fastest computing system in the world AI also has broad applications in the DOE mission space In materials science AI helps researchers speed the experimental process and discover new compounds faster than ever before In high energy physics AI finds patterns in atomic and particle collisions previously unseen by scientists 19 In fusion energy research AI modeling predicts plasma behavior that will assist with building tokamak reactors making the best of our investment in this space Even in fossil energy production AI systems will optimize efficiency and predict needed maintenance at power generating facilities AI technology has the potential to improve computational science methods for any big data problem And with the next generation of supercomputers the exascale computing systems that DOE is expected to field by 2021 American researchers utilizing AI technology will be able to tackle even bigger challenges We can't afford to fall behind in this compelling area of research-and big investments in AI by China and Europe already threaten U S dominance in this field With the immense potential for AI technology to answer fundamental scientific challenges it's clear we should prioritize this research I want to thank our accomplished panel of witnesses for their testimony today and I look forward to hearing what role Congress should play in advancing this critical area of discovery science ### 20 The prepared statement of Full Committee Ranking Member Eddie Bernice Johnson 21 OPENING STATEMENT Ranking Member Eddie Bernice Johnson D-TX House Committee on Science Space and Technology Subcommittee on Research and Technology Subcommittee on Energy Artificial Intelligence With Great Power Comes Great Responsibility June 26 2018 Thank you Chairwoman Comstock Ranking Member Lipinski Chairman Weber and Ranking Member Veasey for holding this important hearing and thank you to the witnesses for being here today Artificial intelligence or AI is already affecting the way we make decisions and interact in our world When fully developed AI will likely have a large impact across all sectors industries and demographics This committee should support policies that ensure safe responsible development of AI that will benefit society and make the world a better place The Science committee should also be stewards for bridging the digital divide and supporting efforts to include broad participation in AI and to prevent algorithmic bias AI's current and potential applications are numerous including virtual personal assistants and predictive technologies that enhance customer interactions From healthcare to national security and everything in-between advancement of AI depends on building foundational knowledge in the many fields that contribute to AI including the social behavioral and economic sciences Federal investment in basic research for AI is critical The top spot for world leadership is open for the taking Our counterparts across the globe have already staked their claim and are investing heavily in R D and in producing a well-qualified talent pool Strategic robust federal investments in AI R D and in educating and training our future AI workforce will be critical if we are to remain a leader in AI These efforts should be directed toward attracting students to the STEM fields and to ensuring our seasoned workforce gains the skills necessary to thrive in an AI-driven economy This committee has been at the forefront of supporting game-changing research that advances the boundaries of science and creates completely new industries that could not have been imagined by earlier generations We have the opportunity to do that with AI if we commit to sustainable funding of research education and coordination activities I look forward to hearing more about what the future holds for this exciting technology Thank you and I yield back 22 Chairwoman COMSTOCK Thank you And I will now introduce today’s witnesses Our first witness today is Dr Tim Persons Chief Scientist at the U S Government Accountability Office He also serves as a Director for GAO’s Center for Science Technology and Engineering Dr Persons received a Bachelor of Science in physics from James Madison University and a Master of Science in nuclear physics from Emory University He also earned a Master of Science in computer science and Ph D in biomedical engineering both from Wake Forest University Next we have Mr Greg Brockman our second witness who is Cofounder and Chief Technology Officer of OpenAI a nonprofit artificial intelligence research company Mr Brockman is an investor in over 30 startups and a board member of the Stellar digital currency system He was previously the CTO of Stripe a payments startup now valued at over $9 billion And he studied mathematics at Harvard and computer science at MIT And our final witness is Dr Fei-Fei Li Chairperson of the Board and Cofounder of AI4ALL In addition Dr Li is a Professor in the Computer Science Department at Stanford and the Director of the Stanford Artificial Intelligence Lab In 2017 Dr Li also joined Google Cloud as Chief Scientist of AI and machine learning Dr Li received her Bachelor of Arts in physics from Princeton and her Ph D in electrical engineering from the California Institute of Technology I now recognize Dr Persons for five minutes to present his testimony TESTIMONY OF DR TIM PERSONS CHIEF SCIENTIST U S GOVERNMENT ACCOUNTABILITY OFFICE Dr PERSONS Good morning Thank you Chairwoman Comstock Chairman Weber Ranking Members Lipinski and Veasey and Members of the Subcommittee I’m pleased to be here today to discuss GAO’s technology assessment on artificial intelligence To ensure the United States remains a leader in AI innovation special attention will be needed for our education and training systems regulatory structures frameworks for privacy and civil liberties and our understanding of risk management in general AI holds substantial promise for improving human life increasing the nation’s economic competitiveness and solving some of society’s most pressing challenges Yet as a disruptive technology AI poses risks that could have far-reaching effects on for example the future labor force economic inclusion and privacy and civil liberties among others Today I’ll summarize three key insights arising from our recent work First the distinction between narrow versus general AI second the expected impact of AI on jobs competitiveness and workforce training and third the role the federal government can play in research standards development new regulatory approaches and education Regarding narrow versus general AI narrow AI refers to applications that are task-specific such as tax preparation software voice and face recognition systems and algorithms that classify the content of images General AI refers to a system exhibiting intelligence 23 on par with or possibly exceeding that of humans While science fiction has helped general AI capture our collective imaginations for some time it is unlikely to be fully achieved for decades if at all Even so considerable progress has been made in developing narrow AI applications that outperform humans in specific tasks and are thus invoking crucially important economic policy and research considerations Regarding jobs competition and the workforce there is considerable uncertainty about the extent to which jobs will be displaced by AI and how many—how much any losses will be offset by job creation In the near term displacement to certain jobs such as call-center or retail workers may be particularly vulnerable to automation However in the long term demand for skills that are complementary to AI is expected to increase resulting in greater productivity To better understand the impact of AI on employment moving forward several experts underscored the need for new data and methods to enable greater insight into this issue Regarding the role of the federal government it will continue its crucial role in research and data-sharing contributions to standards development regulatory approaches and education One important research area of the federal government could support is enhancing the explainability of AI which could help establish trust in the behavior of AI systems The federal government could also incentivize data-sharing including federal data that are subject to limitations for how they can be used as well as creating frameworks for sharing data to improve the safety and security of AI systems Such efforts may include supporting standards for explainability data labeling and safety including risk assessment and benchmarking of AI performance against the status quo It’s always risk versus risk Related to this new regulatory approaches are needed including the development of regulatory sandboxes for testing AI products services and business models especially in industries like transportation financial services and health care GAO’s recent report on fintech found for example that regulators use sandboxes to gain insight into key questions issues and unexpected risks that may arise out of the emerging technologies New rules governing intellectual property and data privacy may also be needed to manage the deployment of AI Finally education and training will need to be reimagined so workers have the skills needed to work with and alongside emerging AI technologies For the United States to remain competitive globally and effectively manage AI systems its workers will need a deeper understanding of probability and statistics across most if not all academic disciplines that is not just the physical engineering and biological sciences as well as competency and ethics algorithmic auditability and risk management In conclusion the emergence of what some have called the fourth industrial revolution and AI’s key role in driving it will require new frameworks for business models and value propositions for the public and private sectors alike Even if AI technologies were to cease advancing today no part of society or the economy would be directly or indirectly untouched by its transformative effects 24 I thank the committee leadership of the committees Thanks to the members here for your holding a hearing on this very important topic today for such a time as this Madam Chairwoman Mr Chairman Ranking Members this concludes my prepared remarks I would be happy to respond to any questions that you or other Members of the Subcommittees have at this time The prepared statement of Dr Persons follows 25 United States Government Accountability Office Testimony Before the Subcommittees on Research and Technology and Energy Committee on Science Space and Technology House of Representatives For Release on Delivery Expected at 10 30 a m ET Tuesday June 26 2018 ARTIFICIAL INTELLIGENCE Emerging Opportunities Challenges and Implications for Policy and Research Statement of Timothy M Persons Chief Scientist Applied Research and Methods GA0-18-644T 26 Chairwoman Comstock Chairman Weber and Ranking Members Lipinski and Veasey Thank you for the opportunity to discuss our work on artificial intelligence AI My testimony today summarizes our March 2018 technology assessment entitled Artificial Intelligence Emerging Opportunities Challenges and Implications 1 According to experts AI holds substantial promise for not only improving human life and economic competitiveness in a variety of capacities but also helping to solve some of society's most pressing challenges At the same time AI poses new risks and has the potential to displace workers in some sectors requires new skills and adaptability to changing workforce needs and could exacerbate socioeconomic inequality Our March 2018 report and my statement today address the following topics How has AI evolved over time According to experts what are the opportunities and future promise as well as the principal challenges and risks of AI According to experts what are the policy implications and research priorities resulting from advances in AI For our March 2018 report the Comptroller General of the United States convened a Forum on Artificial Intelligence a meeting of 21 expert participants held on July 6 and 7 2017 with the assistance of the National Academy of Sciences 2 The work for the report also included a review of relevant literature and consultation with additional subjectmatter experts Additional information about our scope and methodology can be found in our report We performed the work on which this testimony is based in accordance with all sections of GAO's Quality Assurance Framework that are relevant to technology assessments GAO currently has work underway on how automation is affecting labor markets which we expect to publish in early 2019 Because of the 1 GAO Artificial Intelligence Emerging Opportunities Challenges and Implications GA0-18-142SP Washington D C March 28 2018 2 Forum participants were from academia business government and nonprofit organizations For a complete list of participants see Appendix I of GA0-18-142SP Page 1 GA0 18-644T Artificial Intelligence 27 strategic importance of AI in the health care sector GAO is planning a forum to develop a technology assessment of this area as well The Evolution and Characteristics of AI Several Definitions and Taxonomies of AI Exist The field of artificial intelligence can be traced back to a 1956 workshop organized by John McCarthy held at Dartmouth College The workshop's goal was to explore how machines could be used to simulate human intelligence Numerous factors primarily the trends underlying big data i e increased data availability storage and processing power have contributed to rapid innovation and accomplishments in AI in recent years ' As we noted in our March 2018 technology assessment there is no single universally accepted definition of AI but rather differing definitions and taxonomies In addition to defining AI overall researchers have distinguished between narrow and general AI Narrow AI refers to applications that provide domain-specific expertise or task completion whereas general AI refers to an AI application that exhibits intelligence comparable to a human or beyond across the range of contexts in which humans interact While there has been considerable progress in developing AI that outperforms humans in specific domains some observers believe that general AI is unlikely to be achieved for decades in the future AI Has Been Conceptualized as Having Three Waves of Development In our March 2018 work we noted that rather than focusing on a specific definition of AI it can be understood in terms of the waves in which the technology has developed Launchbury 2016 provides a framework that conceptualizes AI as having three waves based on differences in capabilities with respect to perceiving learning abstracting and reasoning 4 3 For more on trends underlying big data see for example GAO Highlights of a Forum Data and Analytics Innovation Emerging Oppottunities and Challenges GA0-16-659SP Washington D C Sept 20 2016 4John Launchbury A DARPA Perspective on Artificial Intelligence 2016 Page2 GA0·18-644T Artificial Intelligence 28 The first wave of AI is represented by expert knowledge or criteria developed in law or other authoritative sources and encoded into a computer algorithm which is referred to as an expert system Examples of expert systems include programs that schedule logistics or prepare taxes Second-wave AI technology is based on machine learning or statistical learning and includes natural-language processing e g voice recognition and computer-vision technologies among others In contrast to first-wave systems second-wave systems are designed to perceive and learn Examples of second-wave systems include voiceactivated digital assistants applications that assist healthcare workers in selecting appropriate treatment options or making diagnoses and self-driving automated vehicles Third-wave AI technologies combine the strengths of first- and second-wave AI and are also capable of contextual sophistication abstraction and explanation An example of third-wave AI is a ship that can navigate the sea wrthout human intervention for a few months at a time while sensing other ships navigating sea lanes and carrying out necessary tasks As described by Launchbury we are just at the beginning of the third wave of AI and further research remains before third-wave technologies become prevalent An important part of third-wave AI will be developing systems that are not only capable of adapting to new situations but also are able to explain to users the reasoning behind these decisions Forum Participants Identified Several Benefits of Artificial Intelligence and Challenges to Its Development The increased adoption of artificial intelligence will bring with it several benefits as well as a number of challenges According to participants at the forum we convened for our March 2018 technology assessment both benefits and challenges will need to be carefully considered alongside one another Figure 1 summarizes selected questions benefits and challenges regarding the use of AI in four high-consequence sectors Participants also stressed that there may be benefits related to AI that cannot yet be predicted or may even be hard to imagine Page 3 GA0-18-644T Artiflciallntelllgence 29 Figure 1 Selected Questions Regarding the Use of Artificial Intelligence AI in Four High-Consequence Sectors Selec1ed Quest1ons Cybersecurity AI applications face threats from cybersecurtty attacks but AI also may be used as a tool for detecting and defending agafnst attacks How can ·aUtonomOuS sy m$ 'be made stffii lg without in' ' ion · ·· · · · · ' How us$tul a ri k-ba'sed pproacl to cfeteim f inO r ri Je rning algOrithms ad ore to legal ' requirements or cethicat nOJm$' ' '•' •' Automated Vehicles Automated vehicles hold promise for increasing driving safety and providing enhanced mobility but pose challenges for assuring increased safety ' whati ur itof¥ rn newort 10r automaied safety' asSilfaiti •' Wt at r · · ' f · n nand ofliabi ity a nd Criminal Justice The use of AI in criminal justice may improve the allocation of law enforcement resources and has the potential to reduce crime and jail populations but also raises concerns about privacy and civil rights violations $ Financial Services The use of AI in financial services could improve client services and enhance surveillance monitoring but also poses challenges to ensuring fair lending attracting and retaining staff with requisite skills and maintaining hardware and software What are th ·options for assessing accuracy arid the potential fo bias in AI data and algoritt Jms • WhataresotutiOnafor safeQuarding privacy in the ection lind use of personal Information· by AI systems What are the mechanisms to address ethical considerations tradeoffs and protectiOns How can regulatory sandboxes be used to test new AI products services and business models SOI Irce GAOForumonArbfie alln elhgenee I GA0-1S 644T Benefits Identified by Forum Participants Improved economic outcomes and increased levels of productivity It may be difficult to accurately predict what Al's impact on the economy could be according to one forum participant In previous periods large investments in automation have been highly correlated with improvements in productivity and economic outcomes which according to one forum participant has led some to believe that transformations as a result of AI could have the same outcome This same participant noted however that no one collects the data needed to measure the impact AI Page4 GA0·18 644T Artificial Intelligence 30 or other types of advanced automation may have on the economy According to another participant whatever the effect that AI will have on productivity in particular and the economy in general the changes will occur quickly and be difficult to predict Improved or augmented human decision making AI can be used to gather an enormous amount of data and information from multiple locations characterize the normal operation of a system and detect abnormalities much faster than humans can In addition AI could be used to create data-informed policy that may help prevent inappropriate or harmful human bias-be it from political pressure or other factors-from creating undesirable results according to one participant However as another participant at the forum noted AI is no guarantee of freedom from bias The participant stressed specifically that if the data being used by AI are biased the results will be biased as well AI can help prevent inappropriate or harmful human bias according to this same participant if it is carefully used if the assumptions of the models are thoughtfully considered and most importantly if the outputs of the model are constantly and closely verified Insights into complex and pressing problems Some of the participants at our forum believed that AI has the potential to provide insights into-and even help solve-some of the world's most complex and pressing problems For example one participant stated that as the number of elderly Americans continues to grow AI could be used to provide medication management mobility support housework meal preparation and rehabilitation services to a growing number of people who need assistance with day-to-day activities In addition there are other complex and pressing problems that may eventually be solved by the adoption of AI According to one participant AI could eventually be used to assure regulatory compliance in the financial sector without unnecessary burden on those being regulated PageS GA0-18-644T Artificial Intelligence 31 Challenges Identified by Forum Participants Barriers to collecting and sharing data While not all applications of AI require massive amounts of data certain applications that use machine learning algorithms do 5 This can be a problem in sectors where data are not easily aggregated or interpreted or readily available Such is the case with criminal justice where the ways in which data are collected and organized vary from jurisdiction to jurisdiction Such is also true with most vulnerable populations and developing countries where data have not yet been collected Lack of access to adequate computing resources and requisite human capital Forum participants told us that AI researchers and developers need access to storage and processing both of which are expensive and sometimes difficult to access at the necessary scale Some forum participants also shared concerns that the accelerated pace of change associated with AI is straining the education and workforce systems' capacity to train and hire individuals with the appropriate skill sets leaving many companies struggling to find workers with relevant knowledge skills and training Adequacy of current laws and regulations The widespread adoption of AI may according to some forum participants have implications regarding the adequacy of current laws and regulations For example one participant noted that current patent and copyright laws provide only limited protection for software and business methods and questioned whether these laws will protect the products created by AI At the same time one of the participants at the forum raised concerns about ways in which AI could be used to violate civil rights This participant cautioned for example that if law enforcement considers race class or gender in AI that is used to assess risk there is the possibility that a defendant's equal protection rights under the 14th Amendment may be violated as well as their due process rights under the 5th and 14th Amendments Ethical Framework for and Explainability and Acceptance of AI The adoption of AI also introduces ethical implications According to a forum participant there is a need for a system of computational ethics to help AI choose options that reflect agreed-upon values Moreover some of the participants at the forum noted that before humans will understand Bauguess The Role of Big Data Machine Learning and AI in Assessing Risks a Regulatory Perspective U S Securities and Exchange Commission Keynote Address to OpRisk North America 2017 New York New York June 21 2017 Page 6 GA0-18-644T Artificial Intelligence 32 appropriately trust and be able to effectively manage AI an AI application or system needs to explain why it took certain actions and why it valued certain variables more than others Forum Participants Identified Several Cross-Cutting Policy Considerations Related to AI and Several Areas Where More Research Is Needed After discussing the benefits and challenges associated with AI the participants at the forum we convened for our March 2018 technology assessment highlighted a number of policy considerations and areas of future research see fig 2 Figure 2 Implications of Artificial Intelligence AI for Policy and Research Source GAO Forum on Artificial lnte1 rgence I GA0·18 S44 T lncentivizing data sharing Forum participants emphasized the need for establishing a safe space to protect sensitive information e g intellectual property and brand information while sharing data Another participant cautioned that for such a safe space to succeed it will need to start with a few manufacturers and clearly define the data that are needed and the specific scenarios in which the data will be used Page7 GA0 18 644T Artificial Intelligence 33 Certain forum participants also expressed concerns that many potentially useful data are guarded by federal agencies that do not provide access to researchers Participants noted successful data-sharing efforts through entities such as MITRE and the National Institute of Standards and Technology NIST In particular some participants highlighted datasharing efforts to improve safety outcomes For instance one participant mentioned that researchers at MITRE had credited data-sharing efforts in the aviation industry employing a safe space with reducing the number of accidents Another participant emphasized the importance of sharing data to better understand safety outcomes associated with automated vehicles stating i f we're going to trust that these vehicles can go out on the road we need to verify that in fact out on the road they are as safe as we think they are Forum participants highlighted other proposed future data-sharing efforts citing the benefits of assessing data from multiple sources to improve outcomes According to one forum participant the National Science and Technology Council Subcommittee on Machine Learning and Artificial Intelligence is working collaboratively among federal departments and agencies to promote the sharing of government data to help develop innovative solutions for social good This sharing may include creating training environments-safe spaces-in which sensitive data are protected among other things Another participant noted that in the criminal-justice sector the federal system could be used as a test bed for various reforms-including data sharing reforms-because the federal system is unified This participant argued that if the federal system could find a way to share data related to risk assessments and other areas and show that the data are being utilized in an evenhanded way the reforms pioneered by the federal system would likely migrate down to the individual state systems This same participant also stated that the Bureau of Justice Assistance and the Bureau of Justice Statistics may be the best positioned to initiate any nationwide data standardization and collection projects Improving safety and security Participants highlighted challenges and opportunities to enhancing the safety and security of system applications from cyber attacks including those with AI features One participant said that the costs of cybersecurity in all forms of network computing are not being shared appropriately and that security breaches are much costlier than the security measures that are needed to prevent breaches This participant said that policymakers will need to consider creating some kind of framework that ensures costs-and liabilities-are appropriately Page 8 GA0-18-644T Artificial Intelligence 34 shared between manufacturers and users In addition two participants said that policymakers should consider creating a new regulatory structure to better ensure the safety of automated vehicles Updating the regulatory approach The widespread adoption of AI will have implications for regulators and lawmakers will need to consider policy options to address these issues according to multiple forum participants One participant reinforced the need for regulators to be proactive including a commitment of resources because change is occurring so rapidly and in unanticipated ways For example as a policy matter going forward one participant explained a new regulatory structure for automated vehicles needs to evolve and that accordingly the federal government should avoid setting standards prematurely Another interrelated issue raised by a participant about automated vehicles concerned how liability would be regulated Currently according to this participant the manufacturer of the automated vehicle bears all responsibility for crashes even if these vehicles improve overall public safety Some of the participants at the forum also raised concerns about privacy including ways in which AI could be used by law-enforcement agencies to violate civil liberties and said that this is an area that needs policy solutions In addition one of the forum participants said that policymakers should consider allowing financial regulators to explore alternative regulatory approaches and reporting mechanisms leveraging technology to improve and reduce the burden of regulation In this regard one participant discussed the merits of regtech that is linking regulation with technology Another participant noted that other laws and regulations may need to be adapted to account for the fact that humans may not always be behind decisions that are made by automated systems For example this participant discussed laws where intent plays a key role as is the case in financial market manipulation If someone programs AI to make money and it does so in a nefarious way it is not clear how current laws could be used to prosecute the creator of the AI Assessing acceptable risks and ethical decision making Policymakers need to decide how they are going to measure or benchmark the performance of AI and assess the trade-offs according to one participant who stressed that the baseline is current practice not perfection i e how humans are performing now absent AI As this participant emphasized i f we have to benchmark AI against Page 9 GA0-18-644T Artificial Intelligence 35 perfection the perfect will be the enemy of the good and we get nowhere Several participants at the forum emphasized that such regulatory questions should be resolved by a variety of stakeholders including economists legal scholars philosophers and others involved in policy formulation and decision making and not solely scientists and statisticians Participants at our AI forum also highlighted several areas they believe deserve more research in terms of new regulatory frameworks data labeling employment and education and explainable AI and computational ethics Establishing regulatory sandboxes In finance there is a worldwide movement to create so-called regulatory sandboxes according to one participant where regulators can begin experimenting on a small scale and empirically testing new ideas As this participant explained regulatory sandboxes would provide a safe haven to assess the results of alternative regulatory approaches Developing high-quality labeled data One participant emphasized the importance of data collection and how to obtain high-quality labeled data This encompasses improving the quality of the data during data collection Another participant we spoke with highlighted the merits of developing adequately labeled data sets As data become more comprehensive and organized or labeled in a manner that facilitates machine learning AI tools can produce more accurate outcomes Understanding Al's effect on employment and reimagining training and education Some forum participants offered mixed views concerning the impacts associated with AI on employment while acknowledging the uncertainties For instance some forum participants noted that job losses in some areas were likely while noting the potential for job increases in other areas One participant advocated for research to better understand how jobs have been changing There is currently no comprehensive federal data source with information on the employment effects AI may have in manufacturing and other segments of the economy Further according to two participants in the absence of a comprehensive datacollection effort it is unclear which jobs will be created by AI which jobs may be augmented or which jobs are likely to be displaced by AI The widespread adoption of AI also brings with it a need to reevaluate and reimagine training and education according to some of the participants Page 10 GA0-18-644T Artificial intelligence 36 Exploring computational ethics and explainable AI According to one participant we will have to design systems that are going to operate in environments where we cannot anticipate in advance all the things that could go wrong Explainable AI and computational ethics are relevant for all places where AI systems are interacting with the physical world As for computational ethics AI researchers have begun establishing rules of their own For example some groups of technologists have created sets of ethical considerations• In addition researchers from six institutions recently formed a group called PERVADE Pervasive Data Ethics for Computational Research whose mission is to develop a clearer ethical process for big-data research for use by both universities and private companies However as one participant noted the current and future developers of AI systems may operate by ethical standards or adhere to certain morals or values that may not be compatible with the rest of society or representative of those who will use the AI In conclusion in our March 2018 technology assessment we noted that AI technologies are already impacting a wide array of economic sectors Our technology assessment also provides an overview of developments in the field of AI focusing on the challenges opportunities and implications of these developments for policy making and research and further helps clarify the prospects for the near-term future of AI and identifies areas where changes in policy and research may be needed Chairwoman Comstock Chairman Weber and Ranking Members Lipinski and Veasey this concludes my statement I would be pleased to respond to any questions you or other Members may have GAO Contact and Staff Acknowledgments If you or your staff have any questions about this testimony please contact Timothy Persons at 202 512-6522 or personst@gao gov Contact points for our Offices of Congressional Relations and Public Affairs may be found on the last page of this statement Individuals making key contributions to this testimony include Stephen Sanford Assistant Director Virginia Chanley Analyst-in-Charge and David Chrisinger Key contributors to the prior work on which this testimony is based are listed in the product Mehr Artificial Intelligence for Citizen Services and Government H Ash Center for Democratic Governance and Innovation Harvard Kennedy Schoo August 2017 102880 Page 11 GA0 18-644T Artificial Intelligence 37 This is a work of the U S government and is not subject to copyright protection in the United States The published product may be reproduced and distributed in its entirety without further permission from GAO However because this work may contain copyrighted images or other materia permission from the copyright holder may be necessary if you wish to reproduce this material separately 38 GAO's Mission The Government Accountability Office the audit evaluation and investigative arm of Congress exists to support Congress in meeting its constitutional responsibilities and to help improve the performance and accountability of the federal government for the American people GAO examines the use of public funds evaluates federal programs and policies and provides analyses recommendations and other assistance to help Congress make infonmed oversight policy and funding decisions GAO's commitment to good government is reflected in its core values of accountability integrity and reliability Obtaining Copies of GAO Reports and Testimony The fastest and easiest way to obtain copies of GAO documents at no cost is through GAO's website https lwww gao gov Each weekday afternoon GAO posts on its website newly released reports testimony and correspondence To have GAO e-mail you a list of newly posted products go to https llwww gao gov and select E-mail Updates Order by Phone The price of each GAO publication reflects GAO's actual cost of production and distribution and depends on the number of pages in the publication and whether the publication is printed in color or black and white Pricing and ordering information is posted on GAO's website https llwww gao gov ordering htm Place orders by calling 202 512-6000 toll free 866 801-7077 or TDD 202 512-2537 Orders may be paid for using American Express Discover Card MasterCard Visa check or money order Call for additional information Connect with GAO To Report Fraud Waste and Abuse in Federal Programs Connect with GAO on Facebook Flickr Twitter and YouTube Subscribe to our RSS Feeds or E-mail Updates Listen to our Podcasts Visit GAO on the web at https www gao gov Contact Website https l www gao gov fraudnetlfraudnet htm Automated answering system 800 424-5454 or 202 512-7470 Congressional Relations Orice Williams Brown Managing Director WilliamsO@gao gov 202 512-4400 U S Government Accountability Office 441 G Street NW Room 7125 Washington DC 20548 Public Affairs Chuck Young Managing Director youngc1@gao gov 202 512-4800 U S Government Accountability Office 441 G Street NW Room 7149 Washington DC 20548 Strategic Planning and External Liaison James-Christian Blackwood Managing Director spel@gao gov 202 512-4707 U S Government Accountability Office 441 G Street NW Room 7814 Washington DC 20548 Please Print on Recycled Paper 39 Timothy M Persons Ph D Chief Scientist United States Government Accountability Office Biography Dr Timothy M Persons is the Chief Scientist of the United States Government Accountability Office as well as a Director for GAO's Center for Science Technology and Engineering CSTE In these roles he is an expert on advanced analytics and cutting-edge science and technology key highly-specialized national and international systems engineering policies best practices and original research studies in the fields of engineering computer and the physical and biological sciences to ensure efficient effective and economical use of science and technology in government programs He has led and conducted advanced scientific and technical studies for the U S Congress on topics ranging from artificial intelligence data analytics bioforensics oversight of high containment laboratories 3D printing nonmanufacturing biodetection systems homeland security imaging and nuclear detection systems and freshwater conservation technologies 40 Chairwoman COMSTOCK Thank you And I now recognize Mr Brockman for five minutes TESTIMONY OF MR GREG BROCKMAN CO-FOUNDER AND CHIEF TECHNOLOGY OFFICER OPENAI Mr BROCKMAN Chairwoman Comstock Chairman Weber Ranking Member Lipinski Ranking Member Veasey members of both subcommittees thank you for having me today to deliver testimony I’m Greg Brockman Cofounder of OpenAI a San Francisco-based nonprofit with a mission to ensure that artificial general intelligence which we define as systems—the highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity Now I’m here to tell you about the generality of modern AI why AGI might actually be in reach sooner than commonly expected and what action policymakers can take today So first what’s OpenAI We’re a research company with one of the world’s most advanced AI research and development teams Yesterday we announced major progress towards a milestone that we Alphabet’s subsidiary DeepMind and Facebook have separately been trying to reach which is solving complex strategy games which start to capture many aspects of the real world that were just not seen in board games like chess or Go We built a system called OpenAI Five which learned to devise long-term plans and navigate scenarios far too complex to be programmed in by a human in order to solve a massively popular competitive game called Dota 2 Now in the past AI-like technology was written by humans in order to solve one specific problem at a time It was not capable of adapting to solve new problems Today’s AI it’s all based on one core technique which is the artificial neural network a single simple idea that as it’s run on faster computers is proving to match a surprising amount of human capability And this was in fact something that was shown in part by my fellow witnesses Dr Li’s work in image recognition And artificial neural networks can be trained to perform speech recognition or computer vision It just depends on the data that they’re shown Now further along the spectrum of generality is AGI Rather than being developed for any one use case AGI would be developed for a wide range of important tasks and AGI would also be useful for noncommercial applications including thinking through complex international disputes or city planning Now people have been talking about AGI for decades and so how should we think about the timeline Well all AI systems they’re built on three foundations That’s data computational power and algorithms Next-generation AI systems are already starting to rely less on conventional data sets where a human has provided the right answer For example one of our recent neural networks learned by reading 7 000 books We also recently released a study showing that the amount of computation powering the largest AI training runs has been doubling every 3–1 2 months since 2012 That’s a total increase of 300 000 times And we expect this to continue for the next five 41 years using only today’s proven hardware technologies and not assuming any breakthroughs like quantum or optical Now to put that in perspective that’s like if your phone battery which today lasts for a day started to last for 800 years and then five years later started to last for 100 million years It’s this torrent of compute this tsunami of compute We’ve never seen anything like this And so the open question is will this massive increase in combinational power combined with near-term improvements in algorithmic understanding be enough to develop AGI We don’t know the answer to this question today but given the rapid progress that we are seeing we can’t confidently rule it out And so now what should we be thinking about today What can policymakers be doing today And so you know the first thing to recognize is the core danger of AGI is that it has fundamentally the potential to cause rapid change whether that’s through machines pursuing goals that are mis-specified by their operator whether it’s through malicious humans subverting deployed systems or whether it’s an economy that grows in an out-of-control way for its own sake rather than in order to improve human lives Now we spent two years worth of policy research to create the OpenAI Charter which in fact is a document I have right here in front of me This contains three sections defining our views on safe and responsible AGI development So that’s—one is leaving time for safety and in particular refusing a race to the bottom on safety in order to reach AGI first The second is to ensure that people at large rather than any one small group receive the benefits of this transformative technology And the third is working together as a community in order to solve safety and policy challenges Now our primary recommendation to policymakers is to start measuring progress in this field We need to understand how fast the field is moving what capabilities are likely to arrive when in order to successfully plan for AGI challenges That moves towards forecasts rather than intuition Measurement is also a place where international coordination is actually valuable and this is important if we want to spread safety and ethical standards globally So thank you for your time and I look forward to questions The prepared statement of Mr Brockman follows 42 Testimony of Mr Greg Brockman Chairwoman Barbara Comstock Chairman Randy Weber Ranking Member Daniel Lipinski Ranking Member Marc Veasey members of both subcommittees thank you for having me today to deliver testimony on this important topic I'm Greg Brockman co-founder of OpenAI a non-profit artificial intelligence development organization Our mission is to ensure that artificial general intelligence AGI - by which we mean highly autonomous systems that outperform humans at most economically valuable work - benefits all of humanity OpenAI OpenAI has three main arms capabilities safety and policy All three of these areas working in concert are crucial to achieve our mission Our capabilities arm is one of the most advanced AI research and development teams in the world Our safety arm is responsible for developing techn'tques to ensure that AGI-Ievel systems will operate as their human operators intend Our policy arm is responsible for researching AGI's social challenges and providing information to policymakers On the capabilities side one milestone in the field is solving complex strategy games 1 which capture many of the aspects of the real world not seen in previous milestones like Chess or Go We recently announced OpenAI Five 2 a system we've created which has reached the semiprofessional level at one of the most complex games played by humans a ten-player team strategy video game called Dota This system devises long-term plans and navigates scenarios far too complex to be programmed in by any human We are aiming to play against top professionals during the Dota world championships in August OpenAI Five taught itself the rules of the game by playing 180 years worth of games against itself each day For comparison top human professionals have at least 12 000 hours of gameplay so our system sees as many games each day as 100 human professionals have seen in their lifetimes On the safety side we recently developed a proof-of-concept technique 3 for allowing humans to monitor the behavior of advanced AI systems We have also collaborated with Alphabet's subsidiary DeepMind to design AI systems which can learn from the implicit preferences of human trainers 1 Gershgorn Dave The massive global race to teach an AI to beat Starcraft II is under way https qz com 1051052 deepmind-goog-and-facebook-fb-have-started-the-global-sprint-for-ai-to-beatstarcraft-ii 2 OpenAI Five https bloq openai corn openai-five 3 Irving Geoffrey et al AI Safety via Debate https bloq openai com debate 43 On the policy side we recently co-authored a report 4 forecasting how malicious actors could misuse AI including recommendations of how to mitigate these threats We're helping to develop the AI Index an AI measurement and analysis initiative as part of the Stanford One Hundred Year Study on AI Our goal is to use this experience to make recommendations about how policymakers can measure and analyze the impact of AI on society And we are attempting to nurture the field of AI policy to ensure there is a deep bench of thinkers about AI across all important actors- companies research labs and governments Narrow vs general AI People often talk about narrow vs general AI in terms of whether they apply to one task narrow or many tasks general But there's also a dimension of competence can they solve only easy tasks or can they solve hard tasks In practice to build AI systems that solve harder problems we've ended up creating increasingly general learning systems- since we let the machine learn more on its own rather than having a human provide knowledge or guidance Specifically in the past Al-like technology was written by humans to solve one specific problem It wasn't capable of adapting to solve new problems In contrast today's AI is all based on one core technique the artificial neural network in a form devised in the 1g8os This is a single simple idea that is as it scales is proving itself to be able to match a surprising amount of human capability Our neural networks still have a lot of room to grow- to give a sense of scale though the numbers are not directly comparable today they usually have around the same number of artificial neurons as an insect has biological ones In the 1980s computers could only run tiny neural networks so the resulting systems couldn't solve interesting problems In 2012 computers were fast enough for a team of researchers including my co-founder llya Sutskever to train a large enough neural network to perform well on the task of categorizing images- performing far better than any other method The neural network learned to categorize images by being shown many examples of already-categorized images and this is now the dominant approach in the field rather than the previously handcrafted rules which were limited in performance Since then neural networks have become the standard tool for solving problems in a variety of fields such as speech recognition and machine translation To give you a sense of progress here are some AI advances from recent years Image recognition Al's ability to correctly categorize images has gone from 75% preneural network 2011 accuracy to around 98% accuracy neural network 2017 on a difficult standard benchmark on which human accuracy is around 95% 4 Clark Jack et al Preparing for Malicious Uses of AI https blog openai com preparing-formalicious-uses-of-ai 44 Fake images videos AI techniques are increasingly able to generate convincing fake images and videos- including fakes of politicians such as Vladimir Putin and President Trump In 2014 the best generated images were low-resolution images of fake people by 2017 they were photorealistic faces that humans have trouble distinguishing from real ones5 Also in 2017 free software became available allowing people to create their own deepfake images Translation In 2014 researchers developed neural machine translation - where computers learn to translate between languages using only large datasets lacking any of the specific rules which human translators use to do their work In 2016 Google Translate performance significantly improved by switching to it in 2017 Facebook improved its site translation by doing the same Speech recognition Due to switching to neural networks over the past few years speech recognition went from barely working such as we've all experienced when calling an automated phone tree to running on smartphones with much higher accuracy Sophistication The complexity of games where neural network-based Als can rival the top human players has increased in complexity from 1970s Atari games like Space Invaders or Breakout 2013 to rich strategic games like the board game Go 2015 to modern real-time strategy games like Dota in both 1-versus-1 2017 and 5-versus-5 2018 formats These advances with neural networks at the core of all of them are more general than past systems and when trained properly can achieve unprecedented performance at one or more interesting tasks A system that can learn image recognition at record-setting levels could also learn to do the same in speech recognition The tools used to generate fake images of politicians could also be used to synthesize new artistic paintings or imaginary architectural plans Neural machine translation systems can learn to translate between any pair of languages - provided we have the training data The next step along the spectrum are future AI systems that can accomplish very hard valuable real-world tasks such as • • • • automatically devising and performing scientific experiments in chemistry or neuroscience helping us design better or cheaper drugs cars computer software and hardware and public infrastructure like transport or logistics systems performing surgery with more precision safety and efficiency than is possible for human surgeons to orchestrating the movements of thousands of self-driving cars and drones across a city or rural area whether to deliver ordinary goods or provide emergency supplies during an extreme weather event 5 Brundage Miles et al The Malicious Use of Artificial Intelligence Forecasting Prevention and Mitigation page 15 https arxiv org pdf 1802 07228 pdf 45 WhatisAGI AGI is even further along the spectrum of generality Our working definition of AGI is systems that are sufficiently advanced that they can outperform humans at most economically valuable work- which includes starting companies making business deals and writing books Such technology will need to be smart in a way unlike our traditionally literal uncreative computers The system's generality means it wouldn't be limited to commercial applications- it could also help with reasoning through complex international disputes city planning and even lawmaking or running countries Rather than being developed for any one use-case AGI would be developed for an entire spectrum of important tasks Why AGI could be developed sooner than commonly expected AI systems are built on three foundations algorithms data and compute or amount of computational resource Next generation AI systems being developed today are relying less on conventional datasets since they can either consume freely-available unlabeled data like a recent state-of-the-art language model we released which learns from an open dataset of books or can expend compute to generate data For example by simulating a robot we can create training data in quantities limited only by the number of computers available to run the simulation We recently released a study6 showing that the amount of compute powering the largest AI training runs has been doubling every 3 5 months for the past six years a total increase of 300 000x This growth is significantly faster than the historic driver of hardware progress Moore's Law which had an 18 month doubling period a 12x increase over the same period AI compute progress is driven partly by faster computers and partly by figuring out how to effectively train AI systems on many computers simultaneously This means that data and compute are rapidly becoming less significant bottlenecks on AI progress We expect this trend to continue We track over 45 hardware startups most in the US that are building next-generation AI computers Most are building on proven technologies that do not require further breakthroughs like quantum or optical As these computers hit the market and as we figure out how to use many such computers at once we expect the rate of breakthrough results to continue apace or even accelerate Our current computational paradigm allows for substantial increases in compute each year for at least the next five years Will the incoming tsunami of compute combined with near-term improvements in algorithmic understanding be enough to develop AGI or will we need to wait for some future algorithmic or hardware breakthrough We don't know the answer to this question yet but given today's rapid progress it seems unwise to be too confident in either direction- at least before uncovering further evidence 6 Amodei Dario et al AI and Compute https blog openai com ai-and-compute 46 The post-AGI future Investment in AI research is increasing rapidly due to how quickly AI advances can be deployed into products Transformative applications on the horizon like self-driving cars promise to save lives increase efficiency and generate huge value with the potential to add trillions of dollars to US GDP 7 • After AGI is created we expect economic and technological growth to accelerate markedly - with the new growth driven primarily by teams of creative computers partnering with creative humans We'll have the technological means to not just generate but also distribute essential resources and hopefully much more to ensure no one falls through the cracks and will be able to concentrate on efforts like education re-training and re-skilling to help people navigate the new economy The benefits will be vast and OpenAI believes those benefits should be equitably distributed rather than locked up with any one entity Technological progress has been accelerating rapidly for the past few hundred years and we expect the post-AGI world to add another jolt to the rate of progress We should expect advances in curing disease life extension transportation and space travel and communication Challenges of AGI development Each stage of AI development will bring its own challenges Narrow AI challenges are easiest to understand and act on because they apply to existing systems These are also the ones that today's corporations are most incentivized to fix These challenges include issues such as fairness transparency privacy and bias- all of which require serious attention if we want even more advanced AI technologies to have a positive impact We expect narrow AI progress to increase the rate of technological progress across the board further challenging today's policy machinery to keep pace as we approach AGI we expect the rate of change to increase further AGI challenges are harder to understand and foresee partly because they apply to systems that have not yet been developed OpenAI focuses on AGI challenges because we believe that they are simultaneously neglected and may happen sooner than is commonly believed The core danger with AGI is that it has the potential to cause rapid change This means we could end up in an undesirable environment before we have a chance to realize where we're even heading The exact way the post-AGI world will look is hard to predict- that world will likely be more different from today's world than today's is from the 1500s Some open questions 7 Lanctot Roger Accelerating the Future The Economic Impact of the Emerging Passenger Economy page 5 https newsroom intel com newsroom wp-content uploads sites 11 2017 05 passengereconomy pdf cid em-elg26916 utm source elq utm medium email utm campaign 26916 elg cid 1494219 47 What is the nature of international society in the post-AGI world We've already seen technology amplify the impact that states and small groups have in the world We expect AGI to further this trend What will people do with their time as economic work becomes an increasingly smaller part of one's life How can we help people live meaningful enjoyable lives in such a world How can AGI be deployed into our economic and social systems in a way that amplifies human preferences Will all humans remain meaningful actors in society Only humans in countries with powerful AGI deployments Only the humans who own a share in the technology itself How do we ensure that AGis operate in line with the values of their operators How do we avoid creating systems that cause social harm in blind pursuit of a poorly-specified goal- the technological version of what happened to Bear Stearns in 2008 AGI will affect every sector of global society and given the difficulty of these questions we don't have the luxury of waiting to see how AGI starts affecting society before addressing its challenges One example issue worth considering today is the possibility of a military arms race toward AGI A military arms race would put pressure on deploying an AGI without adequately verifying that it is safe AGI deployment will be challenging enough without pressure to gamble with safety Similar considerations also apply to pre-AGI AI technologies Safe and responsible AGI development Our views on safe and responsible AGI development are captured in three of the four sections of our Charter'l Broadly Distributed Benefits Long-Term Safety and Cooperative Orientation Safety We do not yet know how hard it will be to make sure AGis act according to the values of their operators Some people believe it will be easy some people believe it'll be unimaginably difficult but no one knows for sure- which is why OpenAI believes that safety research is critically important At the very least any AGI project should leave enough time to get safety right This includes taking steps- well in advance of the development of AGI -to avoid an uncoordinated race In this vein our Charter commits us to assisting rather than competing with a value-aligned safety-conscious project that comes close to building AGI before we do Broadly Distributed Benefits AGI will create unprecedented economic benefits If AGI can truly produce not just a Microsoft-sized amount of value but 100 Microsofts or more then returns beyond some point should not exclusively belong to a small group of people The rest of humanity will have assumed the risks of developing and deploying AGI and everyone deserves a fair share in the post-AGI future 8 OpenAI Charter https bloq openai com openai-charter 48 Cooperative Orientation AGI has the potential to be the most socially beneficial technology humans ever create The world is bigger than any one project and any society which successfully builds safe AGI will win collectively Thus it's important that value-aligned AGI projects view themselves in friendly competition Today we are all competing for talent and prestige But we need the ability to come together under one roof in some form before building such a powerful system bringing together companies and hopefully governments to ensure the resulting technology benefits everyone Policy recommendations 1 Measurement Many other established voices in the field have tried to combat panic about AGI by instead saying it not something to worry about or is unfathomably far off We recommend neither panic nor a lack of caution Instead we recommend investing more resources into understanding where the field is how quickly progress is accelerating and what roadblocks might lie ahead We're exploring this problem via our own research and support of initiatives like the AI Index But there's much work to be done and we are available to work with governments around the world to support their own measurement and assessment initiatives- for instance we participated in a GAOled study on AI last year 2 Foundation for international coordination AGI's impact like that of the Internet before it won't track national boundaries Successfully using AGI to make the world better for people while simultaneously preventing rogue actors from abusing it will require international coordination of some form Policymakers today should invest in creating the foundations for successful international coordination in AI and recognize that the more adversarial the climate in which AGI is created the less likely we are to achieve a good outcome We think the most practical place to start is actually with the measurement initiatives each government working on measurement will create teams of people who have a strong motivation to talk to their international counterparts to harmonize measurement schemes and develop global standards It's easy to imagine the post-AGI world as a destination- but it is more of an arbitrary marker denoting a world with transformative AI technologies There are many open questions around AGI and the more we can understand where the field is how fast we are moving and what is likely to happen in upcoming years the better prepared we will be to answer them And perhaps the most important question AGI raises is that once the world has been fundamentally transformed by systems that perform tasks we'd historically thought of as human - what then 49 Greg Brockman is co-founder and CTO of OpenAt a non-profit artificial intelligence research company working to ensure that artificial general intelligence benefits all of humanity He's an investor in over 30 startups and a board member of the Stellar digital currency system He was previously the CTO of Stripe a payments startup now valued at over $9B USD Greg studied mathematics at Harvard and computer science at MIT 50 Chairwoman COMSTOCK Thank you And we now recognize Dr Li TESTIMONY OF DR FEI-FEI LI CHAIRPERSON OF THE BOARD AND CO-FOUNDER AI4ALL Dr LI Thank you for the invitation Congresswomen and Congressmen My name is Fei-Fei Li I’m here today as the Cofounder and Chairperson of AI4ALL a national nonprofit organization focusing on bringing hands-on experience in AI research to high school students that have been traditionally underrepresented in the field of—in the STEM fields such as girls people of color and members of low-income communities Our program began at Stanford University in 2015 This year AI4ALL are expanded across North America to six university campuses I often like to share with my students that there’s nothing artificial about artificial intelligence It’s inspired by people it’s created by people and most importantly it has an impact on people It’s a powerful tool we’re only just beginning to understand and that’s a profound responsibility I’m here today because the time has come to have an informed public conversation about that responsibility With proper guidance AI will make life better but without it it stands to widen the wealth divide even further making technology even more exclusive and reinforce biases we’ve spent generations trying to overcome This will be an ethical philosophical and humanistic challenge and it will require a diverse community of contributors It’s an approach I call human-centered AI It’s made of three pillars that I believe will help ensure AI plays a positive role in the world The first is that the next generation of AI technology must reflect more of the qualities that make us human such as a deeper understanding of the context we rely on to make sense of the world Progress on this front will make AI much better at understanding our needs but will require a deeper relationship between AI and fields like neuroscience cognitive science and the behavior sciences The second is the emphasis on enhancing and augmenting human skills not replacing them Machines are unlikely to replace nurses and doctors for example but machine learning assistive diagnosis will help their job tremendously Similar opportunities to intelligently augment human capabilities abound from health care to education from manufacturing to agriculture Finally AI must be guided by a concern for its impact We must address challenges of machine biases security privacy as well as at the society level Now is the time to prepare for the effect of AI on laws ethics and even culture To put these ideas in practice governments academia and industry will have to work together This will require better understanding of AI in all three branches of government AI is simply too important to be owned by private interests alone and publicly funded research and education can provide a more transparent foundation for its development Next academia has a unique opportunity to elevate our understanding and development of this technology Universities are a 51 perfect environment for studying its effect on our world as well as supporting cross-disciplinary next-generation AI research Finally businesses must develop a better balance between their responsibility to shareholders and their obligations to their users Commercial AI products have the potential to change the world rapidly and the time has come to complement this ambition with ethical socially conscious policies Human-centered AI means keeping humans at the heart of this technology’s development Unfortunately lack of diverse representation remains a crisis in AI Women hold a fraction of high-tech positions even fewer at the executive level and this is even worse for people of color We have good reasons to worry about bias in our algorithms A lack of diversity among the people developing these algorithms will be among its primary causes One of my favorite quotes comes from technology ethicist Shannon Vallor who says that ‘‘There’s no independent machine values Machine values are human values ’’ However autonomous our technology becomes its impact on the world will always be our responsibility With the human-centered approach we can make sure it’s an impact we’ll be proud of Thank you The prepared statement of Dr Li follows 52 l4ALL AI4ALL 2148 Oakland contact@ai-4-all org House Science Committee Testimony Dr Fei-Fei Li Introduction As I often say to my students there's nothing artificial about AI It's inspired by people it's created by people and-most importantly-it impacts people It's a powerful tool we're only just beginning to understand and that's a profound responsibility I'd like to talk about what this technology looks like today the challenges we'll face in the coming years and what we can do now to ensure we're building a future worth Jiving in Artificial intelligence emerged in the mid 20th century as a quest to build machines with intellectual capabilities similar to those of the human mind As a science it draws on fields like cognitive science neuroscience statistics and mathematics As a technology it represents some of the most active developments in computer science and engineering including machine learning ML -a family of techniques that use statistical modeling to learn from data Today AI and ML are part of a vibrant interdisciplinary pursuit with fields like robotics natural language processing computer vision speech recognition and even philosophy playing ever-growing roles Ifs also changing the world Thank ' to the recent convergence of three key factors-powerful algorithms fast computing hardware and the era of big data-A has rapidly become a driving force in what some are calling the fourth industrial revolution It can help drive our cars assist with radiology diagn oses optimize energy consumption and even track deforestation from satellites All told AI spending is expected to rise irom 20 l7's $ 2B to $57 6B by 2021 according to International Data Corporation DC 53 We have good reason to be excited about AI But this is a nascent field and as a scientist that humbles me After all we've never created a technology to mimic human qualities so closely and we know little about the impact it will have on the world Guiding its development will be an ethical philosophical and humanistic challenge and it will require a diverse community of contributors I call this ' Human-Centered AI and it consists of three simple ideas Three Pillars of Human-Centered AI Pillar I AI must be more inspired by human intelligence Much of Al's recent success lies in a narrow range of problems associated with an abundance of annotated example data-such as accurately labeled photos or text-analyzed by algorithms in process called supervised learning But this approach has its limits For instance in my lab an image-captioning algorithm once fairly summarized a photo as 'a man riding a horse but failed to note the fact that both were statues Similar lapses are found in speech recognition natural language processing robotics and more Simply put AI technologies such as machine perception remains shallow compared to our own visual intelligence Algorithms lack our talent for extrapolating from few examples they tend to miss the big picture when given disorganized information and they're devoid of our remarkable capacity for context and emotion Researchers are working hard to address these shortcomings and progress will deliver tools that understand our needs with newfound depth But their work underscores the growing need for collaboration between AI and neuroscience cognitive science and the behavioral sciences Pillar II AI should strive to enhance us not replace us A recent McKinsey report found that balf of all current work activities can be theoretically automated by technology that already exists This is an alarming fact and there's no denying that job displacement will be a central challenge of the coming decade But what if we imagine AI as a tool for augmenting human capabilities rather than replacing them For example consider the problem of hospital-acquired infections Although they claim over 90 000 lives each yearmany times more than traffic fatalities-their cause is often simple human error In response my Stanford colleagues and 54 collaborators are partnering witb Stanford Children's Hospital to explore the use of smart sensors to track hand hygiene automatically with the goal of improving health outcomes and giving caregivers greater peace of mind So while a nurse or doctor's true skills are unlikely to be replaced by machines--complex problem solving in uncertain situations bedside manner and an emotional connection with patients-this research demonstrates how technology can enhance them Similar opportunities to intelligently augment human capabilities abound from healthcare to education manufacturing to search and rescue and city planning to entertainment It's true that automation is a challenge to certain forms of human labor but it's also an opportunity to elevate our notion of work Pillar III AI must be guided by a concern for its human impact The sheer power of AI means we have an obligation to consider its unintended consequences as thoughtfully as its desired effects This begins with the technology itself where fairness is among the most pressing concerns For example a recent study revealed that widely used facial recognition software recognizes white male faces with considerably higher accuracy than darker-skinned and or female faces Addressing such inequities will force us to confront a host of related problems One is the issue of interpretability or an AI system's capacity to explain its reasoning As they play a growing role in deliberations ranging from parole hearings to loan applications this ability will be critical And because AI relies on data to learn about the world tluough large quantities of images text video and personal records privacy and security will pose major challenges as well The picture only gets more complex at the societal level First some degree of labor disruption is inevitable and now is the time to prepare for its effects with proactive measures like retraining Next licensing and certification will have to expand to include technology that can perform medical legal and other highly regulated tasks without human involvement And our ethics not to mention the laws that codify them will have to evolve to answer questions of accountability as their role in our lives grows more consequential Society will soon face changes we can't even imagine and these predictions are only the beginning We can expect an exciting but uncertain future and a sense of humility will serve us well 55 What Can Be Done In Government Like nuclear technology biotech and energy security AI should be an active topic of discussion among policymakers This will require a far greater baseline of AI proficiency in all three branches of government however as well as ongoing dialogues with industry and academia This technology is simply too important to be owned by private interests alone and publicly-funded research can provide a transparent ethical foundation for its development However both public and private research will depend on a growing field of teclmologists with cutting-edge expertise in subjects like machine learning and data science To ensure we can meet demand it's vital we provide funding for education and diversity outreach in the STEM fields with an emphasis on AI starting as early as possible In Academia Academia has a unique opportunity to bring clarity and rigor to our understanding of AI The rare mix of intellectual resources found in our universities is a perfect environment for studying the effects of AI on our world as well as supporting the cross-disciplinary demands of the next generation of research Current campuses often separate computer science from other fields like the humanities and social sciences but the time has come for this practice to change Academic leaders can encourage discussions across departments in the fonn of classes seminars institutions symposiums and joint projects It's encouraging to note that academic venues are already playing a unique role in combating bias Organizations like FAT ML Fairness Accountability and Transparency in Machine Learning IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems the Partnership on AI and AI Now Institute are active sources of thought leadership in the ethical implementation of AI I hope to see this trend continue 56 In Industry Finally businesses must develop a better balance between their responsibility to shareholders and their obligations to their users community and society at large After all commercial products tend to change the world fastest So while the tech industry's aggressive investment in AI promises rapid advances the time has come to complement that ambition with ethical responsible policies I'm happy to say some companies are already taking the first steps with published guidelines intended to establish ethical boundaries and best practices for the field But this is just a start Across the tech industry and beyond much more can be done to promote the responsible development of AI including relationships with academia nonprofit' and governments to encourage the study of its human impact The Common Thread More Humanity Even after nearly two decades in this field I'm still heartened by the creativity intelligence and diversity of the people driving it I introduce high school students to AI research through AI4ALL a non-profit I co-founded advise doctoral students at Stanford and work with Google researchers worldwide I've learned that no age ethnicity or gender has a monopoly on innovation In fact more and more of the best work in our field comes from women people of color and other groups that have traditionally been underrepresented in STEM Unfortunately a lack of representation remains a crisis in AI A 2014 report revealed that women in tech feel stalled encounter persistent bias and are statistically more likely to leave their jobs As a result females hold only a third of all high-tech positions and even fewer at the executive level In fact only 13% of AI companies have female CEOs For people of color the numbers are even worse The tech world has a long way to go to address its lack of diversity starting with hiring practices Improvements are likely to yield immediate benefit as research 1 from McKinsey and Credit Suisse to the Journal of Personality and Social Psychology have quantified the benefits of diversity on workplace performance But studies also suggests this problem 1 https hbr org 20 16 11 why-diverse-teams-are-smarter 57 has far earlier origins girls for instance tend to lose interest in STEM curricula around age 15 usually due to a lack of peers and role models that look like them How should we expect the world's most powerful algorithms to behave if they remain ignorant to the diversity of human experience Organizations like Al4ALL were founded to reverse this trend Our summer program gives students from underrepresented communities first-hand experience in real AI labs at Stanford and UC Berkeley and we're expanding across North America this year 91% of our graduates feel confident they'll pursue a career in the field and 97% feel they're part of a community in AI and computer science In the words of one recent graduate used to think I wasn't smart enough to do computer science and AI But now I've gained so much confidence because of all the support and being around other girls who are into the same thing Conclusion As computational ethicist Shannon Vallor said there are no independent · machine values machine values are human values A human-centered approach to AI means these machines don't have to be our competitors but partners in securing our wellbeing However autonomous our technology becomes its impact on the world will always be our responsibility Now's the time to start making sure its an impact we can be proud of 58 Or Fei-Fei Li Biography Fei-Fei Li is a Professor in the Computer Science Department at Stanford and the Director of the Stanford Artificial Intelligence Lab In 2017 she also joined Google Cloud as Chief Scientist of AI and Machine Learning Dr U's main research areas are in machine learning deep learning computer vision and cognitive and computational neuroscience She has published approximately 200 scientific articles in top-tier journals and conferences including Nature PNAS Journal of Neuroscience New England Journal of Medicine CVPR CCV NIPS ECCV IJCV and IEEE-PAM Dr Li obtained her B A degree in physics from Princeton in 1999 with High Honors and her PhD degree in electrical engineering from California Institute of Technology Caltech in 2005 She joined Stanford in 2009 as an assistant professor and was promoted to associate professor with tenure in 2012 Prior to that she was on faculty at Princeton University 2007-2009 and University of Illinois Urbana-Champaign 2005-2006 Dr Li is the inventor of Image Net and the Image Net Challenge a critical large-scale dataset and benchmarking effort that has contributed to the latest developments in computer vision and deep learning in AI In addition to her technical contributions she is a national leading voice for advocating diversity in STEM and AI She is a co-founder of Stanford's renowned SAILORS outreach program for high school girls and the national non-profit AI4ALL For her work in AI Dr Li has been a keynote or invited speaker at many conferences including the World Economics Forum the Grace Hopper Conference in 2017 and TED2015 main conference She is a recipient of the 2017 Athena Academic Leadership Award APR 2016 J K Aggarwal Prize the 2016 NVIDIA Pioneer in AI Award 2014 IBM Faculty Fellow Award 2011 Alfred Sloan Facully Award 2012 Yahoo Labs FREP award 2009 NSF CAREER award the 2006 Microsoft Research New Faculty Fellowship and a number of Google Research awards Work from Dr Li's lab have been featured in a variety of popular press magazines and newspapers including New York Times Wall Street Journal Fortune Magazine Science Wired Magazine MIT Technology Review Financial Times and more She was selected as a 2017 Women in Tech by the ELLE Magazine a 2017 Awesome Women Award by Good Housekeeping a Global Thinker of 2015 by Foreign Policy and one of the Great Immigrants The Pride of America in 2016 by the Carnegie Foundation past winners include Albert Einstein Yo-Yo Ma Sergey Brin and more 59 Chairwoman COMSTOCK Thank you And I now recognize myself for five minutes for questions Dr Li there’s a generally accepted potential for AI-enabled teaching to a minimum to provide a backup for traditional classroom education So as AI technology advances it seems reasonable to assume that your traditional education vocational training homeschooling and even college coursework will need to change and adapt Could you maybe comment about how education in general and for specific groups and individuals might be transformed by AI and how we can make that positive and really sort of have more of a democratization of education particularly higher education and in STEM and in science Dr LI Thank you for the question Of course I feel passionate about education So I want to address this question in—from two dimensions One is how could we improve the education of AI and STEM in general to more students and general community Second is what can AI as a technology do to help education itself On the first dimension as evidenced by our work in the AI4ALL we really believe that it’s simultaneously a crisis and an important opportunity that we involve more people in the development of AI technology AI represents—humanity has never created a technology so similar or trying to resemble who we are and we need AI to—we need technologists and leaders of tomorrow to represent this technology So personally I think we need to democratize AI’s education to reach out to more students of color girls women from traditionally underrepresented minority groups At AI4ALL for the past four years we’ve already created an alumni population of more than 100 students and through their own community and route outreach effort we have touched the lives of more than 1 400 youth ranging from middle schoolers to high schoolers in disseminating this AI knowledge and we need more of that in higher education The second dimension that I want to answer your question is AI as a technology itself can help improve education itself In the machine learning community I’m sure Greg you also agree with me that there is an increasing recognition of the opportunity for lifelong learning using technology as an assistive technology I have colleagues at Stanford who focus on research and reinforcement learning and education of how to bring more technological assistance into the teaching and territorialization of education itself and I think this could become a huge tool as I was saying to augment human teachers and human educators to—so that our knowledge can reach to more students and wider community Chairwoman COMSTOCK Excellent And for other witnesses could you maybe comment on how academic institutions and industry could work with government on AI Mr BROCKMAN All right So you know for—OpenAI’s recommendation is really about starting with measurement right to really start to understand what’s happening in the field I think it’s really about for example the study that we did showing the 300 000 times increase We need more of that We need to understand where things are going where we are I think the government is uniquely positioned to set some of the goalposts as well and we’ve been pretty encouraged by seeing some of the work that 60 is happening at GAO and also DIUx has had some success with us So we think it’s really about starting a low touch way for the dialogue to start happening because I think right now the dialogue is not happening to the extent that it should Dr PERSONS All right Thank you for the question I do think that as the committees have pointed out this is a whole-of-society issue It’s going to be government in partnership with the private sector with academia to look at things So I think there is room for thought about how to learn by doing creating internships and ways to try and solve real-world problems so that you have a mix of the classroom experience as well as making and building—you’ll fail a lot of course with these things but learning in a safe environment and then being able to grow expertise in that way Chairwoman COMSTOCK Thank you And Dr Li did you have anything you wanted to add to that also Okay Well thank you And I now will recognize Mr Lipinski for five minutes Mr LIPINSKI Thank you This is a fascinating topic And there’s—I want to try to move through some things quickly but I’ll get some good answers here It seems to me that Mr Brockman you have a different view of how—of AGI the possibilities of AGI and how quickly it can come then the GAO report Is there a reason for this Is there something you think that GAO is missing And if Dr Persons could respond to that Mr BROCKMAN So I don’t know if I can comment directly on the report just not being familiar enough with all the details in there but I can certainly comment on our perspective on AGI and its possibility And a lot of it really comes down to rather than—you know I think that there’s been a lot of more emotion or intuitionbased argument And to your opening remarks you know I think that science-based reasoning in order to project what’s happening in this field is extremely important and that something that we’ve spent quite a lot of effort on since starting OpenAI almost three years ago And so looking at the barriers to progress as compute data algorithms data is something that’s changing very rapidly in terms of what data we can use the computation the power there is growing at a rate that we’ve just never seen over the course of this decade We’re going to be talking you know I think about ten orders of magnitude and that something where if you were to compare that to the typical growth of compute something like Moore’s law that the—over the period where we saw 300 000X increase in the past six years we would’ve only seen 12X right That’s a huge gap and this is somewhere where we’re sort of being projected into the future a lot faster than people realize Now it doesn’t mean that it’s going to happen soon It means that we can’t rule it out It means that for the next five years as long as this hardware growth is happening we’re in a fog and it’s hard to make confident projections And so my position is that we can’t rule it out We know that this is—you know we’re talking about a technological revolution on the scale of the agricultural revolution something that could be so beneficial to everyone in this world And if we aren’t careful in terms of thinking ahead and trying to be prepared it really could be caught unaware 61 Mr LIPINSKI And thank you Dr Persons do you have a response on that Dr PERSONS Sure I think—and with all respect for our Silicon Valley innovators who are upstarts and challenge the status quo I think it’s great that we have this system The key thing that we’re seeing is the convergence of these technologies that was mentioned by my panelists of the exponential power in computing the ubiquitous nature of data the sophistication of algorithms are all coming in But that said many folks in the community are mildly skeptical about the rate at which general AI may come in this area because—for several reasons One is just the way that we think about the problem now the super complexity that is manifest in addressing the various challenges You’re looking at large data sets and looking at all the facets of them It’s much easier to say than to do And again I think a lot of the—as you pointed out the driving force here is the concern about general AI and taking over the world kind of thing and it’s just much harder to mimic human intelligence especially in an environment where intelligence isn’t even really defined or understood And I think as Dr Li pointed out a lot of this is really about augmentation That’s something else we heard from our experts It wasn’t a replacement of humans it was a how can we become better humans more functional humans in doing these things So a lot of it just gets down to the—— Mr LIPINSKI Let me—because I have a short time sorry Dr PERSONS Thank you Mr LIPINSKI I just want to throw out quickly the—there have been very different—vastly different opinions and—about the replacement of jobs and the disappearance of jobs and what the impact’s going to be Mr Brockman what do you think the impact will be Mr BROCKMAN So I think that with new technologies in the short term we always overestimate the degree to which they can make rapid change but I think in the long-term that they do I think technology is change in that we’ve seen with things like the internet that there’s been a lot of job displacement both creation and destruction And I think AI will be no different I think the question of exactly which jobs and when I think we don’t have enough information yet and I think that that’s where measurement really starts to come in So we view it as an open question and a very important one Dr PERSONS And sir if I can just say as a bottom line nobody really knows the impact on this and of course our experts are saying to know more we might need to be able to encourage—let’s— for example our Bureau of Labor Statistics a data-type agency that—out of the federal government to help provide more data or different data or things to help try and answer the question of what is the impact as this technology continues to unfurl That said there’s also a history of—when you—it goes back and attributed to Ned Ludd in the era of British industrialization and the concern of destroying the machines for the concern about loss of jobs and yet—and many times throughout history it’s happened 62 in an array of technologies where net jobs actually increased It just—they were more sophisticated jobs They were toward higher value creation and more productivity So there is hope with this technology as well Mr LIPINSKI And if the Chairwoman will allow I want to hear from Dr Li Dr LI I just want to say that technology inevitably throughout human civilization has an impact to change the landscape of jobs but it’s really really critical like my fellow panelists said that we need to invest in the research of how to assess this change It’s not a simple picture of replacement especially when this technology has a much greater potential and power to augment it I just spent days in the hospital ICU with my mother in the past couple of weeks and—with my own health care and AI research you recognize that a nurse in a single shift is doing hundreds of different tasks in our ICU unit where they’re fighting for life and death for our patients and these are not a simple question of replacing jobs but creating better technology to assist them and to make their jobs better and make the lives better for everyone And that’s what I hope we focus on using this technology Mr LIPINSKI Thank you Chairwoman COMSTOCK Thank you Dr Li That’s a wonderful example of really vividly explaining to us how that can be used because certainly as we’re an aging population in this country that’s a challenge we’re all facing And so the quality of life and improvement in each of those employees and nurses being able to do a better job thank you for outlining that I now recognize Mr Weber Mr WEBER Thank you Madam Chair Dr Li is your mom okay We hope that she is and pray and hope that she is okay Dr LI Thank you I’m here That means she’s better Mr WEBER Okay Otherwise we were going to be missing two witnesses Good Dr LI She’s watching me right now Mr WEBER Well good Chairwoman COMSTOCK Hi Mom She’s doing a great job Mr WEBER You’re doing excellent She’s a proud mom and that’s some good medicine in and of itself right there Dr LI Thank you Mr WEBER So we’re glad for that Dr Brockman you in your statement say that your mission was to actually make sure that artificial intelligence benefited people and was better for the most economically valuable work Do you remember that Mr BROCKMAN So are—— Mr WEBER It’s in your written statement Mr BROCKMAN That’s right So the—our definition of what AGI will be whether created by us or anyone else but just the milestone is a system that can outperform humans and economically valuable work Mr WEBER Okay Well let me read it to you real quick ‘‘I’m Greg Brockman Cofounder a nonprofit development organization Our mission is to ensure that artificial general intelligence by 63 which we mean highly autonomous systems that outperform humans at ’’ quote ‘‘most economically valuable work ’’ end quote ‘‘benefits all humanity ’’ How would you define most economically valuable work Mr BROCKMAN So I think that again—and first of all I just— you know the question of—you know AGI is something that the whole field has been working towards for—you know really since the beginning of the field 50 years ago and so the question of how to define it I think is something that is not entirely agreed-upon that our definition is this—and when we think of it we think of— you think about things like starting companies or very high intellectual work like that—— Mr WEBER Right Mr BROCKMAN —and you know also to things like going in cleaning up disaster sites or things that humans would be unable to do very well today Mr WEBER Okay Well I noticed that in your disagreement that Congressman Lipinski referred to with the report and you call them Silicon Valley upstarts At least you didn’t call them young upstarts so that’s an advantage Thank you for doing that But you’re literally looking at a new industry that even though it’s shifted—bless you—even though the shift is going to be changing you’re actually creating jobs for another industry And going back to Dr Li’s example with her mom in the IU talking about much the nurses do how do you train for those jobs if it’s moving as fast as you think it is Mr BROCKMAN Yes and so you know one thing I think is also very important is that I don’t think we have much ability to change the timeline to this technology I think that there are a lot of stakeholders there are a lot of different pieces of the ecosystem And that—what we do is we step back and we look at the trends and we say what’s going to be possible when And so I think that the question of how to train—and again that’s going to be something— we’re not the only ones that are going to have to help answer that question But I think that the place to start it really comes back to measurement right If we don’t know what’s coming if we can’t project well then we’re going to be taken by surprise And so you know I think that there are going to be lots of jobs and already have been created jobs that are surprising in terms of—you think about with autonomous vehicles that we need to label all this data we need to make sure that the systems are doing what we expect and that all of that—that there’s going to be humans that are going to help make these systems—— Mr WEBER But we would all agree I hope—and this question is for all three panelists—all three witnesses that the jobs they’re going to create are well worth the transformation into all of that technology Dr Persons would you agree with that Dr PERSONS I would agree to that I’ll—let me give you a quick example if I may Speaking with a former Secretary of Transportation recently just a simple example of tollbooth collectors we have now a system where you get the E–ZPass you drive through and you have less of a workforce there that could have had an im- 64 pact at that time for short period on the number or loss of jobs for tollbooth collectors and yet it freed them up It enabled them to perhaps do other things that were needed and large problems Mr WEBER Okay And Mr Brockman you were shaking your head You would agree with that statement Mr BROCKMAN Absolutely I think that the purpose of technology and improving—— Mr WEBER Sure Mr BROCKMAN —it is to improve people’s lives Mr WEBER So Dr Li I see you shaking your head too Dr LI Yes absolutely In addition to the example Mr Persons provided I think deeply about the jobs that are currently dangerous and harmful for humans from fighting fires to search and rescue to you know natural disaster recovery Not only we shouldn’t put humans in harm’s way if we can avoid it but also we don’t have enough help in these situations and they are—this is where technology should be of tremendous help Mr WEBER Very quickly I’m out of time just yes or no If we lose dominance in AI that puts us in a really bad spot in worldwide competitiveness would you agree Dr PERSONS Yes Mr BROCKMAN Yes Mr WEBER Yes Thank you Dr LI Yes Mr WEBER Madam Chair I yield back Chairwoman COMSTOCK Thank you Good question Now I recognize Mr Veasey for five minutes Mr VEASEY Thank you Madam Chair We have heard about already from your testimony some of the advantages of AI and how it can help humankind how it can help advance us as a nation and a country But as you know there are people also that have concerns about AI There’s been a lot of sort of doomsday-like comparisons about AI and what the future of AI can actually mean To what extent do you think this scenario this sort of you know worst-case scenario that a lot of people have pointed out about AI is actually something that we should be concerned about And if there is a legitimate concern what can we do to help establish a more ethical you know responsible way to develop AI And this is for anybody on the panel to answer Mr BROCKMAN So I think thinking about artificial general intelligence today is a little bit like thinking about the internet in maybe the late ’50s right If someone was to describe to you what the internet was going to be how it would affect the world and the fact that all these weird things were going to start happening you’re going to have this thing called Uber which you’re going to be able to—you’d just—you’d be very confused It’d be very hard to understand what that would look like and the fact that oh we forget to put security in there and that we’d be paying for that for you know 30 years’ worth of trying to fix things And now imagine that that whole story which played out over really the course of the past 60 almost 70 years now was going to play out in a much more compressed timescale 65 And so that’s the perspective that I have when it comes to artificial general intelligence is the fact that it can cause this rapid change and it’s already hard for us to cope with the changes that technology brings And so the question of is it going to be malicious actors is it going to be that the technology itself just wasn’t built in a safe way or is it just that the deployment that who owns it and the values that it’s given aren’t something that we’re all very happy with All of those I think are real risks and again that’s something that we want to start thinking about today Dr PERSONS Thank you sir So I agree with that I think the key thing is being clear-eyed about what the risks actually are and not necessarily being driven by the entertaining and yet this science-fiction-type narrative sometimes on these things projecting or going to extremes and assuming far more than where we actually are in the technology So it’s—there are risks It’s understanding the risks as they are and there are always contextual risks Risks in automated vehicles are going to be different than risks in this technology in financial services let’s say So it’s really working again symbiotically with the community of practice and identify what are the things there What are the opportunities And there’s going to be opportunities And then what undesirable things do we want to focus on and then optimize from there on on how to deal with them Thank you Mr VEASEY Mr Brockman in your testimony you had referenced a report outlining some malicious actors in this area Could you sort of elaborate on some of your findings in these areas Mr BROCKMAN That’s right So OpenAI was a collaborator on this research report projecting not necessarily today what people are doing but looking forward what are some of the malicious activities that people could use AI for And so that report—let’s see Yes I think maybe the most important things here you start thinking about a lot of things around information privacy the question of how we actually ensure that these systems do what the operator intends despite potential hacking You think about autonomous systems that are taking action on behalf of humans that are subverted and whether again it’s—you know that this report focuses on active action You think about autonomous vehicles and if a human hacker can go and take control of a fleet of those some of the bad things that could happen And so you know I think that this report should really be viewed as we need to be thinking about these things today before these are a problem because a lot of these systems are going to be deployed in a large-scale way and if you’re able to subvert them then you know the—all of the problems that we’ve seen to date are going to start having a very different flavor where it’s not just privacy anymore it’s also systems that are deployed in the real world that are actually able to affect our own well-being Mr VEASEY Thank you Madam Chair I yield back Chairwoman COMSTOCK Thank you And I now recognize Mr Rohrabacher Mr ROHRABACHER Thank you very much Madam Chairman This as in all advances in technology can be seen as the great hope for making things better or the new idea that there might be new dangers involved or the new technologies will help certain 66 peoples but be very damaging to others And I think that where that fear would be most recognizable is in terms of employment and how in a free society people earn a living And are we talking about here about the development of technology that will help get the tedious and remedial or the lower-skilled jobs that can be done by machine or are we talking about the loss of employment by machines that are designed to really perform better than human beings perform in high-level jobs What are we talking about here Dr LI Okay So I can—I’m still going to use health care as an example because I’m familiar with that area of research So if you look at recent studies by McKinsey and other institutions on employment and AI there is a recognition that we need to talk a little more nuanced than just entire job but the tasks under each job The technology has the potential to change the nature of different tasks Again for example take nurse—a job of a nurse as an example It—no matter how rapidly we develop the technology and the most optimistic assessment it’s very hard to imagine the entire profession of nurse—nursing would be replaced yet within the nursing jobs there are many opportunities that certain tasks can be assisted by AI technology For example a simple one that costs a lot of time and effort in nursing jobs is charting Our nurses in our again ICU rooms our patient rooms spend a lot of time typing and charting into a system into a computer while that’s time away from patients and other more critical care So these are the kind of tasks under a bigger job description that we can hope to use technology to assist them and augment—— Mr ROHRABACHER So are we talking about robots here or a box that thinks and is able to make decisions for us What are we talking about Dr LI So AI technology is a technology of many different aspects It’s not just robot In this particular case for example natural language understanding the speech recognition and possibly in the form of a voice assistant would help charting But maybe delivering of simple tools on the factory floor will be in the form of a small simple delivery robot So there are different forms of machines Mr ROHRABACHER I see Well there are many dangerous jobs that I could see that we’d prefer not having human life put at risk in order to accomplish the goal And for example at nuclear power plants it would be a wondrous thing to have a robotic response to something that could cause great damage to the overall community but would kill somebody if they actually went in to try to solve a problem And I understand that and also possibly with communicable diseases where people need to be treated but you’re putting people at great risk for doing that However with that said when people are seeking profit in a free and open society I would hate to think that we’re putting out of work people with lower skills and we need the dignity of work and of earning your own way once we know now that when you take that away it really has a major negative impact on people’s lives So I want to thank you all for giving us a better understanding of what we’re facing on this and let’s hope that we can develop this technology in a way that helps the widest variety of people and not 67 just perhaps a small group that will keep their jobs and keep the money So thank you very much Chairwoman COMSTOCK Thank you And I now recognize Ms Bonamici for five minutes Ms BONAMICI Thank you so much Thank you to our witnesses First I want to note that our nation has some of the best scientists and researchers and engineers in the world but without stronger investments in research and development especially longterm foundational research we risk falling behind especially in this important area I hope the research continues to acknowledge the socioeconomic aspects as well of integrating AI technologies In my home State at the University of Oregon we have the Urbanism Next center They’re doing some great work bringing together interdisciplinary perspectives including planning and architecture and engineering and urban design and public administration with public private and academic sectors to discuss how leveraging technology will shape the future of our communities Their research has been talking about emerging technologies like autonomous vehicles and the implications for equity health the economy and the environment and governance Dr Persons can you discuss the value of establishing this type of partnership between industry academia and the private sector to help especially identify and address some of the consequences intended and unintended of AI as it becomes more prevalent And I do have a couple more questions Dr PERSONS Sure I’ll answer quickly The short answer is yes Our experts and what we’re seeing is the value in public-private partnerships because again it would be a mistake to look at this technology in sort of isolated stovepipes and it would need to be an integrated approach to things the federal government having its various roles but key—like your mentioning at University of Oregon key academic and research questions There’s many many things to research and questions to answer and then of course industry which has an incredible amount of innovation and thinking and power to drive things forward Ms BONAMICI Terrific Thank you Dr Li I have a couple questions You discuss the labor disruption and I know that’s brought up a couple of times and the need for retraining We really have sort of a dual skills gap issue here because we want to make sure there are enough people who have the education needed for the AI industries but we also are talking about the workers like you mentioned the workers in tollbooths who will be displaced But with the rapid development of technologies and the changes in this field what knowledge and skills are the most important for a workforce capable of addressing the opportunities and the barriers to the development I serve on the Education and Workforce Committee and this is a really important issue is how do we educate people to be prepared for such rapid changes Dr LI So AI is fundamentally a scientific and engineering discipline and to—as an educator I really believe in more investment in STEM education from early age on We look at—in our experience at AI4ALL when we invited these high school students in the age of 14 15 16 to participate in AI research their capabilities 68 and potential just amazes me We have high school students who have worked in my lab and won best-paper award at this country’s best AI academic conferences And so I believe passionately that STEM education is critical for the future for preparing AI workforce Ms BONAMICI Thank you And as everyone on this committee knows I always talk about STEAM because I’m a big believer in educating both halves of the brain and students who have arts education tend to be more creative and innovative Also Dr Li in your testimony you talk about how AI engineers need to work with neuroscientists and cognitive scientists to help AI systems develop a more human feel Now I know Dr Carbonell is not here today but I noted in his testimony he wrote ‘‘AI is the ability to create machines who perform tasks normally associated with human intelligence ’’ I’m sure that was an intentional choice to humanize the machine but I wanted to ask you Dr Li about— he’s not here to explain but I have no doubt that was intentional In your testimony you talk about the laws that codify ethics How is this going to be done Can you go into more depth about how would these laws be done Who would determine what is ethical And would it be a combination of industry government determining standards How is—how are we going to set the stage for an ethical development of AI Dr LI Yes so thank you for the question I think for technology as impactful as AI is to human society it’s critical that we have ethical guidelines And different institutions from government to academia to industry will have to participate in this dialogue together and also by themselves Ms BONAMICI Are they already doing that though You said they’ll have to but is somebody convening all of this to make sure that there are—— Dr LI So there are efforts I’m sure Greg can add to this Industries in Silicon Valley we’re seeing company starting to roll out AI ethical principles and responsible AI practices in academia We see that ethicists and social scientists coming together with technologists holding seminars symposiums classes to discuss the ethical impact of AI And hopefully government will participate in this and support and invest and these kind of efforts Ms BONAMICI Thank you I see my time is expired Thank you Madam Chair I yield back Oh Mr Chairman thank you Mr WEBER Presiding I thank the gentlelady And the gentlelady from Arizona is recognized for five minutes Mrs LESKO Thank you Mr Chair I want to thank the testifiers today very interesting subject and something that kind of spurs the imagination about science fiction shows and those type of things What countries are the major players in AI and where does the United States rank in competition with them And that’s to any panelist or all panelists Mr BROCKMAN So you know today I think that the United States actually ranks possibly top of the list You know I think there are lots of other countries that are investing very heavily You know China is investing heavily lots of countries in Europe are investing heavily The—you know DeepMind is subsidiary of 69 a U S company but located in London And I think that you know it’s very clear that AI is going to be something of global impact and I just think the more that we can understand what’s happening everywhere and figure out how we can coordinate on safety and ethics in particular the better it’s going to go Dr PERSONS Yes I—thank you for the question I think wherever there is large amounts of computing large amounts of data and a strong desire to innovate and continue to develop again in this sort of fourth industrial revolution that we’re moving on then you—it drives toward certainly China and then our allies and colleagues in Western Europe and developed worlds Thank you Mrs LESKO And is there—did you want to answer Mr BROCKMAN Sorry—— Mrs LESKO Go ahead Mr BROCKMAN If I could just add that you know the most important thing to continue to lead in the field it’s really about the talent And right now we’re doing a great job of bringing all the talent in At OpenAI we have a very wide mix of national backgrounds and origins and I think as long as we can keep that up that we’ll be in very good shape Mrs LESKO Thank you And Mr Chair I have one more question and I think this has been asked in different ways before but what steps are we guarding against espionage from let’s say—you said China is involved in this and that’s basically my question— espionage hacking those type of things What guidelines are currently taking place and who’s preventing this Is it the private companies themselves Is government involved Thank you Mr BROCKMAN So one thing that’s a very atypical about this field is because it really grew out of an academic—very small number of academic labs that the overarching ethos in the field is actually to publish And so all of the core research and development is actually being shared pretty widely And so I think that as we’re starting to build these more powerful systems and this is one of the parts of our charter that we need to start thinking about safety and keeping—you know thinking about things that should not be shared and so I think that this is a new muscle that’s being built It’s right now kind of up to each company and I think that that something that we’re all starting to develop But I think having a dialogue around what’s okay to share and what things are kind of too powerful and should be kept private that’s just the dialogue that’s starting now Dr PERSONS And certainly IP or intellectual property protection is a critical issue I think of one former Director of the National Security Agency mentioned that we’re—at the time it was unprecedented theft of U S intellectual property at that time just because of the—it’s the blessing and curse of the internet It’s a blessing it’s this open and the curse is it’s open And so AI is going to I think be in that category In terms of what’s being done in terms of cybersecurity it is something our experts pointed out and said it is an issue As this Committee well knows it’s easier said than done and who has jurisdictions in the U S federalist system about particularly a private company and protection of that the role of the federal government versus the company itself in an era where as I think Mr 70 Brockman has pointed out is sort of the big-data era where data are the new oil yet we want to be open at the same time so that we can innovate So managing that dialectical tension is going to be a critical issue and there’s no easy answer Mrs LESKO Thank you Mr Chair I yield back Mr WEBER The Chair recognizes Ms Esty for five minutes Ms ESTY Thank you Mr Chair and I want to thank the witnesses for this extremely informative and important conversation that we’re having here today I hail from the State of Connecticut where we see a lot of innovation at UCONN at Yale at lots of spinoffs on the sort of narrow AI question But I think for us really the issue is more about that general AI And Mr Brockman your discussion of the advances which makes Moore’s law look puny in comparison is really where I want to take this conversation about Dr Li your discussion which I think is incredibly important about diversity We saw what happened to Lehman Brothers by not being diverse I am extremely concerned about what the implications are for teaching a— as it were if it’s garbage in it’s going to be garbage out If it’s a very narrow set of parameters and thought patterns and life experiences that go into AI we will get very narrow results out So first I want to just talk—get your thoughts on that And the second is on this broader ethical question We’ve looked for many years—I remember back when I was a young lawyer working on bioethical issues The Hastings Center got created to begin to look at these issues This Committee has been grappling with CRISPR and the implications with CRISPR I think about this being very similar that AI has many similar implications for ethical input So if you can opine on both of those questions and recognize we have got two minutes—three minutes left—about both the ethical— whether we need centers to really bring in ethicists as well as technologists and then the importance of diversity on the technology side so that we get the full range of human experience represented as we’re exciting—our exciting new entry into this fourth revolution Thanks Dr LI Yes in fact when—just now—thank you for asking that question Just now when somebody is using the term doomsday scenario to me I think if we wake up 20 years from now whatever years and we see the lack of diversity in our technology and leaders and practitioners that would be my doomsday scenario So it’s so important and critical to have diversity for the following three reasons like you mentioned One is shared jobs that we’re talking about This is a technology that could have potential to create jobs and improve quality of life and we need all talents to participate in that Second is innovation and creativity like you mentioned in Connecticut and other places We need that kind of broad talent to add in the force of AI development And the third is really justice and moral values that if we do not have this wide representation of humanity representing the development of this technology we could have face-recognition algorithms that are more accurate in recognizing a male—white male faces And these are—we could have dangers of out—biased algo- 71 rithms making unfair loan application decisions You know there are many potential pitfalls of a technology that’s biased and not diverse enough Which brings us to this conversation and dialogue of ethics and ethical AI You’re right Previous disciplines like nuclear physics like biology have shown us the importance of this I don’t know if there is a single recipe but I think the need for centers institutions boards and government committees are all potential ways to create and open this dialogue And again we’re starting to see that but I think you’re totally right It’s critical issues Ms ESTY Mr Brockman Mr BROCKMAN If I may so I agree completely with my fellow witness So diversity is crucial to success here So actually—so we have a program called OpenAI scholars where we brought in a number of people from underrepresented backgrounds into the field and provided mentorship and they’re working on projects and spinning up One thing that we found that I think is very encouraging is it’s actually very easy to take people who do not have any AI or machine learning background and to make them into extremely productive first-class researchers and engineers very quickly And that’s you know one benefit of this technology being so new and nascent is that in some ways it—we’re all discovering as we go along too so becoming an expert there just isn’t that high of a bar But—so I think that everyone putting effort in to places where the expertise is I think it’s on them to make sure that they’re also bringing in the rest of the world On the ethical front that’s really core to my organization That’s the reason we exist that we do think that you know for example when it comes to the benefits of who owns this technology who gets it—you know where did the dollars go we think it belongs to everyone And so one of the reasons I’m here is because I think that this shouldn’t be a decision that’s made just in Silicon Valley I don’t think that the question of the ethics and how this is going to work should be in the hands solely of people like me I think that it’s really important to have a dialogue and again that something where you know I hope that that will be one of the outcomes of this hearing Ms ESTY Thank you very much Mr WEBER The gentleman now recognizes Mr McNerney Mr MCNERNEY Well I thank the Chair for holding this and the Ranking Member and I thank the witnesses really very interesting testimony and diverse in its own right One of the things that I think that’s important here is—with this committee is how does the government react to AI Do we need to create a specific agency Does that agency report to Congress or to the Administration Those sorts of things I think are very important Dr Brockman you said—I think one of the most important things was that we need a measure of AI progress Do you have a model or some description of what that would look like Mr BROCKMAN Yes I do Thank you for the question And so first of all I don’t think that we need to create new agencies for this I think that existing agencies are well set up for this And I 72 was actually very encouraged to hear that people are talking about giving NIST a remit to think about these problems Again GAO and DIUx are already starting to work on this For example DIUx had a satellite imagery data set hosted a public competition The kind of thing that we think would be great for government to do as well is to have standardized environments where academics and private sector can test robotic approaches setting up competitions towards specific problems that various agencies and departments want to be solved All of those I think can be done without any new agency and I think that that’s something that you can both get benefits directly to the relevant agencies also understand the field and also start to build ties between private sector and public sector Mr MCNERNEY I’m one of the founders of the Grid Innovation Caucus What are the most likely areas we’ll see positive benefits to the grid to electric grid stability and resiliency Who would be the best to answer that Mr Persons Dr PERSONS Sure Thank you for the question I think one of the ways—GAO has done a good deal of work on this issue but it’s just protection of the electrical grid in the cybersecurity dimension So as one of our scenarios or profiles that we did in this report what our experts and what folks are saying and again based on the leadership of this Committee and the importance of cyber is that it’s a—without which nothing—AI is going to be a part of cyber moving forward and so protection of the grid in the cyber dimension is there Also I think as the Chairman mentioned earlier the word optimization so how we optimize things and how algorithms might be able to compute and find optimums faster and better than humans is an opportunity for grid management and production Thank you Mr MCNERNEY So AI is also going to be used as a cyber weapon against infrastructures or potentially used as a weapon is that right Dr PERSONS There are concerns now when you look at a broad definition of AI and you look at bots now that are attacking networks and doing distributed—what are—DDOS or distributed denial of service attacks and things like that that exists now You could—unfortunately in the black hat assumption you’re going to assume that as AI becomes more sophisticated the white hats and so too unfortunately the black hat side of things the bad guys are going to also become more sophisticated And so it’s going—that’s going to be the cat-and-mouse game I think moving forward Mr MCNERNEY Another question for you Dr Persons In your testimony you mentioned that there’s considerable uncertainty in the jobs impact of AI Dr PERSONS Yes Mr MCNERNEY What would you do to improve that situation Dr PERSONS Our experts were encouraging specific data collected on this Again we have important federal agencies like BLS Bureau of Labor Statistics that work on these issues what’s going on in the labor market for example and it may just be an update to what we collect what questions we ask as a government how we provide that data which of course is very important to our understanding of unemployment metrics and so on 73 So there are economists that have thoughts about this We had some input on that There’s no easy answer at this time but the idea that there is an existing agency doing that sort of thing is there The key question is how could we ask more or better questions on this particular issue on artificial systems Mr MCNERNEY Thank you Dr Li you gave three conditions for progress in AI being positive Do you see any acceptance or general wide acceptance of those conditions How can we spread the word of that so that the industry is aware of them and the government is aware of them and that they follow those sorts of guidelines Dr LI Thank you for asking Yes I would love to spread the word So I think I do see that—the emergence of efforts in all three conditions The first one is about more interdisciplinary approach to AI and ranging from universities to industry we see the recognition of neuroscience cognitive science to cross pollinate with AI research I want to add we’re all very excited by this technology but as a scientist I’m very humbled by how nascent the science is It’s only a science of 60 years old compared to traditionally classic science that’s making human lives better every day physics chemistry biology There’s a long long way to go for AI to realize its full potential to help people So that recognition really is important and we need to get more research and cross-disciplinary research into that Secondly is the augmenting human and again a lot of academic research as well as industry startup efforts are looking at assistive technology from disability to you know helping humans And the third is what many of us focus on today is the social impact from studying it to having a dialogue to having—to working together through different industry and government agencies So all three are the elements of human-centered AI approach and I see that happening more and more Mr MCNERNEY Thank you I yield back Mr WEBER The Chair now recognizes the gentleman from New York No The Chair now recognizes the gentleman that’s not from New York Mr Palmer Mr PALMER Thank you Mr Chairman I’d like to know if AI can help people who are geography-challenged Mr WEBER The gentleman’s time has expired Mr PALMER I request that that question and response be removed from the record I do have some questions In my district we have the National Computer Forensics Institute which deals with cybercrime and what I’m wondering about is with the emergence of and evolution of AI What are we putting in place because of the potential that that creates for committing crime and for solving crime Dr Persons do you have any thoughts on that Dr PERSONS Well certainly in one of the areas we did—thank you for the question One of the areas we did look at in general was just criminal justice So I mean just the risks that are there in terms of the social risks making sure the scales are balanced exactly as they ought to be that justice is blind and so on was the focus of that 74 However I think in terms of the criminal forensics AI could be a tool that helps suss out what—you know in a retrospective sense what happened But again it’s an augmentation that’s helping the forensic analyst who would know what things look like And the algorithm would need—in the machine-learning sense of things would need to learn what the risks might be going forward so that you perhaps could identify things more proactively and perhaps in near or at real-time So that’s the opportunity for this Again AI is a tool and cyber was a key message we heard moving forward Mr PALMER Any thoughts on that Mr BROCKMAN So today you know we’re already starting to see some of the security problems with the methods that we’re creating for example that there’s a new class of attack called adversarial examples where researchers are able to craft like a physical patch that you could print out and put on any object They’ll make a computer vision system think that it’s whatever object you want it to be so you could put that on a stop sign and confuse a selfdriving car for example So these sorts of ways of subverting these powerful systems is something we’re going to have to solve and going to have to work on just like we’ve been working on computer security for more conventional systems And I think that the way to think about if you could successfully build and deploy an AGI what that would look like in many ways it’s kind of like the internet in terms of being very deeply integrated in people’s lives but also having this increasing amount of autonomy and representation and taking action on people’s behalf And so you’ll have kind of this question of how do you make sure you know first of all that’s something that could be great for security if these systems are well-built and have safety in their core and are very hard to subvert But also if it’s possible for people to hack them or to cause them to do things that are not aligned with the value of the operator then I think that you can start having very large-scale disruption Mr PALMER It also concerns me in the context of—it was announced a couple of weeks ago that the United States plans to form a space corps We know that China has been very aggressive in militarizing space If you have any thoughts on that discussion of how artificial intelligence will be used in regard to space Communication systems that are highly vulnerable already I think that there’s some additional vulnerability that would be created Any thoughts on that And any one of the three of the panelists Dr PERSONS Yes sir So in terms of the risk in space obviously one of the key concerns for AI is weaponization and—which I think is part of that and so much less the space domain or any other one And so I know our Defense Department has key leadership thinking on this and working strategically on how do we operate in an environment where we have to assume there’s going to be the adversary that might not operate in the ethical framework that we do and to defeat that but there’s no simple answer at this time other than our Defense Department is thinking about it and working on it Mr PALMER And he’s not here obviously to testify but Dr Carbonell’s testimony he made a statement that we need to produce or AI researchers especially more U S citizen or perma- 75 nent resident AI researchers And I think that kind of plays into that issue of how do we deal with AI in space That’s one of the reasons why I’ve been pushing for a college program like an ROTC program to recruit people into the space corps in these areas start identifying students when they’re maybe even in junior high and scholarship them through college to get them into these positions Any thoughts on that Dr PERSONS I’ll just answer quickly and just say I think as Dr Li has I think elegantly pointed out before this is really an interdisciplinary thing I think there’s going to be a need for sort of the STEM STEAM specialist that’s particularly focused on this but I think any particular vocation is going to be impacted in one way or the other just like you could imagine rewinding a couple decades or a few decades—I’ll date myself—but when the advent of the personal computer the PC coming in and how that affected Now we walk into any vocation and somebody’s using a PC or something like that and it’s not unusual but at the time you had to learn how to augment yourself or your task with that And I think that will be the case Mr PALMER Well we’re—if I may Mr Chairman—just to add this final thought that we’ve had to deal with some major hacks federal government systems that are hacked and what we’re faced with we’re competing with the private sector for the best and brightest in terms of cybersecurity We’re going to find ourselves in the same situation with AI experts the truly skilled people and that’s why I’m suggesting that we may need to start thinking about how do we recruit these people and get them as employees of the federal government And that was my thoughts on setting up an ROTC-type program where we would recruit people in we’d scholarship them whether it’s for cybersecurity or for AI and with a four- or five-year commitment to work for the federal government because there’s going to be tremendous competition And the federal government has a very difficult time competing for those type people So with that Mr Chairman I yield back Mr WEBER Now the Chair recognizes the gentleman from New York Mr TONKO It’s okay We’re patient I thank our respective Chairs and Ranking Members for today’s very informative hearing And welcome and thanks to our witnesses I’m proud to represent New York’s 20th Congressional District where our universities are leading the way in artificial intelligence research and education initiatives SUNY Polytechnic Institute is currently the home of groundbreaking research developing neuromorphic circuits which could be used for deep learning such as pattern recognition but are also useful for AI or machine learning In addition the institute has established an ongoing research program on restive memory devices Rensselaer Polytechnic Institute RPI is pushing the boundaries of artificial intelligence in a few different areas In the healthcare front RPI is focusing on improving people’s lives and patient outcomes by collaborating with Albany Medical Center to improve the performance of their emergency department by using AI and analytics to reduce the recur- 76 rence of costly ER visits by patients And RPI researchers are also collaborating with IBM to use the Watson computing platform to help people with prediabetes avoid developing the disease In our fight to combat climate change and protect our environment researchers at RPI and Earth and Environmental Science are working with computer science and machine learning researchers to apply cutting-edge AI to climate issues In the education space RPI is exploring new ways to use AI to improve teaching as well as new approaches to teaching AI and data science to every student at RPI With all that being said there are tremendous universities across our country that are excelling in AI research and education And what are some of the keys to helping AI institutions like them to excel What do we need to do What would be the most important That’s to any one of our panelists Dr LI So thank you for asking this question I think just like we recognize AI really is such a widespread technology that I think one thing to recognize is that it is still so critical to support basic science research and education in our universities This technology is far from being done Of course the industry is making tremendous investment and effort into AI but it’s a nascent science It’s a nascent technology We have many unanswered questions including the socially relevant AI including AI for good including AI for education healthcare and many other areas So one of the biggest things I see would be investment into the basic science research into our universities and encouraging more students thinking in interdisciplinary terms taking courses You know they can be STEM students STEAM students AI is not just for engineers and scientists it could be for students with policymaking mind for students with law interests and so on So I hope to see universities participating in this in a tremendous way like many great schools in New York State Mr TONKO Thank you Dr Persons or Mr Brockman either of you Mr BROCKMAN Sorry I agree with Dr Li but I also point out that I think it is also becoming increasingly hard to truly compete as an academic institution because if you look at what’s happening industry right now is actually doing fundamental research It’s very different from most scientific fields in that the salary disparity between what you can get at one of these industrial labs versus what you can get in academia it’s a very very large And there’s a second piece which is in order to do the research you need access to massive computational resources And for example the work that we just did with this you know game breakthrough that required basically a giant cluster of you know something around 10 000 machines And that’s something where in an academic setting it’s not clear how you can get access to those resources And so I think for the playing field to still be accessible I think that there needs to be some story for how people in academic institutions can get access to that and I think the question of you know where is the best research going to be done and where are the best people going to be I think that’s something that it’s you know playing out right now in industry’s favor but it’s not necessarily set in stone 77 Mr TONKO Thank you Dr Persons Dr PERSONS Yes sir Thank you for the question And I would just add to my fellow panelists the fact that our experts had said that real-world testbeds are important to this You don’t know what you don’t know so not only in addition to adding access to data but being able to test and do things these—one thing for sure and I learned in fact from OpenAI that a lot of the times these things come out with surprising results and so that’s the whole reason of creating safe environments to try things out and de-risk those technologies And that’s something that I think is going to be important to enable that base of research to have an avenue to perhaps move up the technology maturity scale possibly into the market and certainly hopefully to solve critical complex real-world problems Mr TONKO Thank you Very informative Mr Chair I yield back Mr WEBER The Chair now recognizes the gentleman from Illinois Mr FOSTER Thank you Mr Chairman And thank you for coming to testify today You know I’ve been interested in artificial intelligence for quite a long time Back in the 1990s working in particle physics we were using neural network classifiers to have a look at trying to classify particle physics interactions And when I couldn’t stand it during the government shutdown and not so long ago I went and downloaded TensorFlow and worked through the—part of the tutorial on it And you know the algorithms are not so different than what we were using back in the 1990s but the computing power difference is breathtaking And I very much resonated with your comments on the huge increase in dedicated computer power for deep learning and similar—and that is likely to be transformative given the recent—and so that—you know we have to think through that because even with no new brilliant ideas on algorithms there’s going to be a huge leap forward So thank you for that That’s a key observation here You know in Congress I’m the co-Chair of the New Democrats Coalition on the Future of Work Task Force where we have been trying to think through what this means for the workplace of the future And so I’d like to—if—Mr Chairman I’d like to submit for the record a white paper entitled ‘‘Closing the Skills and Opportunity gaps ’’ without objection Mr WEBER Without objection so ordered The information appears in Appendix II Mr FOSTER Thank you And I will be asking for the record if you could have a look at this and see if—you know how—what sort of coverage you think this document has for the near-term policy responses because it’s—you know this is coming at us I think faster than a lot of people in politics really understand And also I will be asking for the record—I guess you may not have to respond right now—where the best sources of information on how quickly this will be coming at us You know there are conferences here and there but you attend and your friends attend a lot of them I’d be interested in where you think—you really come 78 together to get the techno-experts the economic experts you know the labor economists people like that all in the same room I think it’s something we should be putting more effort into On another track I’ve been very involved in Congress in trying to resurrect something called the Office of Technology Assessment You know and what the J O did here is very good which is to bring—we had a conference of the experts and you brought in a good set of experts And a year later now we are getting a report on this And you know you need more bandwidth in Congress than that just you know on all technological issues but this is a perfect example A year-old group of experts and then AI you know is—those are opinions that are sort of dated a little bit even a year in the past And so the Office of Technology Assessment for decades provided immediate high-bandwidth advice to Congress on all sorts of technological issues And so we’re coming closer and closer every year in getting it refunded after it was defunded in the 1990s And so I think—well to ask you a question here is there anyone on the panel who thinks that Congress has enough technological capacity as it currently stands to deal with issues like this Dr PERSONS So—— Mr WEBER I can answer that Mr FOSTER Yes—no it’s a huge problem and it’s been aggravated by the fact that people have decided in their wisdom to cut back on the size and salaries available for Congressional staff One of my—the previous members who talked about—here talked about the difficulty the federal government will have in getting real professionals top-of-the-line professionals in here and you know we’re seeing Members of Congress who are willing to do anything but give them the salaries that will be necessary to actually compete for those jobs Let’s see I am now—let’s see Oh Mr Brockman you had— your—I would advocate everyone have a look at the reference 5 in—which is your malicious use of AI your—reference 5 in your testimony which I spent—I stayed up way too late last night reading that and it is real Along the same lines Members of Congress have access to the classified version of a National Academies of Science study on the implication of autonomous drones for—and this is something that I think you know has to be understood by the military We’re about to mark up a military authorization bill an appropriations bill that is spending way too much money fighting the last war and not enough fighting the wars of the future And then finally Dr Li the—in the educational aspects of this one thing I struggle for—I guess if you look through the bios of people who are the heroes of artificial intelligence they tend to come from physics math places like that And in theoretical physics or mathematics a huge fraction of the progress comes from a tiny fraction of people It’s just a historical truth And I was wondering is AI like that Are they—you know are there a small number of heroes that really do most of the work and everyone else sort of fills in the thing Dr LI So like I said Dr Foster AI is a very nascent field so even though it is collecting a lot of enthusiasm worldwide soci- 79 etally as a science it’s still very young And as a young science it starts from a few people As a—I was also trained as a physics major and I think about early days of Newtonian physics and that was a smallish group of people as well I mean it’s—it would be too much to compare directly but what I really do want to say is that we—maybe in the early even pre-Newtonian days of AI we are still developing this so the number of people are still small Having said that there are many many people who have contributed to AI Their names might not have made it to the news to the blogs to the tweets but these are the names that as students and experts of this field we remember them And I want to say many of them are members of the underrepresented minority group There are many women in the first generation of AI experts So—— Mr FOSTER Yes And when I was—— Dr LI —we need to hear more from them Mr FOSTER —you know two or three clicks down in the references cited by your testimony and you look at the papers there and the author lists it’s pretty clear that our dominance in AI is due to immigrants okay And Dr Li I suspect you might not have come to this country under the conditions that are now being proposed by our President And I won’t ask you to answer that but it’s important when we talk about what it is that makes this country dominant in things like AI It is immigrants okay And I’ll just leave it at that and I guess my time is up Mr WEBER I thank the gentleman I thank the witnesses for their testimony and the Members for their questions The record will remain open for two weeks for additional written comments and written questions from Members The hearing is adjourned Whereupon at 12 24 p m the Subcommittees were adjourned Appendix I ANSWERS TO POST-HEARING QUESTIONS 81 82 ANSWERS TO POST-HEARING QUESTIONS Responses by Dr Jaime Carbonell HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence With Great Power Comes Great Responsibility Dr Jaime Carboneil Director Language Technologies Institute and Allen Newell Professor School of Computer Science Carnegie Mellon University Questions submitted by Chairwoman Barbara Comstock House Committee on Science Space 1 and Technolo Are the discoveries made in arti ciai narrow intelligence thus far re ective of a focus on certain types of arti cial intelligence research For example is there a need for more human cognitive research to help facilitate development of arti cial general intelligence What would be the most effective research path forward to advance AGI developments and who should conduct such research government industry academic institutions Answer First the term arti cial narrow intelligence is not commonly used by most researchers and practitioners in the eld Instead they call it applied Al or just Second yes indeed much of AI is focused on solving speci c important problems such as self-driving vehicles speech recognition image recognition Special purpose robotics There is also considerable work on broader aspects of intelligence such as natural ianguagc understanding autonomous robots underlying computational architectures such as deep neural networks and so on There is less work on trying to build a single uni ed intelligence Many Al researches eschew such a goal in that they see Al as being a powerful suite of methods to improve just about everything rather than to substitute for humans A few however pursue the still distant goal of true general intelligence My belief and that of many but far from ail AI researchers is that the best path towards AGI is largely the one we are on solve rst the dif cult components of intelligent behavior spinuoff useful applications to keep the research funding whether based on the underlying science or the economic or military utility or more typically a combination and gradually build up to broader aspects of intelligence This approach also gives us ample time to weigh the ethical implications as we evaluate the emerging AGE systems 83 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Dr Jaime Carbonell Director Language Technologies Institute and Allen Newell Professor School of Computer Science Carnegie Mellon University Questions submitted by Representative Neal Dunn House Committee on Science Space and Technology 1 With what confidence can we guarantee that we can prevent A I s that are psychopaths such as MIT's Norman Answer Norman is a publicity stunt by a few members of the MIT Media Lab MIT of course conducts much top-quality research so my criticism is not leveled at the institution my alma mater but at this specific project I do not know the inner workings of Norman but from the information they provide it appears to be based on machine learning algorithms being purposely fed false and misleading data The result of course is a confused AI system Humans especially children can be driven crazy indoctrinated or brainwashed if exposed to volumes of false and misleading data with no true or accurate data to counteract it The same is true in a more rudimentary way for machine learning systems So there is no possible guarantee of sanity other than exposure to information that is mostly true and balanced and not meant to deceive An interesting line of new research in AI is to help identifY fake data or postings but that requires exposure to large amounts of normative mostly correct data for contrast This line of research identifying false reports is particularly timely in light of Russian trolling and meddling in US elections but still has quite a way to go 2 Will we be able to teach them to act like they have compassion or empathy What compels them to act this way Yes if we categorize behaviors as demonstrating degrees of compassion or empathy then train machine learning models on contextually selecting said behaviors for appropriate circumstances in which to exhibit them- e g knowing when the AI system should say it is sorry or when to say better luck next time or when to recall and retell a similar situation that caused concern is definitely possible by extending current methods However these approaches simulate empathy or compassion even if done in contextually appropriate ways and that is not the same as truly feeling such emotions No one yet knows how or whether the latter may be doable Page 2 of7 84 3 Will we as flawed humans be able to create a perfect consciousness What will the ramifications of that be Answer I will have to pass on this question because I do not know what constitutes a perfect consciousness and likely no one else really knows either 4 Will we create an algorithm that will override data input Answer Presuming the question refers to data-driven machine learning algorithms this is very much an active area of research Isaac Asimov the late well-known science fiction writer got it right with his three laws of Robotics wherein his imaginary future robots learned adapted and behaved in complex ways but always obeyed the three laws overriding input data when necessary to do so My own research includes this questionhow to create flexible neural-network learning systems that obey constraints rules even as they adapt based on input data Neural nets without added mechanisms will forget earlier information if given overwhelming new data- the real question is how to not forget and always respect important constraints and thus override data if necessary but that requires creating additional structure for these algorithms So the short form of the answer is yes Page 3 of7 85 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Dr Jaime Carbonell Director Language Technologies Institute and Allen Newell Professor School of Computer Science Carnegie Mellon University Questions submitted by Representative Bill Foster House Committee on Science Space and Technology 1 Your testimony mentioned China's goal of dominating the AI industry and that it has already surpassed the US in terms of patents granted for AI technologies Do you have any thoughts on whether our current standard for what is eligible for patent protection helps or hinders AI innovation and American leadership in this space Answer I am not an expert on the legal framework for patentability but the 2014 Supreme Court decision on Alice Corp v CLS Bank International dampened the patentability of software based methods and algorithms including AI methods However more recent interpretations of the Alice decision have been better balanced Some of my colleagues opine that patenting AI methods is more difficult these days than in preAlice times Page4 of7 86 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence -With Great Power Comes Great Responsibility Dr Jaime Carbonell Director Language Technologies Institute and Allen Newell Professor School of Computer Science Carnegie Mellon University Questions submitted by Representative Jacky Rosen House Committee on Science Space and Technology I Research grants and cooperative research agreements are the most common ways we support science and technology research at major universities and federal labs In my district the University of Nevada Las Vegas is working on various applications of artificial intelligence in areas including health care transportation big data and cyber security Currently through an NSF grant UNLV is partnering with the local Clark County School District to mentor high school teachers on computer science cybersecurity and Big Data This program will equip teachers to introduce Big Data and AI curriculum to high school students and hopefully inspire them to pursue further computing education and related careers Across the country we are continuing to see a huge demand for workers in STEM fields but despite these increasing opportunities not enough Americans have the necessary STEM education and skills a What more can we do to prepare and train enough individuals with the essential skills to meet the needs of an AI-driven economy today and in the future Answer This is a very timely question As a countJy we must ramp up the training of AI-practitioners as well as AI researchers The demand far exceeds the supply just ask any hiring manager in the AI industry Part of the problem is that universities have great difficulty retaining AI faculty who are being offered outsize compensation packages by industry thereby reducing the pool of AI educators The issue is not just salary but research funding and facilities- and there the Government can help with increased NSF and other funding to entice the cream of the crop to remain in academia and train the next generations of AI engineers and scientists At CMU we recently created an AI major- that is AI at the undergraduate level to supplement AI programs at the MS and PhD levels But we too suffer challenges retaining our best educators The creation of other such AI programs at other universities would be very useful if they can hire and retain the appropriate faculty in today' s very tight market So in shmt the responsibility is shared between academia create better advanced and undergraduate training in AI and Government create more research funding opportunities to both further the science and to retain the top academic talent in positions where they prepare the next generations of AI experts Page 5 of7 87 2 Research has shown that AI has the potential to reinforce existing biases We need to ensure that the datasets used to input information about our society represent the diversity of the world we live in and address and prevent algorithmic bias a How can we ensure more diversity in datasets that feed into AI systems Answer Indeed the question is timely Biased data leads to biased results- this is true across science and in AI it may also lead to biased AI systems It is our burden to understand the biases and when appropriate to counteract them For instance an AI system trained only on crime reports might conclude that most activities are criminal If the purpose is to analyze patterns when crime is known to occur then that bias is not harmful as the results only apply to criminal activity But if the purpose is to determine whether crimes occurred then the bias will be very harmful- the system would need a diverse and representative example of non-criminal activities as well Hence coping with bias is more a burden on the researchers and AI engineers to either provide unbiased data or to interpret the results in light of the bias and only apply said results when the same conditions exist as were present for the data collection The old adage garbage in garbage out applies universally my version is bias in bias out AI systems are not immune b Could the fact that we still have a significant underrepresentation of women and minorities in the fields that feed into AI- such as computer science physics and engineering- contribute to this bias problem Answer I believe that underrepresentation is not the central problem in this specific data-bias problem The central issue with training machine learning algorithms at the heart of AI systems is bias - deliberate or accidental- in the input training data Any good statistician can vouch for that observation The gender race or ethnicity of the researcher is not directly a data-biasing factor That said however the selection of which problems to address with powerful AI techniques can be more reflective of the researcher's background Problems of concern to minorities may not be selected for study by non-minority researchers And that is a real bias which would be to a significant extent overcome by greater gender race ethnicity and cultural balance among AI researchers It is indeed the case that women and minorities are significantly underrepresented in the fields that feed AI graduate programs and those gaps should be narrowed and eventually closed for many good reasons beyond data bias or problemselection bias I am proud to say that Carnegie Mellon University has achieved close to gender parity in its undergraduate computer science student population Other universities are striving to reach that goal but there is a quite a way to go Page 6 of 88 And we must all work harder to be more inclusive of minority students in these fields My own personal experience having advised some eight PhD students over two dozen of whom being women and a small handful of minority PhD students is that role models and peers play a strong role in their success Being of Hispanic descent has helped me better understand their perspectives and concerns as they are similar to what mine were at the time I studied physics and mathematics on the road to becoming an AI researcher Page 7 of7 89 Responses by Dr Tim Persons HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Dr Tim Persons Chief Scientist U S Government Accountability Office Questions submitted by Chairwoman Barbara Comstock House Committee on Science Space and Technology 1 Are the discoveries made in artificial narrow intelligence thus far reflective of a focus on certain types of artificial intelligence research For example is there a need for more human cognitive research to help facilitate development of artificial general intelligence AGI What would be the most effective research path forward to advance AGI developments and who should conduct such research- government industry academic institutions Answer Our technology assessment on artificial intelligence AI did not explicitly focus on narrow AI versus AGI However our assessment does describe how AI has been defined in a variety of ways an approach for conceptualizing the development of narrow AI and AGI and policy and research areas that should be addressed going forward Specifically we highlighted several policy issues that the experts who participated in our 2017 forum on AI believe require further attention including incentivizing data sharing improving safety and security updating the regulatory approach and assessing acceptable risks and ethical decisions As indicated by the range of forum participants research should be conducted by government industry and academics as all have a role to play in this emerging area 2 What are some ways in which individuals or nation states might use artificial general intelligence AGI maliciously in a physical or cyber-attack Could AGI cyber-attacks wreak more havoc and damage than the types of major cybersecurity hacks reported in recent history such as the Office of Personnel Management data breach Equifax data breach Yahoo data breach etc While we have not done work to compare potential damages caused by AGI versus other types of cyber attacks participants from our forum highlighted challenges and opportunities to enhancing the safety and security of system applications from cyber attacks including those with AI features One participant emphasized that security of the system is only as good as the security of the computer that it is running on so if an AI 90 system is running on a computer that is then hacked the system itself is also vulnerable In addition some participants stated that protecting against hackers is not something an individual company can do on its own Rather efforts to combat hackers need to be industry-wide Page 2 of8 91 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Dr Tim Persons Chief Scientist U S Government Accountability Office Questions submitted by Representative Neal Dunn House Committee on Science Space and Technology 1 With what confidence can we guarantee that we can prevent A I s that are psychopaths such as MIT's Norman Answer While our technology assessment on artificial intelligence did not explicitly address this issue our assessment does highlight concerns policy implications and research areas to consider related to ethical decision making in the context of AI 2 Will we be able to teach them to act like they have compassion or empathy What compels them to act this way Answer Our technology assessment on artificial intelligence did not address how closely AI will be able to learn compassion and empathy if ever However some of the experts who participated in our 2017 forum on AI highlighted the need to develop an ethical framework to govern the use of AI and ensure the actions and decisions of AI systems can be adequately explained and accepted by those who interact with such systems For example some of the participants noted that before humans will understand appropriately trust and be able to effectively manage AI an AI application or system needs to explain why it took certain actions and why it valued certain variables more than others Accordingly a key policy consideration will be to assess acceptable levels of risk and ethical considerations of AI systems 3 Will we as flawed humans be able to create a perfect consciousness What will the ramifications of that be This is another area that could be considered in terms of developing an ethical framework Forum participants highlighted the need for research in exploring computational ethics and explainable AI whereby systems can reason without being told explicitly what to do and inspect why they did something making adjustments for the future Participants noted the need to develop and adopt an appropriate ethical framework to govern the use of AI in research as well as explore factors that govern how quickly society will accept AI systems in their daily lives Page 3 of8 92 4 Will we create an algorithm that will override data input Answer Our technology assessment did not specifically address this issue This is another issue that would have to be considered as part of developing an ethical framework as this area continues to develop Page 4 of8 93 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Dr Tim Persons Chief Scientist U S Government Accountability Office Questions submitted by Representative Bill Foster House Committee on Science Space and Technology 1 Your testimony briefly touched on a forum participant's concerns about the adequacy of current laws to protect products created by AI Do you have any thoughts on whether our current standard for what is eligible for patent protection helps or hinders AI innovation and American leadership in this space Answer In our technology assessment we pointed out that the widespread adoption of AI raises questions about the adequacy of current laws and regulations For example deploying AI requires contractual agreements with the users of the AI according to one participant from our 2017 forum on AI and contracts fall under state law which could be an issue when state laws create rights through contracts that are not protected by the federal government In another example a participant noted that current patent and copyright laws provide only limited protection for software and business methods and questioned whether these laws will adequately protect the products created by AI This same participant also was concerned that current patent and copyright laws pose challenges for AI in at least three key areas • An inventor does not have to reveal very much about their software code to secure a patent or copyright which may be problematic where public safety liability or criminal justice is concerned • Patent protection lasts 20 years and copyrights created by corporations last 120 years both of which are considered too long a time horizon for AI according to this participant Specifically because advancements in AI have moved at such an extraordinary pace this participant believes that data protection for pharmaceuticals may be a better model for protecting innovations in AI In that context a brand-name drug company receives 4-5 years of exclusivity in exchange for making its safety and efficacy clinical trial data available to potential competitors Thus pharmaceutical companies receive this period of exclusivity enforced through the context of regulatory approval in exchange for data openness Page 5 of8 94 • If AI derives its creative results in part through the collective actions of numerous humans according to this participant it is not clear whether that creativity is attributable solely to the program or to the program's creators In our technology assessment we discussed concerns and potential approaches for a government role in mitigating such concerns For instance participants emphasized the need for establishing a safe space to protect sensitive information e g intellectual property privacy and brand information Another participant reiterated optimism that government could get the data it needed to properly protect the public while maintaining proprietary data protections In this regard participants noted successful data-sharing efforts through entities such as MITRE and the National Institute of Standards and Technology NIST Page 6 of8 95 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Dr Tim Persons Chief Scientist U S Government Accountability Office Questions submitted by Representative Jacky Rosen House Committee on Science Space and Technology I Research grants and cooperative research agreements are the most common ways we support science and technology research at major universities and federal labs In my district the University ofNevada Las Vegas is working on various applications of artificial intelligence in areas including health care transportation big data and cyber security Currently through an NSF grant UNLV is partnering with the local Clark County School District to mentor high school teachers on computer science cybersecurity and Big Data This program will equip teachers to introduce Big Data and AI curriculum to high school students and hopefully inspire them to pursue further computing education and related careers Across the country we are continuing to see a huge demand for workers in STEM fields but despite these increasing opportunities not enough Americans have the necessary STEM education and skills a What more can we do to prepare and train enough individuals with the essential skills to meet the needs of an AI-driven economy today and in the future Answer Experts who participated in our 2017 forum on AI noted the need to have research to reevaluate and reimagine training and education as well as better understand the implications of AI on employment Additionally a participant suggested that further research is needed to explore new means to encourage students from low-income backgrounds women and minorities to pursue science technology engineering and math STEM subjects where the majority of jobs will be going into the 21st century In recent prior work we found that certain minority groups made gains in representation in the technology workforce from 2005 to 2015 but that no growth occurred for other minority groups 1 Researchers interviewed for that work identified several factors that may have contributed to lack of growth for some groups including fewer women and minorities graduating with technical degrees and company hiring and retention practices In addition to R D grants each year the federal government manages a portfolio of grant programs to promote STEM education In our most recent work in this 1 GAO Diversity in the Technology Sector Federal Agencies Could Improve Oversight of Equal Employment Opportunity Requirements GA018·69 Washington D C Nov 16 2017 Page7 of8 96 area we noted that the interagency body responsible for implementing the federal STEM education strategic plan has not reviewed programs' performance assessments or reported the participation rates of underrepresented groups in these programs as required by law 2 Reviewing and reporting this information is important for assessing whether programs are broadening access to groups historically underrepresented in STEM fields In our forum we also highlighted the importance of bringing those with STEM skills together with those from other disciplines For instance one participant emphasized that students who want to work with AI will need to master STEM disciplines as well as the social sciences because the technology that is developed is going to interact with humans 2 Research has shown that AI has the potential to reinforce existing biases We need to ensure that the datasets used to input information about our society represent the diversity of the world we live in and address and prevent algorithmic bias a How can we ensure more diversity in datasets that feed into AI systems Although our assessment did not directly address this issue forum participants emphasized that if the data being used by AI are biased the results will be biased as well One participant noted that with AI there is no guarantee of freedom from bias AI can help prevent inappropriate or harmful human bias according to this same participant if it is carefully used if the assumptions of the models are thoughtfully considered and most importantly if the outputs of the model are constantly and closely verified Another participant stated that the baseline is current practice not perfection and that the goal should be to become less biased and more accurate not perfect b Could the fact that we still have a signiticant underrepresentation of women and minorities in the fields that feed into AI - such as computer science physics and engineering- contribute to this bias problem This is not an issue that we have looked at directly as it relates to AI However in a report we issued on women representation in STEM research 3 we noted that research has found that diversity brings different information opinions and perspectives to a group leading to better decision-making than non-diverse groups 4 1 GAO Science Technology Engineering and Mathematics Education Actions Needed to Better Assess the Federal Investment GA0-18-290 Washington D C March 23 2018 3 GAO Women in STEM Research Better Data and lriformation Sharing Could Improve Oversight of Federal Grant-making and Title IX Compliance GA0-16-14 Washington D C Dec 3 2015 4 Moss-Racusin Corinne A et al Science faculty's subtle gender biases favor male students Proceedings of the National Academy of Sciences 109 no 41 2012 Page 8 of8 97 Responses by Mr Greg Brockman HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Mr Greg Brockman Co-Founder and Chief Technology Officer OpenAI Questions submitted by Chairwoman Barbara Comstock House Committee on Science Space and Technology l Are the discoveries made in artificial narrow intelligence thus far reflective of a focus on certain types of artificial intelligence research For example is there a need for more human cognitive research to help facilitate development of artificial general intelligence AGI Answer Most AI research today involves a machine learning technique called deep learning As compared to previous techniques deep learning is more focused on empirical results than theory That said AI researchers are constantly combining deep learning techniques with techniques from other scientific disciplines including neuroscience and graph theory We at OpenAI- and AI researchers everywhere- still probably have many blind spots about how best to build safe and broadly beneficial AGI In the corning years I expect many AGI-relevant insights to come out of scientific disciplines other than machine learning What would be the most effective research path forward to advance AGI developments and who should conduct such research- government industry academic institutions Answer This question can be considered from the perspective of research effectiveness and from the perspective of values or incentives From the research-effectiveness perspective Progress toward AGI benefits from lessconventional institutional settings Progress requires both scientific progress- i e rapidly testing hypotheses and running experiments- and large-scale engineering efforts We've deliberately structured OpenAI to strike this balance between resources and nimbleness As a result we've been able to rapidly iterate- i e carry out a large number of hypothesis- experiment- result cycles -while simultaneously using substantial amounts of computing infrastructure Properties of more-conventional institutional settings are nonetheless critical to AI progress Specifically progress toward AGI will likely require the intellectual 98 environment of academic centers the resources of industry and the flexibility and experimental mentality that is the hallmark of a Silicon Valley startup From the values incentives perspective It's important that any institutions developing advanced AI technology are structurally aligned with the interests of humanity That is to say they are committed to the broad distribution of benefits from AI and they have a safety-first mindset for developing AI OpenAI is structured to serve this function but in principle other institutions with different structures could serve this function 2 What are some ways in which individuals or nation states might use artificial general intelligence AGI maliciously in a physical or cyber-attack Could AGI cyber-attacks wreak more havoc and damage than the types of major cybersecurity hacks reported in recent history such as the Office of Personnel Management data breach Equifax data breach Yahoo data breach etc Answer I'll refer in this answer to the risks of advanced AI more broadly which are on a spectrum of generality and capability beyond today's AI systems Advanced AI like all technologies can be used to amplify our human ability to achieve our goals whatever those goals might be Hopefully the goals most of us want to pursue are generally agreeable- e g be more economically productive Unfortunately some people have less-agreeable goals and- absent social or technical safeguards- advanced AI will empower them as well Using advanced AI to augment illegal or unjustified physical- or cyber-attacks is just one such example Advanced AI might not only increase the risk of these types of attacks it might also make it more difficult to identify the source of the attack and lower the barriers to entry for attacks We recently co-authored a report on the risks of the malicious use of AI which goes into substantial detail about actions that can be taken today to address these risks including for example research into automated exploit generation 1 1 https arxiv org pdf 1802 07228 pdf Page 2 of8 99 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Mr Greg Brockman Co-Founder and ChiefTeehnology Officer OpenAI Questions submitted by Representative Neal Dunn House Committee on Science Space and Technology l With what confidence can we guarantee that we can prevent A l s that are psychopaths such as MIT's Norman Answer Antisocial or unsafe behavior from an AI can be a result of either deliberate or inadvertent training Both of these safety issues need to be addressed but in different ways MIT Media Lab's Norman experiment was a deliberate attempt to highlight how sensitive the characteristics of an AI system are to the data used to train it 2 The MIT researchers chose to train an AI system on a variety of frightening images they then asked the trained system to label new images that were not part of its training dataset The system selected many frightening or disquieting labels for these images precisely because of how it was trained This experiment highlights an important risk but it's critical not to misinterpret the experiment as evidence that psychopath-like properties might be emergent in AI systems That said there are also unintentional safety risks related to the training and reward structure of AI systems such as bias in an algorithm used in hiring or parole decisions MIT's Media Lab is working on solutions to these risks So are for example Microsoft's Fairness Accountability Transparency and Ethics in AI FATE group 3 and Google Research 4 OpenAI has a team devoted to working on how best to avoid a subclass of these risks specific to more-advanced AI systems For example see our collaboration with DeepMind on Learning from Human Preferences 5 which explores a training technique that might become more critical as AI systems become more advanced 2 https www media mit edu projects normanloverview https www microsofi com en-us research group fate See e g Mitchell Algorithmic Bias in Artificial Intelligence The Seen and Unseen Factors Influencing Machine Perception ofimages and Language https www clsp jhu edu events margaret-mitchell-googleresearch # Wl Zu9hKiL9 5 https lblog openai com deep-reinforcement-learning-from-human-preferences 3 4 Page 3 of8 100 2 Will we be able to teach them to act like they have compassion or empathy What compels them to act this way Answer AI systems' values are a function of their learning procedure and training environment AI systems acquire values implicitly similar to the way humans learn values they develop internal biases and subjective judgments which reflect the data they are exposed to during training This is cause for both optimism and concern It's cause for optimism in that it means we should in principle be able to train AI systems to display the types of compassion and empathy that we ourselves display It's cause for concern to the extent the level of empathy and compassion among humans today is lower than we would like it to be As noted above this is an area of active research at OpenAL 3 Will we as flawed humans be able to create a perfect consciousness What will the ramifications of that be Answer This is a complicated question that could easily be misinterpreted To accurately answer it would likely require further clarification and discussion With that caveat as is the case for any technology it's important that we build safe AI systems the behavior of which is aligned with human goals including human values OpenAI works toward this goal in several respects For example it does fundamental research related to developing advanced AI systems that can be aligned with human goals It also strives to cultivate a broader interest in the machine learning community in safe and socially responsible AI research and development For a high-level overview ofOpenAI's commitment to these principles see its Charter 6 4 Will we create an algorithm that will override data input Answer AI systems' actions are the result of their learning procedure and training environment Sometimes the actions a system takes in pursuit of an apparently straightforward objective can be concerning or surprising For example we once trained an AI system to play a simulated boat race game by trying to maximize the game score 7 We assumed the system would learn to complete the race quickly believing that was the best way to achieve a high score But the system found another way It found a lagoon where it could turn in a large circle and repeatedly knock 6 7 https blog openai com openai-charter https blog openai com faulty-rcward-functions Page 4 of8 101 over targets timing its movement so as to always knock over the targets just as they repopulate while catching on fire and crashing into other boats The boat race experiment is an example of an AI system achieving an apparent straightforward goal in a surprising- and potentially dangerous- way The problem is that the actual goal of the human designers was not simply to maximize game score but to do so in a way consistent with the spirit of the game But programming the spirit of the game as part of the program's formal objective is much more complicated This example evidences a potentially concerning safety risk but it's important to distinguish this risk from the risk of an algorithm overriding data input The algorithm did not ignore the data rather it found a new unexpected way of achieving a particular outcome We've co-authored a paper which describes several technical problems AI researchers will need to solve to ensure AI systems act predictably and in accordance with our actual objectives 8 8 https arxiv org abs 1606 06565 Page 5 of8 102 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Mr Greg Brockman Co-Founder and Chief Technology Officer OpenAI Questions submitted by Representative Jacky Rosen House Committee on Science Space and Technology I Research grants and cooperative research agreements are the most common ways we support science and technology research at major universities and federal labs In my district the University ofNevada Las Vegas is working on various applications of artificial intelligence in areas including health care transportation big data and cyber security Currently through an NSF grant UNLVis partnering with the local Clark County School District to mentor high school teachers on computer science cybersecurity and Big Data This program will equip teachers to introduce Big Data and AI curriculum to high school students and hopefully inspire them to pursue further computing education and related careers Across the country we are continuing to see a huge demand for workers in STEM fields but despite these increasing opportunities not enough Americans have the necessary STEM education and skills a What more can we do to prepare and train enough individuals with the essential skills to meet the needs of an AI-driven economy today and in the future Answer One way to encourage more people to learn AI-relevant skills is to recognize and reward excellence in AI research e g through prizes awards fellowships or special programs For example the vanguards of early space development were highly regarded by their communities which in tum inspired a generation of children to dream of becoming astronauts Similarly the physicists of the mid-1900s received great renown and prestige leading to many of the brightest minds of a generation to learn physics I hope that we can inspire a new generation of AI researchers in a similar way A more-direct incentive than prestige is funding Funding often comes with conditions that make it less attractive to researchers Increasing the amount of nostrings-attached grants for AI research --- particularly pro-social research that might be insufficiently incentivized in commercial markets -- could be highly effective A related mechanism is direct funding of PhD students or PhD programs But prestige and funding are insufficient if the best minds of the next generation don't have the requisite skills To address that risk we should invest more in Page 6 of8 103 STEM education These investments should target both early-age educational programs as well as more-practical vocational programs for older students and adults 2 Research has shown that AI has the potential to reinforce existing biases We need to ensure that the datasets used to input information about our society represent the diversity of the world we live in and address and prevent algorithmic bias a How can we ensure more diversity in datasets that feed into AI systems Answer We should work toward a research and industry culture that normalizes testing for bias fairness and interpretability- among other risks- throughout the research development and deployment process In practice this takes many forms First we should make sure that the datasets we use to train AI systems reflect the diversity of their real-world environments For example a recent Washington Post analysis showed how widely deployed smart speakers from Google and Amazon Google Home and Alexa favor certain accents likely as a consequence of being trained on data that favors those accents 9 Second we should bear in mind historical human biases that might be deeply ingrained in the data on which AI systems are trained 10 If our goal is to train AI systems that can improve the world we should take care not to entrench the biases of the past Third we should seek international approaches to international problems both as a means of obtaining more-diversified datasets and as means of achieving positive-sum solutions to global challenges Lastly governments might consider creating- or catalyzing the creation ofdatasets that are more representative of the entire population the governments represent If such datasets existed private organizations- such as OpenAI wouldn't depend so heavily on the selective market-driven datasets developed in the commercial sector 9 https www washingtonpost com graphics 2018 businesslalexa-does-not-understand-youraccent noredirect on utrn terrn 32f0e25b4996 10 See e g Crawford TheTrouble with Bias https nips cc Conferences 2017 Schedule showEvent 8742 Page 7 of8 104 b Could the fact that we still have a significant underrepresentation of women and minorities in the fields that feed into AI- such as computer science physics and engineering- contribute to this bias problem Yes AI is both a technology that can transform its enviromnent and a reflection of the environment in which it's created As I note above this fact creates a risk of entrenching current cultural biases in our technology I cite above several current research efforts- at many different organizations - working to solve these problems OpenAI for its part has started an OpenAI Scholars program to create paid mentorship for high-potential people from backgrounds underrepresented in the field 11 But much more is happening than I have cited and much more needs to happen still '' https lblog openai com opcnai-scholars Page 8 of8 105 Responses by Dr Fei-Fei Li HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Dr Fei-Fei Li Chairperson of the Board and Co-Founder AI4ALL Questions submitted by Chairwoman Barbara Comstock House Committee on Science Space and Technology 1 Are the discoveries made in artificial narrow intelligence thus far reflective of a focus on certain types of artificial intelligence research For example is there a need for more human cognitive research to help facilitate development of artificial general intelligence AGI What would be the most effective research path forward to advance AGI developments and who should conduct such research- government industry academic institutions Answer I'm a strong proponent of a cross-disciplinary approach to AI in which complementary but distinct fields play a collaborative role Since its inception our field has benefited from the insights yielded by research in fields like neuroscience and cognitive science and this trend should be encouraged as the technology develops For example today's Deep Learning algorithms have benefited greatly from Nobel Prize winning research in the mammalian primary visual cortex from the 1960's Much of this collaboration will begin at universities of course as so much of the requisite talent is already sharing a campus but our attitude towards the traditional separation of disciplines will have to be modernized in order to foster the level of crosspollination the future of this technology will demand Industry players will have to recognize that the creation of truly beneficial machine intelligence is not merely an engineering problem and be willing to augment their own development teams with expertise that lies beyond computer science Finally governments can use funding public research projects and even legislation to ensure this work remains transparent and committed to the common good I caution against the notion of AGI as the focus ofthis endeavor however Across the field little consensus exists on whether a distinction between AGI and AI is necessary or even whether it makes sense More importantly AI is already playing a pivotal role in our world and its influence is growing apace This impact is being felt in finance the criminal justice system and even healthcare and developments already on the horizon promise a transformative effect on the lives of people all over the world By basing the AI of today and tomorrow on humanistic cognitive and even empathic insights from across the academic spectrum we can ensure its impact is a positive one 106 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Dr Fei-Fei Li Chairperson of the Board and Co-Founder AI4ALL Questions submitted by Representative Neal Dunn House Committee on Science Space and Technology 1 With what confidence can we guarantee that we can prevent A I s that are psychopaths such as MIT's Norman Answer It's important to remember that AI is still a nascent field and that many important questions have yet to be answered This makes it difficult to talk in terms of confidence especially on such speculative topics However MIT's Norman demonstrates the remarkable susceptibility oftoday's AI to the data and experiences from which it learns If we want our algorithms to treat humans with respect and compassion it's up to us to prioritize such values when developing them In fact I often end talks by reminding the audience that the behavior of machines will always be our responsibility no matter how autonomous they become 2 Will we be able to teach them to act like they have compassion or empathy What compels them to act this way 3 Will we as flawed humans be able to create a perfect consciousness What will the ramifications of that be 4 Will we create an algorithm that will override data input Answer to 2 3 and 4 As I mentioned above we're so early in the evolution of this technology that we should approach any speculation with caution especially when such complex ideas are involved Our understanding of consciousness remains both rudimentary and contentious with ongoing debates from philosophy to neuroscience However whether or not consciousness will someday play a role in AI we know that today's machines can already reflect our worst traits-and even amplify them In light of this I believe we have an obligation to learn all we can from such mistakes contain them before they grow and take intelligent steps to help avoid them in the future I liken this to the scientific method itself Human perception is inherently limited-and easily misled especially at very small and very large scales-and the scientific revolution was perhaps the first widespread acknowledgement of this awkward truth Fortunately our blind spots didn't preclude our ability to recognize them and even build a system of reasoning to help us work around them Likewise although we're far from ethically perfect beings I'm optimistic we can develop practices that anticipate those imperfections in the technology we create Page2 ofS 107 We're certain to make mistakes in the coming years Whether or not we respond to them productively will be determined by our ability to talk honestly about them Our greatest asset will be our humility Page 3 ofS 108 HOUSE COMMITTEE ON SCIENCE SPACE AND TECHNOLOGY Artificial Intelligence- With Great Power Comes Great Responsibility Dr Fei-Fei Li Chairperson of the Board and Co-Founder AI4ALL Questions submitted by Representative Jacky Rosen House Committee on Science Space and Teclmology 1 Research grants and cooperative research agreements are the most common ways we support science and technology research at major universities and federal labs In my district the University ofNevada Las Vegas is working on various applications of artificial intelligence in areas including health care transportation big data and cyber security Currently through an NSF grant UNLV is partnering with the local Clark County School District to mentor high school teachers on computer science cybersecurity and Big Data This program will equip teachers to introduce Big Data and AI curriculum to high school students and hopefully inspire them to pursue further computing education and related careers Across the country we are continuing to see a huge demand for workers in STEM fields but despite these increasing opportunities not enough Americans have the necessary STEM education and skills a What more can we do to prepare and train enough individuals with the essential skills to meet the needs of an AI-driven economy today and in the future Answer In addition to encouraging STEM education across the board more can be done to open these fields to traditionally underrepresented groups like girls and people of color For example my own non-profit AI4ALL runs a growing range of summer programs across the country in which high school students from underserved communities get hands-on exposure to AI research in real university labs Our own polling suggests this has a transformative effect on what's possible in the minds of such students and I strongly advise we expand this effort However while diversity is a worthwhile goal by any standard-especially in a country founded on the value of inclusion-there's an argument to be made on purely numerical grounds as well the future will soon present challenges that can only be met with a wide deep pool of talent and we simply can't afford to turn away qualified contributors 2 Research has shown that AI has the potential to reinforce existing biases We need to ensure that the datasets used to input information about our society represent the diversity of the world we live in and address and prevent algorithmic bias a How can we ensure more diversity in datasets that feed into AI systems Page4 ofS 109 b Could the fact that we still have a significant underrepresentation of women and minorities in the fields that feed into AI- such as computer science physics and engineering- contribute to this bias problem Answer to both a and b Bias in data can be addressed in two primary ways First the process by which data is organized must be made more sensitive to our myriad blind spots as humans and more adept at counteracting them This is an area of active research and while there's a long way to go I'm happy to say we're seeing early signs of its adoption in industry Second and equally importantly the people who take part in this process must themselves reflect the diversity of the world at large Despite growing efforts to increase representation in tech it remains predominantly white and male Addressing this disparity will go a long way towards ensuring our technology treats everyone with fairness This is another reason why organizations like AI4ALL are so important more diversity in the AI industry means better experiences for its users far beyond Silicon Valley Page 5 of5 Appendix II ADDITIONAL MATERIAL FOR THE RECORD 111 112 OPENING STATEMENT SUBMITTED BY DR CARBONELL Testimony of Professor Jaime Carbonell before the Committee on Science Space and Technology Subcommittee on Research and Technology and Subcommittee on Energy of the U S House of Representatives on the hearing titled Artificial Intelligence- With Great Power Comes Great Responsibility June 26 2018 Good Morning My name is Jaime Carbonell I am a professor of computer science and the director of the Language Technologies Institute at Carnegie Mellon University I would like to thank Chairman Comstock of the Research and Technology Subcommittee and Chairman Weber of the Energy Subcommittee for inviting me to testify today I would also like to thank Ranking Member Lipinski Ranking Member Veasey and all of the Members of the Committee for your interest in Artificial Intelligence It is a field in which I have been working for over 40 years Background Artificial intelligence AI is the ability to create machines who perform tasks normally associated with human intelligence from autonomous driving and spoken language comprehension to medical diagnosis and scientific discovery Throughout its roughly 60-year history AI has been incubated admired dismissed as infeasible respected maligned feared and more recently mainstreamed Progress has been inexorable over the years based on faster hardware larger memories new computing paradigms such as parallelism in graphical processing units GPUs major algorithmic innovations such as deep neural networks and a plethora of successful applications driving the current industrial AI boom A Brief History The term Artificial Intelligence was coined in 1956 at a conference in Dartmouth University by the four founders of the field Herbert Simon Alan Newell Marvin Minsky and John McCarthy in their shared conviction that human thought could be reduced to symbolic computation on digital computers in the near future Simon and Newell founded the AI efforts at Carnegie Mellon University Minsky at MIT and McCarthy at Stanford Although there were some early successes such as propositional logic provers and checker-playing programs the true magnitude and complexity of human reasoning and the difficulty in replicating aspects of human intelligence in digital computers ensured that AI was a very longterm endeavor not just a single moonshot or even a Manhattan Project The 1960's witnessed the founding of most of the AI subfields computer vision natural language processing speech recognition robotics automated reasoning and planning and so on In the 1970's symbolic approaches to AI dominated including reasoning based on first-order 113 logic and in the 1980's the first substantial commercial benefits of AI were reaped with the advent of rule-based expert-systems These systems encoded narrow human expertise in commercially important areas such as automatically configuring digital computers or deciphering mass-spectrograms Another example of deep narrow expertise was in chess playing machines IBM's Deep Blue based on the earlier Deep Thought chess playing program from Carnegie Mellon University beat world champion Gary Kasparov in 1989 Additionally the field of modem Machine Learning was founded in that decade During the 1990's steady progress was made on virtually all the component AI technologies including neural networks for learning combining vision navigation and manipulation into operational robotics and combining speech recoguition and natural language processing into dialog systems The 90's also heralded the statistical approaches in AI and machine learning reducing many cognitive problems to mathematical optimization Strangely the 1990's are sometimes called the AI winter since much of the progress was not externally visible and other technologies such as the advent of the internet and early electronic commerce captured the limelight After the tum of the century industry started taking AI more seriously The search engine companies for instance worked on semantic associations to fmd information beyond the exact keywords in a query For instance asking for inexpensive vehicle insurance yields results about cheap car insurance Those same companies accessing huge amounts of electronic text in multiple languages to train cross-language associations were able to build the first widely-used general-purpose machine translation engines The rule-based expert systems of earlier decades evolved and became embedded into manufacturing systems and workflow-planning and otherwise entered routine practice and were no longer the subject of academic inquiry Instead academia sought new challenges such as integrated virtual agents autonomous robots multilingual speech-to-speech translation and so forth The movement of Artificial General Intelligence AGI took root where the focus on performing very well at a narrow difficult task was replaced by the desire to exhibit adaptive intelligence to perform reasonably across many tasks i e re-focusing on the initial goal of AI to create human-like intelligence In the current decade interest in AI has exploded in large part due to the resurgence of deep networks in machine learning at the core of AI and their ability to perform many tasks much better than before - tasks ranging from self-driving vehicles and robotic assistants to robust speech understanding to semi-intelligent chatbots such as Siri Cortana and Alexa All the large tech companies have sprouted AI labs or divisions as have many in other disciplines including finance and manufacturing AI and its sub-disciplines Borrowing the title of the famous 1955 movie we can say that human intelligence is a many-splendored thing it is also a many-faceted thing Vision and physical coordination is not a trait unique to humans but inherited from our ancestor species On the other hand complex language writing higher mathematics and abstract reasoning most scientists would attribute uniquely to our species or at least our genus AI attempts to understand and replicate all aspects of human intelligence including those that evolution took many millions 2 114 of years to perfect But to make the ultimate challenge somewhat more tractable AI is divided into multiple somewhat overlapping disciplines • • • • • Sensing Using vision sound touch and super-human sensors such as sonar radar and lidar to sense and form a mental model of the immediate environment and update the model dynamically Communicating Using language written or verbal to convey information questions or commands to other agents in the social group as well as to record knowledge for future reference Acting Effecting changes in the physical world by manipulating objects navigating to other locations observing the outcomes of the actions and recording the results Reasoning Envisioning performing physical or communicative actions in the world inferring their consequences and planning sequences of actions to achieve desired goals as well as inferring the likely actions of other agents Learning Acquiring knowledge and experience to improve other cognitive tasks sensing acting communicating and reasoning including the ability to solve new problems or address new situations Many established AI areas map onto one or a combination of the above disciplines For instance computer vision is primarily sensing but also learning to recognize store and classify patterns Natural language processing is primarily communicating but contains elements of reasoning e g semantic interpretation and learning Robotics while centered on sensing and acting touches on all the AI disciplines Sometimes AI disciplines are confused with the overall endeavor as in are you doing AI or machine learning This is like asking are you a car repairman or an engine mechanic If someone is practicing the latter they are also practicing the former Similar confusion is evident in should we invest in robotics or AI Or Should we build a chatbot rather than doing AI Instead the better question is should we focus in Robotics- or on chatbots- as the most relevant part of AI for our business All of the subfields of AI are reinforced by advances in adjacent subfields and some like machine learning are crosscutting and ubiquitous Applications of AI In many ways AI is the ultimate meta-technology the technology that enhances most other technologies whether they be in strategic planning manufacturing healthcare transportation customer contact entertainment forensics mining agriculture or even scientific research AI extends to defense and national security ranging from intelligence gathering and interpretation to operational warfighter capabilities And AI is playing an increasing role in humanitarian areas such as workplace safety and disaster relief Let us look at a few illustrative applications of AI focusing on the areas this writer knows best • Question Answering Wouldn't it be great if we could bring to bear all the world's knowledge- or at least all the publicly available knowledge- to bear in answering 3 115 • • • • automatically every burning question Starting with DoD funding followed by IBM's Watson in the Jeopardy Challenge to today's open-domain systems Q A has emerged as a major challenge task for AI I am proud to say that Carnegie Mellon has been a central player in Q A from the get-go in government programs through helping IBM build the original Watson to heading the leader boards in current Q A challenges Autonomous Driving Imagine drivers who do not get distracted do not get sleepy always know their routes and can be counted on to stay always sober That's the safety promise of autonomous vehicles- not an absolute safety guarantee but nonetheless a major improvement From the Navlab project at CMU in the 1980's to No hands across America autonomous highway driving in the 1990's to present large-scale commercial endeavors autonomous vehicles are the future of safe transportation Workplace safety What if blue-collar workplace accidents could be predicted and thereby reduced AI in the form of historical workplace data analysis correlating accidents with workplace conditions safety inspections and behavioral indicators can help do just that as evidenced by joint work between our university and a company called Industrial Scientific already helping improve worker safety in many companies across different industries such as construction mining manufacturing etc Massive multilinguality There are some 6 000 languages in the world but only the top 2% or so by population and economic significance have translation software What if there is a natural disaster requiring international assistance or what if the US military needs to coordinate with local authorities for the other 98% NSF ARO and now DARPA are addressing this challenge to build information extraction from low-resource languages and rudimentary translation in a matter of days rather than years Game theory AI has proven itself at full-information games such as chess and go and more recently at partial-information games such as Texas holdem poker beating the human champions in each of these games The same technology is used to optimize kidney exchanges negotiations and other partial knowledge high-payoff decision tasks The above examples represent a very small cross-section of a much larger AI application space How about agriculture AI is optimizing planting and fertilizing plans and schedules as well as creating robot-controlled tractors Law enforcement Voice and facial identification are helping forensic analysis Music AI-based accompaniment and AI-based tutors for different instruments Healthcare AI-based DNA analysis for risk factors robo-surgery for super-human precision analysis of patient records and much more are emerging Finance Hedge-fund analytics based on machine learning investment risk and so on Education AI based tutors in mathematics language and computer programming are emerging There is nary an industry untouched by AI and its impact will significantly grow and become more manifest in the coming years across the board Artificial General Intelligence AGI AI is rapidly embedding itself into just about every aspect of our lives its utility is beyond question However how about the original AI dream of 4 116 creating general human intelligence Many AI researchers have not given up on that goal though many feel the path leads through the various components of intelligence before striving for a grand unification Other AI researchers are content with narrow AI that is task-oriented AI and not general intelligence since that is what drives our economic engine Yet others impatient to return to the AI-genesis goal of creating human-level intelligence coined the phrase Artificial General Intelligence seeking to leapfrog a component-based approach with a more holistic one Some AI researchers such as this writer view AGI as an aspirational goal working to generalize AI methods from the narrow to the broad from the specific to the general For instance consider transfer learning If a person learns to drive a car perhaps through instruction and considerable practice and then she is asked to drive a van or a small truck chances are she can do so albeit with a few awkward moments adjusting to the height and size of the new vehicle However if the software trained for a self-driving car were to be extracted and implanted onto a small truck disaster may ensue- judging distances to other vehicles would be off due to the higher camera angle the breaking distance and turning radius would not behave as expected and even the lanedetection and following might fail The difference is that humans transfer what they learned previously while compensating for the known differences such as height angles size vehicle responsiveness etc Transfer learning strives to do exactly that in an automated manner learning what to keep what to discard what to modifY and how to modifY it across related tasks Currently fielded AI approaches require retraining from scratch insensitive to having learned a very similar task Hence transfer learning is a small but important step towards AGI Another example is deep neural networks whose topology must be currently hand crafted by researchers for a given task e g for face recognition vs reading MRls or for machine translation vs text mining Each task requires determining how many nodes are required arranged into how many layers what type of layer-based combination functions what kinds of connectivity among the layers and so on Researchers including this writer are striving to create self-configurable deep networks that change their structure automatically to optimize task performance That is another small step towards AGI Many other researchers are also investigating how to make AI more general one step at a time in many different ways and in different sub fields such as in robotics natural language processing automated reasoning or machine learning However a few researchers are less patient trying to reach AGI by more direct means and though not yet having achieved the desired breakthroughs are nevertheless determined to pursue the proverbial gold at the end of the long rainbow Some common AI myths Some popular beliefs or claims about AI deserve the label of myth as they are contrary to observations and informed opinion including the following • One stop shop for all ofAI AI is not binary it is not something one has or fails to have The field continues to progress at a good pace over time more sophisticated and capable 5 117 • • • • AI methods and systems are created Any vendor claiming to be the ultimate provider of everything AI is selling a bill of goods or a pig in a poke As more AI capabilities come online we must be ready to see whether and how to employ them and we never know for certain where or by whom they will be created AI deep neural nets Over its history many AI paradigms emerged dominated the field and then fell out of favor to newer methods first-order logic rule-based systems the first coming on neural nets 1980s symbolic machine learning the second coming of neural nets 1990s statistical machine learning probabilistic reasoning and now the third coming of deep neural nets Actually every AI paradigm leaves its mark and many powerful systems use hybrid approaches not just the latest method du· jour even as convolutional and recurrent deep networks are proving to be very powerful AI suite of software tools This is no more true than carpentry being just a box of carpenter's tools one also needs the master carpenter Surgery is not just a suite of scalpels sutures and other surgical implements one needs the experienced surgeon AI tool suites are major enablers of novel AI applications but the skilled AI practitioner is an integral part of the equation AI is impossible This used to be a common way for pundits to gain attention including some philosophers claiming that AI would never create a champion chess player and would never understand human speech and would never drive a vehicle in traffic Over time all were proven wrong Now the more common claim is that AGI is impossible That one will be harder to disprove as AI systems become more general and more powerful the proverbial goal posts will be moved from passing the Turing Test 1 to ever harder tasks as each is accomplished over time Time and much research will tell whether true AGI is indeed achievable AI will cause massive unemployment A variant of this claim has been made after every major technological advance but the opposite is typically true- there was more employment after the industrial revolution than before or after the introduction of information technology and the internet AI is already deployed in many industries yet we have the lowest unemployment rate in recent times Instead AI will likely change the nature of work displacing workers in some fields over time and creating other remunerative jobs Rather than speculation careful studies by economists AI scientists and policy makers are required to make better predictions of labor market effects of AIwe simply do not know them in any detail But it is fairly clear that the countries most advanced in AI technologies will reap its benefits and others will be left behind just like in earlier industrial revolutions as the effects of AI will indeed be global 1 The most popular version of the test designed by Alan Turing the famous code-breaking British mathematician is whether a person asking questions of two hidden responders one a human the other a machine can tell which is which after a brief period of time based on their responses If the questioner cannot tell them apart then the machine is said to have passed the Turing Test 6 118 Glimpsing into the futnre of AI The most reliable prediction one can make relating to AI is extrapolating current trends into the near future • • • • AI will become increasingly ubiquitous in everyday life from general chatbots to individualized health monitoring to personal assistants Turing award winner Professor Raj Reddy from CMU predicts that personal assistants will evolve into guardian angels always-on AI system helping people prepare for meetings monitoring news or social media for items of immediate interest reminding users to exercise or to not touch that temping chocolate dessert These angels or cogs as this writer prefers to call them will be driven by observing learning and receiving instructions from their users The power of underlying AI technologies will increase as machine learning language technologies and robotic sensing improve due to more powerful GPUs improved deep neural network architectures better training algorithms that combine raw data with domain knowledge and so on These will increase the range of practical AI applications across the board AI will expand into new areas including cyber security where research is already establishing that zero-shot mal ware detection is possible via machine learning methods trained on prior cyberattacks and malware profiles This is a major improvement to waiting for a new virus to infect many machines before viral signature is disseminated by the cyber protection provider AI can also help in energy conservation including smartgrid and smart-home power consumption as well as optimizing renewable power installations via predictive local-climate models Artificial general intelligence will witness a resurgence not so much from a major paradigm switch in AI though that is what its proponents desire but rather from evolutionary forces in bringing to bear transfer learning self-configurable neural networks more flexible robotics more natural and general dialog systems and other • • improved AI methods coming together into larger systems of wider capabilities The demand for AI talent will far outstrip the supply Industries engage in bidding wars and employee poaching in attempts to fill open positions in AI raising compensation levels well beyond any previously seen Universities strive to increase the supply but when industry attracts top university faculty- and this is happening with increasing frequency- it is very difficult to maintain let alone expand educational offerings AI in other countries will surge This is already happening in Russia India and especially China see the following section but the trend will expand to many more countries that realize AI is the crucial technology of tomorrow and quickly ramp up efforts in research education and especially industry 7 119 • AI ethics will become a central concern As with virtually all technologies AI can be used for the benefit of human kind or its detriment For instance electricity powers our world but also can electrocute the internet is a major economic driver and wonderfully empowering by democratizing information but it is also used to spread hate-speech and enable cyber bullying AI is no different and we must consider ethical implications where we refers to AI researcher and practitioners ethicists economists and policy makers working together For instance we may want AI to help us establish and maintain healthy lifestyles but we may not want that same AI to breach our privacy and inform others of our individual health problems US Dominance in AI The United States has been the undisputed world leader in AI from its inception in 1956 through the early days of overoptimistic expectations that AI was just around the corner to the realism that AI is difficult- much harder than rocket science- and its ultimate success in a wide variety of challenging tasks In the 1980's Japan's Fifth Generation Project challenged the US AI dominance and the 1990's Europe's Economic Commission attempted to surpass US AI efforts but neither succeeded in large part due to the foresight and perseverance of the US public sector Unlike the above limited-lifetime efforts the National Science Foundation NSF the Defense Advanced Projects Agency DARPA the US intelligence community including the Intelligence Advanced Projects Activity IARP A and the research divisions of each of the armed services ONRINRL AFRL ARL ARO persevered in supporting AI through all of its growing pains its successes its more challenging times and its arrival in the mainstream to become a top candidate for the most versatile and valuable industry of our century The foresight to persevere to strive to win the long game of AI is a credit to the cited government institutions through multiple administrations and through remarkable bipartisan unity in the national interest Hence my top recommendation is kudos to these agencies they do an excellent job please extend them as much support as possible However the US dominance in AI is being challenged like never before Many countries are striving hard to improve their AI know-how work-force and industry including China Russia Korea Japan Germany the UK and India Consider China for instance which has made AI a national priority On May 4 2018 CNBC reported China is determined to steal the AI crown from the US and nothing not even a trade war will stop it China's 2030 plan envisions building a $1 trillion AI industry Wired Magazine reports China will be the world's dominant player in artificial intelligence by 2030 This isn't a prediction by a researcher or academic it's government policy from Beijing Whereas these statements may be on more alarmist than reliable predictions they clearly indicate Chinese intent China's national priority is AI preeminence Even General Secretary Xi Jinping is reported to have AI books on his shelf It is difficult to estimate the very substantial level of AI funding in China but there some components include 1 The city ofTianjin is committing $5 billion to support the new AI industrial park 2 The Feb 20 2018 statement in the Financial Times saying Last year almost half the global investment into AI startups went to China up from a mere 11 3 per cent slice in 8 120 2016 3 On June 22 2018 the South China Morning Post reported China's Ministry o f Science and Technology has funded at least eight A1-related research projects over the past six months to the tune of 2 73 billion yuan US$430 million from the central government budget and The China Academy of Sciences CAS which has over 300 labs and four national research centres received over 2 7 billion yuan for its 11 fundamental science projects last year although it's unclear how many of these are directly-A related China has already far surpassed the US in terms of patents granted for AI technologies according to Quartz May 2 2018 In contrast the United States spends about $2 billion per year on AI research according to OSTP not counting substantial DoD spending on procurement that includes applied AI research The bottom line is that both countries take AI very seriously and that China is striving and investing to achieve AI leadership It will take a greater effort for our country to maintain international leadership in AI Some recommendations The following suggestions address what this writer believes to be national priorities with respect to AI whose neglect will be to our collective peril • Support the US government agencies which helped create and foster AI NSF DARPA ARPA parts ofNRL ONR ARL ARO AFRL etc Without DARPA in the early days AI would not exist in its present form These agencies do a good job of funding and shepherding AI research They deserve increased financial support to do their jobs even better especially in light of the sudden large-scale international competition No matter what other measures the US government takes with respect to AI keep and grow the proven successful processes and agencies • Address the AI personnel shortage We need to produce more AI researchers especially more US-citizen or permanent-resident AI researchers One way to do so is to provide scholarships for US students pursuing AI careers One thought is that these could be funded by asking companies in an expanded merit-based HlB visa program to pay $30K or so per sponsored visa Then if there are some 1OOK new Hl B visas per year that would create a fund of $3 billion per year which would provide something like 30 thousand scholarships Whatever the mechanism or the numbers we should be training more AI researches and engineers The flipside is to retain more the foreign AI talent trained in US universities which is also addressable by increasing HlB visas • Address the exodus ofA faculty from universities Industry is making offers that few AI faculty can refuse with the inevitable loss of top university AI talent making it very difficult to teach and mentor new generations of students We must stop eating our seed com though it is not clear how to change incentives to attract and preserve more toptalented AI faculty • Encourage AI at the undergraduate leveL AI used to be for PhD students then MS programs in AI or sub-disciplines of AI started including several at CMU Now is the time to start AI as an undergraduate major as we are experimentally doing at CMU 9 121 • Consider funding a national AI center Other countries have national AI laboratories or are creating same as a uexus between government industry and academia The US should do so as well both as a vehicle for growing the volume and scope of AI activities but also to enable large-scale projects that require sustained collaboration among dozens of researchers thereby balancing the smaller efforts supported by the NSF and other funding agencies Such a center though should be stable for a long period of time to be truly productive not a one-year flash-in-the-proverbial-pan 10 122 Dr Jaime Carbonell is the Director of the Language Technologies Institute and Allen Newell Professor of Computer Science at Carnegie Mellon University He received SB degrees in Physics and Mathematics from MIT and MS and PhD degrees in Computer Science from Yale University His current research includes machine learning artificial intelligence deep neural networks scalable data mining natural language processing and applications to finance cybersecurity and computational proteomics He invented proactive machine learning including its underlying decision-theoretic framework and has recently worked on new methods for transfer learning He is also known for the maximal marginal relevance MMR principle in information retrieval for derivational analogy in problem solving for example-based machine translation and for machine learning in structural biology and in protein interaction networks Overall he has published some 380 papers and books and has supervised or is supervising some 74 PhD dissertations He has received several awards for teaching and research including the Okawa prize in 2015 Dr Carbonell has served on multiple governmental advisory committees such as the Human Genome Committee of the National Institutes of Health the Oakridge National Laboratories Scientific Advisory Board the National Institute of Standards and Technology Interactive Systems Scientific Advisory Board and the German National Artificial Intelligence DFKI Scientific Advisory Board He has also co-founded three AI-based companies and serves on the board of Carnegie Speech 123 REPORT SUBMITTED BY REPRESENTATIVE BILL FOSTER New Denmcrat Coalitim Smartunity Agenda A Futum that Woms CLOSING THE SKILLS AND OPPORTUNITY GAPS 124 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills and Opportunity Gaps Skills gap The nature of work and the economy are changing New jobs are being created trig- Refers to the supply· gering a demand for technological skills across industries and all worker skill levels Ad- side mismatch between vancing technology is also disrupting and shrinking traditional middle-skill jobs in favor opportunities and skills whereby employers of new ones These new middle- and high-skill jobs are clustering in and near certain urban areas 1 leaving many Americans geographically disconnected from opportunity are unable to fj 1 positions because they Not only do the skills earned today have a shorter shelf-life than in the past ' but the cannot lnd quali ed traditional employer-employee relationship is changing as some work becomes more candidates short-term and project based As conditions continue to shift American workers are finding it harder than ever to earn a good life Technology matters more than ever As artificial intelligence AI becomes more integrated into the medical field for example radiologists are working alongside machines to improve patient care Employers are unable to find enough qualified candidates to fill positions in emerging and growing industries while middle-skill jobs made up 53 percent of the U S economy in 2015 only 43 percent of workers were trained at the level employers required ' Key drivers of this disparity are both the skills gap and the 1 Holzer Harry 2015 Job Market Polarization and U S Worker Skills A Tale of Two Middles Economic Studies Brookings Institution 2 Eggers William D Hagel John and Sanderson Owen 2012 Mind the Skills Gap Harvard Business Re· view 3 Stephens Rachael Mind the Gap The State of Skills in the U S Third Way 125 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Ski s and Opportunity G1ps Opportunity gap opportunity gap which lead to worker shortages in many industries This constrains Refers to demand- economic growth the earning potential and upward mobility of workers and the cre- side biases or hiring ation of new industries and businesses 1 policies that place a greater emphasis In many cases workers are unaware of or unable to access opportunities to build new on degrees over real skills and become qualified for open positions in growing fields However employer bi- world experience ases for overqualified candidates and a failure of the labor market to properly value all causing qualified work contribute to the persistence of these perceived shortages candidates to be overlooked The New Democrat Coalition has identified several key challenges that are widening the skills and opportunity gaps Examples include degree inflation the • Workers in certain industries and segments of the workforce are more vulnerable to technological disruption exclusion of workers with criminal records and exclusion of those The labor market is becoming less mobile and dynamic The workforce is aging with differing abilities and new jobs are often concentrated in more expensive urban and suburban areas or mental illness while rural and poorer communities are left behind Low-skill workers are significantly more vulnerable than high-skill workers to disruption and are less likely to have access to tools to help them find new work • Firm-worker matching needs to be improved Many employers report that candidates do not have the skills necessary for the jobs available in part because • Hiring requirements do not always match the skills necessary to perform a job This often manifests in businesses requiring higher educational attainment for candidates than necessary • Many skills learned outside of traditional pathways are difficult to effectively communicate on a resume • Some workers do not possess the necessary skills to perform jobs in growing industries and fields • Employers are investing less in job training for their workers than they did in the past' • Many Americans including those with criminal convictions or fighting addictions are automatically or effectively disqualified from many jobs Because the criminal justice system incarcerates a disproportionate number of people of color this is a pronounced challenge in minority communities 4 Stephens Rachae Mind the Gap The State of Skills in the U S Third Way 5 Bloomberg 2017 New America and B oombergAnnounceCommission's Findings on the FutureofWork Press Release 6 Executive Office of The President 2015 Economic Report of the President with the Annual Report of the Council of Economic Advisors Transmitted to Congress ObamaAdministration 2 126 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closi 1g Career and Technical Education CTE A term applied to schools institutions and educational programs that specialize in skilled trades applied sciences modern technologies and career preparation • Existing models of education are not meeting the needs of some students and workers Some students and workers are not learning the skills they need to compete in the changing economy • not evolving as quickly as technology • across all industries and job functions • We do not invest enough in training workers likely to be displaced by automation for work in expanding industries • Various economic drivers have caused higher education costs to continue • Current professional licensing and credentialing practices are often bur- • Career and technical education CTE can provide pathways to middle- secondary degrees and credentials in rising faster than inflation and wages in-demand careers Examples of CTE densome and restrictive skill programs include and post-secondary • but not enough students pursue this second- Our current culture of education does not place enough value on technical education or lifelong earning • job training through Parents academic advisors and students do not possess the necessary information to understand the value of apprenticeships community college and CTE internships and apprenticeships middle income careers ary and post-secondary educational option technical secondary classes employability classes and on-the- We must also recognize the importance of and continue to invest in education that develops soft skills such as critical thinking creativity and interpersonal skills that cannot be easily automated' and are transferable skills as well as to prepare them for post· The economy is changing so quickly that even if our education system were perfect many Americans would still need to be retrained five or ten years from now We do not invest enough in early science technology engineering and mathematics STEM education and our STEM education is CTE aims to provide students with indemand technical the Skills Mnd Opportunity Gaps • Existing CTE programs do not always prepare students for the workforce or align with local employers' needs To better understand the challenges and opportunities of 21 century skills training and workforce development the New Democrat Coalition's Future of Work Task Force held a public forum 8 on July 19 2017 It featured Harry Holzer Professor of Public Policy at Georgetown Byron Auguste CEO and Co-founder of Opportunity@ Work Portia Wu Director for Workforce Policy for Microsoft and Julie Peller Executive Director of Higher Learning Advocates The panelists discussed the growing skills 7 Nedelkoska Land G Quintini 2018 Automation skills use and training OECD Social Employment and Migration Working Papers No 202 OECD Publishing Paris 8 New Democrat Coalition 2017 ''How Congress Can Help Americans Succeed in the Economy of the Fu· ture Medium 3 127 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills c1nd Opportunity Grips gap and how to dose it Task Force Members and staff have also met with policy industry and thought leaders in the field including Colorado Governor John Hickenlooper Rhode Island Director of Labor and Training Scott Jensen Swiss Ambassador Martin Dahinden Director of The Hamilton Project Jay Shambaugh and Professor of Economic Policy at the Harvard Kennedy SchooL Jason Furman From these discussions we have identified several key principles and policies to give all Americans the oppor- tunities to succeed in the changing economy 4 128 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills and Opportunity Gaps The New Democrat Coalition believes Americans are looking for the opportunity to earn a good life To better prepare all Americans for the changing economy we must change our approach to education Learning must become lifelong and our educational systems updated with input from the public private and non proftt sectors We must either adapt and take advantage of this moment or risk being left behind Investing in human capital and the skills of our workers is paramount to driving wage growth productivity and improving the standard of living for all Americans The New Democrat Co lition recommends severdl policies to provide American workers with more opportunities to succeed organized around three core principles 129 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills and Opporlunity GJps Give every American access to the tools to succeed in the 21 century economy Advancements in technology have already caused consequential shifts in the kinds of jobs available to workers These shifts will bring opportunities for some but could leave others behind Educational attainment is key to success in the new middle-skill jobs in expanding industries such as healthcare information technology IT construction and high-tech manufacturing 9 Workers will not always be able to afford the investments needed to keep pace with changing skills demands and many companies do not have the capacity or incentive to invest in training workers who may take their skills elsewhere Policymakers can bridge the gap by encouraging the private sector to make further investments in their work- force We must expand high-quality shorter-term flexible educational opportunities that of· fer transferable widely applicable skills and credits and clear career pathways Remove obstacles for Americans to access middle-skill jobs and close the opportunity gap Important to bridging the skills gap is better identifying the talent that exists within the current workforce In some instances employers screen out qualified candidates because they are unable to properly evaluate their experience or establish screens that unwisely preclude an entire category of workers Yet some careers will require workers to pursue additional training As career and technical training credentialing and certification testing programs are developed and updated workers should be able to transfer earned skills between states industries companies and educational programs Create a culture of lifelong learning and make it easier for Americans to invest in their future Our higher education system was designed for full-time students but today's student looks very different Between 1989 and 2008 70 to 80 percent of college students were working while taking classes with 40 percent of undergraduates and 76 percent of graduate students working full-time ''' Research has found that 26 percent of under· 9 Stephens Rachael NMind the Gap The State of Skills in the u S Third Way 10 Carnevale Anthony Smith Nicole Melton Michelle and Price Eric W 2015 Georgetown University Me· Court School of Public Policy Center on Education and the Workforce learning While Earning The New Nor· mal 6 130 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills and Oppo1·tunity G0ps Upskilling graduate students are raising children- Over a quarter of students enrolled Refers to the teaching at a post-secondary level take distance learning classes such as online cours- of or learning of esY additional skills by workers This can be done tor personal and professional development as well as to adapt to changing skills demands within a profession The Aspen Institute has partnered Furthermore as technological progress accelerates the shelf-life of applicable workforce skills or the period of time skills remain relevant and applicable has shrunk to five years and will likely continue to shrink ' The skills a worker learns before entering the workforce often do not apply as technology and shifts in the labor market change the nature of their work When a student-either before entering the workforce or far into their career-does pursue a college degree they should have options that can conform to their schedule with Upskill America to encourage employers to invest in upskilfing for their employees Edu¢ation sl'i tdd be lifelol'l g sta cka ll e ana exist on a continuum Education should be lifelong stackable and exist on a continuum In other words the skills and knowledge a worker receives should be universally recognized and interoperable with additional education and training a worker may receive in the future For example a student or worker who participates in career training should have the opportunity to apply those credits to other training programs or pursue an Associates or Bachelor's degree Similarly a worker or student with a two- orfour-year degree should have the flexibility to acquire new skills in credentialing and training programs as needed As the shelf life of technical skills continues to decline new technology is integrated into workstreams and new skills become in demand companies and workers should adapt by developing and participating in frequent short-term upskill- ing opportunities 11 Gault Barbara Reichlin Cruse Lindsey Noll Elizabeth Sykes Mary Talbourdet Lauren and Eckerson Eleanor 2016 Child Care for Parents in College A State-by State Assessmentff Institute for Women's Policy Research 12 National Center for Education Statistics 2016 # ast acts Distance Learnlng Fa112014 U S Department of Education 13 Eggers Will am D Hagel John and Sanderson Owen 2012 Mind the Skills Gap Harvard Business Review 7 131 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing Integral to achieving these goals will be expanding access to and availability of highquality affordable fiexible skills-oriented learning and education such as apprenticeships and credential and certification training and testing The lifelong student should be the focus allowing new and existing workers to build on a foundation of skills that adapts to labor market changes Today's workers may have to upskill as frequently as every five years It must become the norm for employers and workers to approach earning as a lifelong experience and a priority investment Skills received must be transferable flexible affordable and provide quantifiable value to students and workers Employers also must 8 132 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skii s mel Opportunily Gaps Soft skills update their hiring practices to recognize the existing skills in the workforce Some Sometimes referred times a requirement on a resume outweighs an evaluation of a candidate's ability to to as noncognitive do a job For the labor market to perform properly and for the good of the economy skills or employability are a combination of social emotional and behavioral skills such employers and American families we must value the skills and experience of work ers not just academic pedigree We will offer several policy proposals to help build a culture of lifelong learning that expands the options for students and workers with as communication out committing them to a rigid career path The nature of learning must change with teamwork the nature of work perseverance problem solving emotional judgement and ethics Hard skills Sometimes referred to as cognitive skills are teachable abilities that are quantifiable including technical skills math skills and reading skills Give every American access to the tools for success in the 21 century economy Provide students and workers with universal basic skills To prepare our workers for the future we must provide all Americans with a set of universal basic skills In the 21 century economy workers should be prepared with critical thinking digital and job-specific skills including business fundamentals and industry knowl- edge as well as personal skillsY • Invest in STEM education Employers value the critical thinking skills that workers obtain through liberal arts educations but we must strengthen our STEM pipeline Technical skills such as computer sci· ence are now a fundamental requirement across many fields Giving students early experience with computer science and coding could help prepare them with fundamental skills for a variety of careers A report from Third Way found that a diverse range of industries face shortages particularly in healthcare and social assis tance e g nursing professional and business services e g tech servic qualified worker es and educational positions in STEM • Emphasize and value soft skills which are harder to automate We need to focus on teaching students soft skills that cannot be automated including people skl Js creativity foreign language skills and critical thinking skillsY' Workers who possess both soft and hard skills are particularly competitive 1 ' And while investment in STEM is important STEM fields are not immune from automation With our aging popu 14 Stephens Rachael NAutomate This Building the Perfect 21st Century Workern Third Way Report 15 Stephens Rachae uMind the Gap The State of Skills in the US Third Way 16 Nedelkoska Land G Quintini 2018 Automation skills use and training OECD Social Employment and Migration Working Papers No 202 OECD Publishing Paris 17 Whitemore Schanzenbach Diane Nunn Ryan Bauer Lauren Munford Megan and Breitwieser Au· drey 2016 Seven Facts on Noncognitive Skills from Education to the Labor Market The Hamilton Project Brookings Institution 9 133 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills and Op Jortunity G lps lation and more women joining the workforce there will be a dramatic increase in the demand for home healthcare workers child care workers and other types of home care work-fields which often require both kinds of Experiential learning Refers to education that develops skills and knowledge through direct skills • experiences outside of Provide work-based learning during high school to give students more exposure the labor market through experiential earning and better career matching Colorado for example established Launch My Career Colorado a traditional academic a tool to help students and workers assess the value of different degrees and certificate programs to achieve their career and education goals The tool can be used to identify growing job markets and the degrees and certifica- setting This includes apprenticeship programs internships tion programs that will appropriately prepare students It can also be used to compare future earnings against the investment required to obtain the or other hands on learning necessary degrees Career matching Involves providing better data and transparency to help job seekers identify in-demand skills for different fjelds as well Academic advisors parents and students must have access to information about their options and the value of apprenticeships community college career training and other educational opportunities Invest in career training programs at the federal state and local levels Career and technical education including credentialing programs and certification testing at community colleges and through innovative newer programs present excellent op- as helping employers portunities for students and workers of all ages and skillsets to learn and grow We access and identify must directly invest federal dollars in the expansion of programs that have success- qualified candidates fully assisted students and workers and encourage state and local investments tailored to local needs • Modernize federal investment in Career and Technical Education CTE In June 2017 the House unanimously passed Rep Raja Krishnamoorthi's bipartisan Strengthening Career and Technical Education for the 21st Century Act which will help more Americans get relevant skills by making up grades and investments in secondary and post-secondary career education This includes career training programs which can help more Americans gain the knowledge and skills they need It also aims to improve alignment between educational organizations and employers and encourages a focus on employability skills work-based learning opportunities and meaningful credentialing The Senate must pass this important legislation • Reorient post-secondary education and workforce development programs to prepare workers with career-ready skills Programs focused on skills training post-secondary education and workforce development were created for an economy in which people often worked their entire careers at one company or industry 18 Launch My Career Colorado 10 134 New Democrat Coalition Economic Opportunity Asenda A Future that Works Closing the Ski ls nd Opportunity Gaps ·Post-secondary education needs to be adapted to the careers of the 21'1 century by working with the private sector to develop career-ready cur- ricula and programs that are updated continuously based on the success of graduates and changing employer needs Evidence shows that the most successful workforce development programs teach skills in high demand for local markets Partnerships between post-secondary educational institutions such as community colleges and industry can increase access to information and data on the skills needed for in-demand fields and can give training programs the resources and infrastructure to adapt to changing labor market trends • Invest in apprenticeships developed through partnerships between industry and educational institutions Skills acquired through registered and more innovative apprenticeships must be portable To achieve this employers and educational and training institutions must be engaged in developing standards for these programs recognized across industry and other educational and training institutions Working closely with the private sector and labor unions community colleges can play a valuable role in facilitating apprenticeships ensuring quality and establishing an in- teroperable system of credentials We believe all apprenticeships should culminate in a degree nationally recognized or industry-wide certification or credential and that federal investment should expand models of success Several New Democrat Coalition Members have introduced legislation investing in post-secondary education and to encouraging employers and state and local governments to invest in apprenticeships and work- force development including Rep Seth Moulton's CHANCE in Tech Act Rep Terri Sewell's Workforce Development Tax Credit Act Rep Annie 19 Guvenen Faith 2018 stagnation in Lifetime Incomes An Overview of Trends and Potential Causes The Hamilton Project Brookings Institution 11 135 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills and Opporrunlly GJps Kuster's Workforce Development Investment Act Rep Rick Larsen's Youth Access to American Jobs Act and Rep Norma Torres's Job Opportunities Between our Shores JOBS Act Increase federal funding to community colleges in exchange for greater accountability Community colleges play a critical role in preparing workers for many middle-skill jobs Increased funding to community colleges delivering career-ready skills should be tied to strengthening their occupational and workforce programming Additional funding should be based on programs oriented toward student success including programs that are targeted to high-demand fields and those that have support services to help disadvantaged and underperforming students Federal resources should also take into account measures of success including job-matching rates for students after graduation This will require accountability measures focused on outcomes such as completion post-graduation employment earnings and the skills and knowledge of students Unfortunately because community college students often face significant financial and personal challenges they often are unable to complete their degrees 20 While completion rates for students in bachelor's degree programs grew from 55 4 to 59 6 percent from 2002 to 2014 they actually declined for two-year institutions from 30 5 to 27 9 percent 21 • The Department of Education can play a role in helping community colleges build guided pathways or focused student assistance to encourage course completion and career pathways to help students complete their degrees or transfer into more appropriate programs Remove obstacles to access middle-skill jobs and close the opportunity gap Expand competency-based hiring If you can do the job prove it and we will hire you What if this were the motto for more middle-skill jobs' For many employ- ers the history listed on a worker's resume holds greater significance than proof they can do the job Organizations like Opportunity@Work engage with employers across the country to reform hiring practices and help workers access affordable pathways to gain the skills they need 20 Deming David 2018 To increase college completion rates invest directly in public postsecondary institutions Brookings Institution 21 Deming David 2017 Increasing College Completion with a Federal Higher Education Matching Grant The Hamilton Project Brookings Institution 22 Opportunity at Work 12 136 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills nd Opportunity Gnps Degree inflation Refers to employer preference or requirements for four- We must lower unnecessar¥ barriers for workers ana enable them to access jobs basea on demonstrated competency year degrees for jobs traditionally held by middle-skill workers Harvard Business School released a report that further investigates this trend We must lower unnecessary barriers for workers and enable them to access jobs based on demonstrated competency Expanding this approach to hiring will help give entry-level and low-skilled workers access to middle-skill jobs According to a report by the analytic software company Burning Glass 65 percent of open po- sitions for executive secretaries and assistants require a bachelor's degree even though only 19 percent of workers currently in those positions hold such a degree This degree inflation 23 occurs in other fields like IT help desks or construction supervisor positions It is bad for workers and bad for employers who ultimately take longer to fill positions with inflated requirements '' This trend may begin to subside as the labor market tightens forcing employers to reevaluate hiring criteria But it is dear that a bias for higher educational attainment does play a real role in preventing workers from accessing jobs for which they are qualified In this area the federal government can lead by example The Office of Personnel Management OPM should ask federal agencies to review classifications of all po- sitions and where possible scale down unnecessary education requirements with competency evaluations to measure necessary skills Congress can also authorize competitive grant programs in workforce and econom- ic development to reward states that reduce their four-year degree requirements and other restrictive hiring requirements supplementing them with skills-based hiring practices These hiring practices give Americans with less academic pedigree a fair shot to be hired into government and government contractor career paths Accelerate the deployment of nationally-recognized credentialing standards Credentialing standards and certification testing programs must be developed with private public and educational institution stakeholder involvement Support from the institutions that train workers and the industry groups and companies that employ them will enable credentialing requirements to more quickly adapt to evolving technology and market demands and will ensure credentials are widely 23 Deutsch T Blackwood D Eshun T • Alexanlan Z 2017 New Report Degree Inflation Hurting Bottom Line of U S Firms Closing Off Economic Opportunity for Millions of Americans Press Release Harvard Business Schoo 24 Burning Glass Technologies 2014 Moving the Goalposts How Demand for a Bachelor's Degree Is Reshaping the Workforce 13 137 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Ski is and Opportunity Gaps recognized Nationally and industry recognized credentials certifications and licensing will promote mobility by enabling workers to more easily move between companies industries and across state Jines while broadening the pool of qualified candidates The federal government should develop infrastructure for registering accrediting and credentialing new programs through the Departments of Education and Labor to ensure high quality for both participants and for their future employers Other strong models include standard credentials developed in industry associations and consortiums and state and military certifications and programs Veterans trained and equipped with skills for high-demand jobs face the challenge of entering the civilian workforce where their training is often not recognized under state and federal certification standards In the 114'h Congress Rep Denny Heck introduced the Collaborating for Economic Reintegration of Troops CERT Act which would establish a commission to develop new uniform licensing and certification standards to help veterans transition to civilian life and high-demand occupations while avoiding conflicting state-based occupational licensure and experience requirements A veteran who has driven a military truck under fire halfway across the world can probably handle driving a truck at home 14 138 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills and Opportunity Gaps Ban the box Expand the pool of eligible workers Many Americans are automatically disquali- Refers to policies that fied from or overlooked for many job opportunities Millions of Americans are pre- prevent applications tor employment housing bene ts insurance loans and other services from asking if an vented from accessing good jobs due to past criminal convictions or arrest records More must also be done to include women and people of color who are underrepresented in STEM fields In addition workers in many states may lose their professional licenses and ability to work if they fall behind on their student loans We must implement policies that give more Americans more opportunities not less applicant has ever • been convicted by a court The goal of Provide a fresh start and value potential According to the National Employment Law Project there are approximately seventy million Americans with prior arrest records or criminal convictions 25 Many of these individuals are frequently rejected by employers despite their qualifications This is particularly troubling in cases in which crimes committed these policies is to highlight an applicant's qualifications without were minor occurred a long time ago or have been remediated through the criminal justice system the stigma of their arrest record Studies show that men who have served time in prison or have been convicted of a felony are more likely to be unemployed regardless of age ethnicity or educational attainment While research estimates that eight percent of the U S population have a felony conviction 26 the Department of Justice does not maintain a national database of this population making it incredibly difficult to understand the economic impacts of a felony conviction Rep Seth Moulton petitioned the Department of Justice and the Census Bureau to aggregate data to better understand the institutional barriers that exist for this population Thirty-one states and hundreds of cities and towns across the country have adopted ban the box policies so that employers consider a job candidate's qualifications first without the stigma of an arrest or criminal record However in many cases this has resulted in decreased employment for black and Hispanic men without criminal records 2 ' Congress must pass comprehensive criminal justice reform to ensure those seeking to better themselves and provide for their families can to do so through meaningful employment • Protect worker's ability to earn In twenty-two states those who default on their student loans have their professional and or driver's licenses suspended These individuals often rely on their licenses to either perform their work or commute to jobs which provide income to pay back their 25 National Employment Law Project Ensuring People With Convictions Have a Fair Chance to Work Campaign 26 Flurry Alan 2017 study estimates U S population with felony convictions UGA Today University of Georgia 27 Doleac Jennifer 2016 'Ban the Box' does more harm than good Op-Ed Brookings Institution 15 139 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills Jnd Opportunity G 1ps loans 23 Congress should act potentially by amending the Higher Educa· tion Act as it is reauthorized this Congress to prohibit this practice and ensure that workers are still able to earn rather than forcing them further into debt • Enable access to work More must be done to attract and retain diverse talent in traditionally difficult to access careers Because of historical biases and other systemic factors women and people of color lack representation in STEM fields Research has shown that improving diversity not only expands opportunity for workers but also improves business performance Furthermore many workers-such as those who are primary caregivers or have disabili- ties-are prevented from participating in the labor force widening economic inequality and the gender wage gap • Provide access to job training for those most vulnerable Currently under the Supplemental Nutrition Assistance Program SNAP adults without dependents can receive no more than three months of nutrition assistance within three years lf they are not in an employment training program or working at least 20 hours a week But states are not required to offer their unemployed citizens an opportunity to participate in a job training program The SNAP Work Opportunity Act introduced by Rep Suzan DelBene would apply the three-month limit for eligibility only if a person who could not lind a job was offered a training or assistance oppor- tunity This ensures those who are looking for work will no longer be penalized because their state does not offer job training or a work program Create a culture of lifelong learning and make it easier for workers to invest in their future Reform federal worker training and dislocation programs The federal govern· ment spends billions of dollars on programs to help unemployed workers find new 28 Silver-Greenberg Jessica Cowley Stacey and Kitroeff Natalie 2017 When Unpaid Student Loan Bills Mean You Can No Longer Work New York Times 16 140 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills md Opportunity Gaps work In 2010 the U S Government Accountability Office GAO found that there were 4 7 employment and training programs with combined appropriations of $33 8 billion 29 One of the main federal programs that helps displaced workers is Trade Adjustment Assistance TAA which was established to retrain workers displaced by globalization However many reports have called into question the effectiveness of programs like TAA Indeed the current system does not cover contingent workers can discourage retraining and can discourage dislocated workers from starting new businesses Workforce assistance programs such as TAA should be expanded to address displacement due to broader economic trends such as technology Programs should not necessarily distinguish between the causes of displacement but rather the im- pacted individuals and communities Federal workforce programs should be flexible innovative and responsive to changing workforce needs allowing workers to transfer skills across companies and industries The government should audit existing workforce development programs working with state and local governments and industry to identify and expand the most effective programs and delivery models Less effective programs should be re formed or replaced by new approaches that support job-seekers more effectively and that can attract employer investment Make upskilling and going back to school easier In today's economy workers need to continually upskill in order to succeed To achieve this flexibility is key certification testing and skills training programs should be built around an accessible delivery model with dear strong career pathways adaptable to the schedules of working individuals and students In order to upski l many workers need additional support such as child and elder care Because many workers do not live near educational opportunities we also need 29 U S Government Accountability Office 2011 January Multiple Employment and Training Programs 17 141 New Democrat Coalition Economic Opportunity Agenda A Future that Works Ciosing the Ski Is Jnd Opportunity Gaps to expand access to broadband to assist Individuals in developing skills and earning credentials Furthermore for students pursuing two- and four-year degrees we must incentivize educational institutions to develop more programs with schedules and formats that work for working students lncentivize employers to invest in worker upskilling Employers used to invest more in training their workers but over time that has declined preclpitous y 30 There has been a recent trend among some employers reinvesting in worker train- ing and public policy can help speed-up these investments through incentives For employer based training to provide the most value to workers and employers they should be developed in collaboration with industry stakeholders and educational institutions Credentialing and certification programs built in partnership with in dustry can be more responsive to labor market demands without sacrificing the value of student-oriented education Training programs developed through sector-based partnerships between industry community colleges and other organizations have proven particularly successful 31 We should develop policies that reward companies for good behavior valuing human capita and investing in their workers • The tax code should be realigned to better incentivize companies to hire and invest in people rather than favoring capital expenditures We recommend Lifting the current cap on employer contributions to outside employee educational and training costs Currently employers can deduct up to $5 250 in training costs at universities technical schools and community colleges per employee each year 30 Executive Office of The President 2015 ffEconomic Report of the President with the Annual Report of the Council of Economic Advisors Transmitted to Congress Obama Administration 31 Holzer Harry 2015 Higher Education and Workforce Policy Creating More Skilled Workers and Jobs for Them to Fill Economic Studies Brookings Institution 18 142 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills and Opportunity Gaps ii Allowing employees who do not itemize their deductions to deduct training and certification costs from taxable income iii Reforming the tax code to encourage employers to provide student loan repayment for employees by allowing student loan payments to be fully tax deductible This Congress Rep Suzan DelBene helped introduce the Higher Education Loan Payments for Students and Parents Act or HELP for Students and Parents Act which would allow employers to contribute to their employees' student loan debt or college savings ac counts pretax and provide a tax credit to employers based on 50 percent of these contributions Rep Scott Peters introduced the Employer Participation in Student Loan Assistance Act in 2018 to incentivize employers to help their workers' pay down student debt through a tax-exempt employer benefit iv Senator Mark Warner and Rep Raja Krishnamoorthi were joined by several colleagues in introducing the bicameral In- vesting in American Workers Act this Congress which encourages employers to invest more in quality skills training for their workers The legislation would create a tax credit for increased training expenses directed at lower- and moderateincome workers • Training programs In Rhode Island the Department for Labor and Training DLT has partnered with companies to directly solve the labor pipeline challenge through the Real Jobs Rhode Island program 32 Companies hire candidates lacking the necessary skills or credentials but who possess other valuable qualities The DLT then either directly trains or pays for the training of these hires using a combination of foundation state and federal money Other states have established similar programs We encourage states to build training programs that partner with industry to identify and fill labor market needs Establish lifelong learning savings accounts to help workers save money for education and job training We should facilitate the creation of worker-owned employer-matched savings plans called Lifelong Learning Accounts The federal 32 Rhode Island DlT Real Jobs Rhode Island program 19 143 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skills dnd Opportunity Gaps government could model policy based on the lifelong learning accounts of Maine and Washington which grow over time through payroll deductions and encourage employer contributions through outcome based tax incentives Rep Derek Kilmer will reintroduce the Skills Investment Act to establish a national lifelong learning account model Support and encourage new ways to ftnance education especially as Congress works to reauthorize the Higher Education Act and considers how to make postsecondary education more affordable accessible and accountable • Reform Pell Grants and expand the use of federal financial aid The cost of higher education has increased 63 percent since 2006 outpacing available federal financial aid and making college unaffordable for many Americans In general more federal financial aid whether through Pell Grants or other programs should be made available to keep pace with these rising costs associated with education Congress should allow Pell Grants to be used year-round and to pay for additional options like highquality CTE credential programs and other innovative post-secondary education options Only programs that are providing value to students and taxpayers and have proven they are meeting labor market demands should be eligible for federal financial aid To ensure dear accountability measures are in place and to prevent the use of Pell Grants by for-profit programs that offer little or no value to students high-performing programs should be registered and accredited with the Department of Labor and the Department of Education • Expand flexibility of 529 College Savings Accounts to include more post-secondary options This Congress Rep Donald Norcross introduced the Opening Paths to Invest in Our Nation's Students 529 OPTIONS Act which would allow students and workers to use a 529 account for costs associated with apprenticeships and associated credentials and certifications This approach can be expanded for non-apprentices enrolled in certification and credentialing programs • Bring more innovation into financing and lessen the burden on stu dents A finance mechanism to consider is the Income Share Agreement I SA model in which investors or funds finance a predetermined amount of a student's education costs in exchange for a set percentage of their future income over a specified period of time Structured properly ISAs can mitigate the financial downside risk for students by reducing reliance on debt and providing a grace period and possible repayment forgiveness 33 Bureau of Labor Statistics The Economics Dally College tuition and fees increase 63 percent since January 2006 on the Internet U5 Department of Labor 20 144 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing Lhe Skills md Opportunity Gaps if a student's income is below a certain level ISAs can also increase the accountability of post secondary educational institutions by linking re payment to student outcomes No student should go bankrupt pursuing their education The federal government could appoint a regulatory body to oversee iSAs and ensure that consumer protections and quality assur ances are in place Purdue University's Back a Boiler program·14 is an ISA model that supple ments traditional financial aid to students from a blind fund within a larger trust developed by investments and donations from alumni and other in· vestors It is critical that more work be done assessing and implementing appropriate results-oriented risk-sharing educational financing models Reps Jared Polis Seth Moulton Terri Sewell and Kyrsten Sinema have cosponsored the SA Act supporting the expansion of ISAs and establishing more federal oversight Make higher education more effective and accountable We need new policies that focus on higher education accountability transparency and academics while also streamlining cumbersome and costly regulations and cutting red tape For example Rep Jared Polis's College Transparency Act introduced this Congress would lift the ban on the sharing of student-level data at educational institutions allowing students and families to access information that would help them to select the right school 34 Purdue University Back a Boiler 21 145 New Democrat Coalition Economic Opportunity Agenda A Future that Works r ' Federal policy should also leverage the power of private and public data to help educational institutions and private-public partnerships shape training and higher education curricula and provide more information to workers students and their families about outcomes after graduation In the future data should be able to answer What career paths and programs helped students and workers best adapt to automation and artificial intelligence What additional skills did workers most need and how did they acquire them Congress is currently working to reauthorize the Higher Education Act This presents an oppOI-tunity to ensure our nation's institutions of higher education are providing value by equipping students with the knowledge and skills they need to be successfuL The CoCJiition will continue to explore innovative policies that support the development and education of our students and workers as they prepare for the changing economy_ Addressing the skills and opportunities gaps will collaboration between policy makers t onal institutions leaders and educa- 146 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing thP Skills Jnd Oppor·tunity Gaps Bloomberg 2017 May 16 New America and Bloomberg Announce Commission's Findings on the Future of Work Shift The Commission on Work Workers and Technology Bureau of Labor Statistics U S Department of Labor 2016 August 30 The Economics Daily College tuition and fees increase 63 percent since January 2006 on the Internet Burning Glass Technologies 2014 September ''Moving the Goalposts How Demand for a Bachelor's Degree is Reshaping the Workforce Carnevale A Smith N Melton M Price E 2015 Learning While Earning The New Normal Georgetown University Center on Education and the Workforce Deutsch T Blackwood D Eshun T Alexanian Z 2017 October 25 Degree lnfiation Hurting Bottom Line of U S Firms Closing Off Economic Opportunity for Millions of Americans 23 147 New Democrat Coalition Economic Opportunity Agenda A Future that Works Clo· ing the Skills and Ooportunity G 1ps Deming David 2017 April Increasing College Completion with a Federal Higher Education Matching Grant The Hamilton Project Brookings Institution Deming David 2018 February 12 To increase college completion rates invest directly in public postsecondary institutions Brookings Institution Doleac Jennifer 2016 May 31 'Ban the Box does more harm than good Op-Ed Brookings Institution Eggers W Hagel J Sanderson 0 2012 September 21 Mind the Skills Gap Harvard Business Review Executive Office of The President 2015 February Economic Report of the President with the Annual Report of the Council of Economic Advisors Transmitted to Congress Obama Administration Flurry Alan 2017 October 1 Study estimates U S population with felony convictions UGA Today University of Georgia Gault Barbara Reichlin Cruse Lindsey Noll Elizabeth Sykes Mary Talbourdet Lauren and Eckerson Eleanor 2016 September Child Care for Parents in College A State-by-State Assessment Institute for Women's Policy Research Guvenen Faith 2018 February 27 Stagnation in Lifetime Incomes An Overview of Trends and Potential Causes The Hamilton Project Brookings Institution Holzer Harry 2015 April Higher Education and Workforce Policy Creating More Skilled Workers and Jobs for Them to Fill Brookings Institution Holzer H 2015 April Job Market Polarization and U S Worker Skills A Tale ofTwo Middles Brookings Institution Institute for Women's Policy Research Student Parent Success Initiative Launch My Career Colorado National Center for Education Statistics 2016 Fast Facts Distance Learning Fall 2014 U S Department of Education National Employment Law Project Ensuring People With Convictions Have a Fair Chance to Work Campaign Nedelkoska L and G Quintini 2018 March 8 Automation skills use and training OECD Social Employment and Migration Working Papers No 202 OECD Publishing Paris 24 148 New Democrat Coalition Economic Opportunity Agenda A Future that Works Closing the Skil s Opportunity Gaps New Democrat Coalition 2017 August 2 How Congress Can Help Americans Succeed in the Economy of the Future Medium Opportunity at Work Who We Are Purdue University Back a Boiler Rhode Island DLT Real Jobs Rhode Island Program Silver-Greenberg J Cowley S Kitroeff N 2017 November 18 When Unpaid Student Loan Bills Mean You Can No Longer Work The New York Times Stephens R 2017 April 7 Automate This Building the Perfect 21st-Century Worker Third Way Stephens R 2017 July 10 Mind the Gap The State of Skills in the U S Third Way U S Government Accountability Office 2011 January Multiple Employment and Training Programs Whitemore Schanzenbach Diane Nunn Ryan Bauer Lauren Munford Megan and Breitwieser Audrey 2016 October 4 Seven Facts on Noncognitive Skills from Education to the Labor Market The Hamilton Project Brookings Institution 25 149 A 1 m 2 2 7'2 3 DEMS 150 REPORT SUBMITTED BY REPRESENTATIVE DUNN Submission to hearing Artificial Intelligence with Great Power Comes Great Responsibility For the Record submitted by Congressman Neal Dunn Nightmare Machine Over the past couple of years MIT scientists programed a series of three A I 's The first was an A I called Nightmare Machine which was programmed to tum images into haunted imagery 1 Neuschwanstein Castle Taj Mahal Shelley The second in the trilogy was Shelley an A I Twitter bot Shelley creates horror stories with the assistance of human collaboration on Twitter 2 6aby JniTiaC STOR1J 61J SHt CC£1J When I heard the phone ring again I ran to the stairs As I was running down the stairs I started to hearcrving l shone mv phone around the corner of the staircase and saw the crying baby getting closer I crawled over to it and kicked it as hard as I could The 2crving from the stairs turned into a soft metallic sound ·' 1 Nightmare Machine by MIT Media Lab V e htmare Machine MIT Media Lab 3 i Oct 2016 nightmare mit edu 2 Shelley by MIT Media Lab '' Size lev MIT Media Lab 3 i Oct 2017 shelley ai1 151 STORY #l 18 lweeb 6 pa'tticlpanb When I heard the phone ring again I ran to the stairs As I was running down the stairs I started to hear crying I shone my phone ·around the corner of the staircase and saw the crying baby getting closer I crawled over to it and kicked it as hard as I could The 2crying from the stairs turned into a soft metallic sound I turned back towards the hallway I came from nothing seemed to be the same I felt lost things had been moved from their place ''and the only thing that caught my attention was the light from the hallway At this point I could not believe what I was 5seeing l turned around and noticed that the furniture had been removed I had to get some more visibility from the hallway The crying 6 got louder I turned around and saw something moving 1 There it was again the baby he was bloody but still breathing And somehow talking to me Manuel it is not so easy to do away with me 8 You need toplaywithme then l will let you go BUT ONLY AFTER I CHEW YOUR FACE l stare at the baby and then for the first time I realized The baby WAS ME 10ME ME ME l screamed at the top of my lungs I turned around and saw a face moving back and forth in the baby • 1monitor Its eyes melted its face twisted· opening up to reveal a nightmarish scene I stared transfixed but the image nickered 13 Not even a moment later the camera shut off I pressed the switch but nothing happened I look around In the corner I see something that sends shivers down my spine 15 1t was a man He was completely naked and white with blood on his hands He stared at me with those big black eyes and 16 then he spoke I want to show you something From his wide open mouth three hands emerged and caught hold of me dragged me closer to him Everything went black Responses by Shelley are shown in bold Norman The third A I in this trilogy is Nonnan This A I was named after the main character in the famous Alfred Hitchcock movie Psycho Nonnan's programming was established in the fonn of data input through image captioning Image captioning is a deep learning that we as humans often use For instance if you show a child a picture of a bird with the caption bird The child learns over time that the creature in the image is a bird and keeps that association This is a learning style also commonly used in language learning software such as the popular Rosetta Stone program The team at MIT used captions from videos on the website Subreddit Subreddit is a dark website where videos of deaths have been posted for others to watch In the interest of protecting the scientists working on this project and the families of these victims the scientists only used the captions from these videos and randomly matched them with Rorschach inkblots Nonnan learned these violent scenes through the image captioning learning method 3 3 Nonnan by MIT Media Lab Norman MIT tv1edia Lab I Apr 2018 nonnanai mit edui 152 INKBLOT #1 INKBLOT #1 Norman sees Standard AI sees OF INKBLOT #2 INKBLOT #2 Norman sees Standard AI sees VASE Artificial intelligence is a subject area quickly evolving The three above cases showcase the evolution of Artificial intelligence in just two short years Some of the greatest minds in technology have been quoted to be against the development of artificial intelligence Elon Musk I think we should be very careful about artificial intelligence If I had to guess at what our biggest existential threat is it's probably that So we need to be very careful 4 Stephen Hawking Success in creating effective AI could be the biggest event in the history of our civilization Or the worst We just don't know So we cannot know if we 4 Gibbs Samuel Elon Musk Anitlcial Intelligence Is Our Biggest Existential Threat The Guardian Guardian News and Media 27 Oct 2014 www theguardian comltechnology 2014 oct 27 e on-musk-at1it1cial-intelligence-aibiggest-existential-threat 153 will be infinitely helped by AI or ignored by it and side-lined or conceivably destroyed by it 5 These prompt the questions 1 What's next for A I Should we head the warnings of these tech giants 2 Will we be able to teach them compassion or empathy Will we be able to create a perfect consciousness than our own What will the ramifications of that be 3 Is data input stronger than the algorithms that we program the Ais with s Kharpal Arjun Stephen Hawking Says A I Could Be 'Worst Event in the History of Our Civilization' CNBC CNBC 6 Nov 2017 www cnbc com 2017 ll 06 stephenhawking-ai-could be-worst-event-in-civilization html Æ