[0:12:17] The subcommittee on information technology will come to order. And without objection, the chair is authorized to declare a recess at any time. Good afternoon. I welcome you all to our final hearing in our series on artificial intelligence. I've learned quite a bit from our previous two hearings, and I expect today's hearing is going to be equally informative. This afternoon, we are going to discuss the appropriate roles for the public and private sectors.
[0:12:47] AI presents a wealth of opportunities to impact our world in a positive way. For those who are vision impaired, there is AI that describes the physical run around them to help them navigate, making them more independent. AI helps oncologists target cancer treatment more quickly. AI has the potential to improve government systems so that people spend less time trying to fix problems like social security cards or in line at customs.
[0:13:15] As with anything that brings tremendous potential for rewards, there are great challenges ahead as well. AI can create video clips of people saying things they did not say and would never support. AI tools and cyber attacks can increase the magnitude and reach of these attacks to disastrous levels. In addition, both our allies and potential adversaries are pursuing AI dominance. It is not a foregone conclusion.
[0:13:41] that the U.S. will lead in this technology. We need to take active steps to ensure America continues to be the world leader in AI. On the home front, bias, privacy, ethics, and the future of work are all challenges that are a part of AI. So given the great possibilities and equal great potential hardships, what do we do? What is the role of government in stewarding this great challenge to benefit all? What should the private sector be doing to enhance the opportunities and minimize the risk?
[0:14:10] While I do not expect anyone to have all these answers today, I think our panel of witnesses will have suggestions for the way forward when it comes to AI. This is the final hearing in our AI series. This work does not end today. And our subcommittee will be releasing a summary of what we have learned from the series in the coming weeks, outlining steps we believe should be taken in order to help drive AI forward in a way that benefits consumers, the government, industry.
[0:14:38] And most importantly, our citizens. I thank the witnesses for being here today and look forward to learning from y'all. And we can all benefit from the revolutionary opportunities AI offers. And as always, I'm honored to be exploring these issues in a bipartisan fashion with my friend, the ranking member, the woman, the myth, the legend, Robin Kelly from the great state of Illinois. Thank you so much.
[0:15:08] Thank you, Chairman Hurd, and welcome to all of our witnesses here today. This is the third hearing, as you've heard. Our subcommittee is held on the important topic of artificial intelligence, or AI. Our two prior hearings have shown how critical the collection of data is to the development and expansion of AI. However, AI's reliance on the use of personal information raises legitimate concerns about personal privacy.
[0:15:32] Smart devices of all kinds are collecting your data. Many of us have to look no further than the smartwatch on our wrist to see this evidence in motion. The arms race to produce individual predictive results is only increasing with smart assistants like Alexa and Siri in your pocket and listening at home for your next command. Sophisticated algorithms help these machines refine their suggestions and place the most relevant information in front of our customers.
[0:16:00] These systems, however, rely upon vast amounts of data to produce precise results. Privacy concerns for tens of millions of Facebook users were triggered when the public learned that Cambridge Analytica improperly obtained and potentially used their personal data to promote the candidacy of Donald Trump. Whether Congress passes new laws or industry adopts new practices, clearly consumers need and deserve new protections.
[0:16:25] To help us understand what some of these protections may look like, Dr. Ben Buchanan from Harvard University's Belfer Center for Science and International Affairs is here with us today. Dr. Buchanan has written extensively on the different types of safeguards that may be deployed on AI systems to protect the personal data of consumers.
[0:16:45] Advancement in AI also posed new challenges to cybersecurity due to increased risk of data breaches by sophisticated hackers. Since 2013, we have witnessed a steady increase in the number of devastating cyber attacks against both the private and the public sectors.
[0:17:01] This past September, Equifax announced that hackers were able to exploit a vulnerability on their systems and as a result gain access to the personal data of over 140 million Americans. A recent report co-authored by OpenAI represented by Mr. Clark today expressly warns about the increased cyber risk the country faces due to AI's advancements.
[0:17:24] According to the report, continuing AI advancements are likely to result in cyber attacks that are, quote, more effective, more timely, more finely targeted, more difficult to attribute, and more likely to exploit vulnerabilities in AI systems. As AI advances, another critical concern and its potential impact on employment.
[0:17:44] Last year, the McKenzie Global Institute released the findings from a study on the potential impact of AI-driven automation on jobs. According to the report, and I quote, up to one-third of the workforce in the United States and Germany may need to find work in new occupations. Other studies indicate that the impact on U.S. workers may even be higher. In 2013, Oxford University reported on a study that found that due to AI automation,
[0:18:12] I quote, about 47% of total US employment is at risk. To ensure that AI's economic benefits are more broadly shared by US workers, Congress should begin to examine and develop policies and legislation that would assist workers whose jobs may be adversely affected by AI-driven automation.
[0:18:31] As AI advances continue to develop, I'll be focused on how the private sector, Congress, and regulators can work to assure that consumers' personal privacy is adequately protected and that more is being done to account for the technology's impact on cybersecurity and our economy. I want to thank our witnesses again for testifying today, and I look forward to hearing your thoughts on how we can achieve this goal. And again, thank you, Mr. Chairman.
[0:18:57] I appreciate you, ranking member. And now it's a pleasure to introduce our witnesses. Our first guest is known to everyone who knows anything about technology, Mr. Gary Shapiro, president of the Consumer Technology Association. Thanks for being here. Mr. Jack Clark is here as well, director at OpenAI. We have Ms. Tara Lyons, the executive director at Partnerships on AI. And last but not least, Dr. Ben Buchanan.
[0:19:26] A postdoctoral fellow at Harvard Kennedy School's Belfer Center for Science and International Affairs. Say that three times fast. I appreciate all y'all's written statements. It really was helpful in understanding this issue. And pursuant to committee rules, all witnesses will be sworn in before you testify. So please stand and raise your right hand. Do you solemnly swear or affirm that you're about to tell the truth, the whole truth, and nothing but the truth? So help you God. Thank you. Please be seated.
[0:19:59] Please let the record reflect that all witnesses answered in the affirmative. And now, in order to allow for time for discussion, please limit your testimony to five minutes. Your entire written statement will be made part of the record. As a reminder, the clock in front of you shows your remaining time. The light turns yellow when you have 30 seconds left, and when it's flashing red, that means your time is up. Also, please remember to push the talk button to turn your microphone on and off. And now, it's a pleasure to recommend...
[0:20:28] to recognize Mr. Shapiro for your opening remarks.
[0:20:33] I'm Gary Shapiro, President and CEO of the Consumer Technology Association, and I want to thank you, Chairman Hurd and Ranking Member Kelly, for inviting me to testify on this very important issue, artificial intelligence. Our association represents 2,200 American companies in the consumer technology industry. We also own and produce the coolest, greatest, funnest, most important, and largest business and innovation event in the world, the CES, held each January in Las Vegas.
[0:20:59] Our members develop products and services that create jobs. They grow the economy and they improve lives. And many of the most exciting products coming to market today are AI products. CTA and our member companies want to work with you to figure out how we can ensure that the U.S. retains its position as a global leader in AI while also proactively addressing the pressing challenges that you've already raised today.
[0:21:24] Last month, we released a report on the current and future prospects of AI, and we found that AI will change the future of everything, from healthcare and transportation to entertainment security. But it will also raise questions about jobs, bias, and cybersecurity. We hope our research, along with the efforts of our member-driven artificial intelligence working group, will lay the groundwork for policies that will foster AI development and address the challenges AI might create. First, consider how AI is creating efficiency and improving lives.
[0:21:54] The U.S. will spend $3.5 trillion on health care this year. The federal government shelters over 28% of that cost. By 2047, the CBO estimates federal spending for people age 65 and older who receive Social Security, Medicare, and Medicaid benefits could account for almost half of all federal spending. AI can be part of the solution. Each patient generates millions of data points every day, but most doctors' offices and hospitals are not now.
[0:22:24] AI can quickly sift through and identify aspects of that data that can save lives. For example, Qualcomm's AlertWatch AI system, which provides real-time analysis of patient data during surgery, significantly lowers patients' heart attacks and kidney failures, and it reduces average hospital stays by a full day.
[0:22:46] Cybersecurity is another area where AI can make a big impact, according to our study. AI technologies can interpret vast quantities of data to prepare better for and protect against cybersecurity threats. In fact, our report found that detecting and deterring security intrusions was the top area where companies are today using AI.
[0:23:09] AI should contribute over $15 trillion to the global economy by 2030, according to PwC. Both the present and prior administrations have recognized the importance of prioritizing AI. But AI is also capturing the attention of other countries. Last year, China laid out a plan to create a $150 billion world-leading AI industry by 2030. Earlier this year, China announced a $2 billion AI research park in Beijing.
[0:23:38] Just unveiled a high-profile plan to foster AI development in France and across the European Union. I was there last week, and it was the talk of France. Today, the U.S. is the leader in AI, both in terms of research and commercialization. But as you said, Mr. Chairman, our position is not guaranteed. We need to stay several steps ahead. Leadership from the private sector, supported by a qualified talent pool and light-touch regulation, is a winning formula for innovation in America.
[0:24:05] We need government to think strategically about creating a regulatory environment that encourages innovation in AI to thrive while also addressing the disruptions we've been talking about. Above all, as we note in our AI report, government policies around AI need to be both flexible and adaptive. Industry and government also need to collaborate to address the impact AI is having and will have on our workforce. The truth is most jobs will be improved by AI, but many new jobs will be created and, of course, some will be lost.
[0:24:35] We need to ensure that our workforce is prepared for these jobs of the future, and that means helping people whose jobs are displaced gain the skills that they need to succeed in new ones. CTA's AI working group is helping to address these workforce challenges. We just hired our first vice president of U.S. jobs, and on Monday, we launched CTA's 21st Century Workforce Council to bring together leaders in our industry to address the significant skills gap in our workforce we face today.
[0:25:04] In addition to closing the skills gap, we need to use the skills of every American to succeed. CTA is committed to strengthening the diversity of the tech workforce. Full representation of our workforce will go a long way to making sure that tech products and services consider the needs and viewpoints of diverse users. We as an industry also need to address data security. We also need to welcome the opportunity to continue to work with you on that and other areas. We believe that the trade agenda, the IP agenda,
[0:25:33] Immigration all tie into our success as well in AI. There's no one policy decision or government action that will guarantee our leadership in AI, but we are confident we can work together on policies that will put us in the best possible position to lead the world in AI and deliver the innovative technologies that will change our lives for the better. Thank you, Mr. Shapiro. Mr. Clark, you're now recognized for five minutes. Chairman Hurd, Ranking Member Kelly, and other members of the subcommittee, thank you for having this hearing.
[0:26:05] I'm Jack Clark, the Strategy and Communications Director for OpenAI. We're an organization dedicated to ensuring that powerful artificial intelligence systems benefit all of humanity. We're based in San Francisco, California, and we conduct fundamental technical research at the frontiers of AI, as well as participating in the global policy discussion. I also help maintain the AI Index, an AI measurement and forecasting initiative, which is linked to Stanford University.
[0:26:35] I'm here to talk about how government can support AI in America, and I'll focus on some key areas. My key areas are ethics, workforce issues, and measurement. I believe these are all areas where investment and action by government will help to increase this country's chances of benefiting from this transformative technology. First, ethics.
[0:26:57] We must develop a broad set of ethical norms governing the use of this technology, as I believe existing regulatory tools are and will be insufficient. The technology is simply developing far too quickly. As I and my colleagues recently wrote in the report, Malicious Use of Artificial Intelligence, Forecasting, Prevention and Mitigation, this unprecedentedly rapid proliferation of powerful technological capabilities brings with it unique threats or worsens existing ones.
[0:27:25] And because of the nature of the technology, traditional arms control regimes or other policy tools are insufficient. So we need to think creatively here. So how can we control this technology without stifling innovation? I think we need to work on norms. And what I mean by norms are developing a global sense of what is right and wrong to do with this technology. So it's not just about working here in America, it's about
[0:27:51] Taking a leadership position on norm creation so that we can also influence how AI is developed worldwide. And that's something that I think the United States is almost uniquely placed to do well. This could include new norms around publication as well as norms around safety research or having researchers evaluate technologies for their downsides as well as upsides and having that be a part of the public discussion.
[0:28:15] I'm confident this will work. We've already seen similar work being undertaken by the AI community to deal with our own issues of diversity and bias. Here, norms have become a product of everyone, and by having an inclusive conversation that's involved a wide set of stakeholders, we've been able to come to solutions that don't require specific regulations but can create norms that condition the way that the innovation occurs. So a question I have for you is, what do you want to know about?
[0:28:43] What are you concerned about, and what conversations can we have to make sure that we are responsive to those concerns as a community? Second, workforce. The U.S. is currently the leader in AI technology, but as my colleague Gary said, that's not exactly guaranteed. There's a lot of work that we need to do to ensure that that leadership remains in place, and that ranges from investment in basic research to also supporting the community of global individuals that develop AI.
[0:29:11] I mean, part of the reason we're sitting here today is because of innovations that occurred maybe 10 to 15 years ago as a consequence of people I could count on these two hands. So even losing a single individual is a deep and real problem, and we should do our best to avoid it. Third, measurement. Now, measurement may not sound hugely flashy or exciting, but I think it actually has a lot of value and is an area where
[0:29:37] Government can have an almost unique enabling role in helping innovation. The reason why we're here is we want to understand AI and its impact on society. And while hearings like this are very, very useful, we need something larger. We need a kind of measurement moonshot so that we can understand where the technology is developing, where it's going in the future, where new threats and opportunities are going to come from, so that we can have not only informed policymakers, but also a more informed citizenry.
[0:30:07] And I think that having citizens feel that the government knows what's going on with AI and is taking a leadership role in measuring AI's progress and articulating that back to them can make it feel like a collective across America effort to develop this technology responsibly and benefit from it. Some specific examples already abound for ways this works. DARPA wanted to measure how good self-driving cars were and held a number of competitions which enabled the self-driving car industry.
[0:30:35] Two years ago, it held similar competitions for cyber defense and offense, which have given us a better sense of what this technology means there. And even more recently, DIUX released the XView satellite data set and competition, which is driving innovation in AI research in that critical area to national security and doing it in a way that's inclusive of as many smart people as possible. So thank you very much. I look forward to your questions.
[0:31:02] Thank you, Mr. Clark. Well, you're in the right place. Measurement may not be flashy, but we talk about IT procurement as well, which isn't sexy either. So you're with good company. Ms. Lyons, you're now recognized for five minutes. Good afternoon.
[0:31:19] Chairman Hurd, Ranking Member Kelly, thank you for the opportunity to discuss a very important set of issues. I am the Executive Director of the Partnership on Artificial Intelligence to Benefit People and Society, a 501c3 nonprofit organization established to study and formulate best practices on AI technologies, to advance the public's understanding on AI, and to serve as an open platform for discussion and engagement about AI and its influences on people and society.
[0:31:44] The partnership is an unprecedented multi-stakeholder organization founded by some of the largest technology companies in conjunction with a diverse set of cross-sector organizations spanning civil society and the not-for-profit community and academia. Since its establishment, the partnership has grown to more than 50 partner organizations spanning three continents.
[0:32:04] We believe that the formulation of the partnership could not have come at a more crucial time. As governments everywhere grapple with the implications of technology on citizens' rights and governance, and as the research community increasingly emphasizes the need for multidisciplinary work focused on not just the question of how we build technologies, but in some cases whether to, and also in what ways, the partnership seeks to be a platform for collective reflection and, importantly, collective action.
[0:32:32] My remarks this afternoon will focus first on some of the potential opportunities and challenges presented by artificial intelligence, and second on how the partnership hopes to engage with policymakers, with industry, the research community, and other stakeholders. Artificial intelligence technologies present a significant opportunity for the United States and for the world to address some of humanity's most pressing and large-scale challenges, to generate economic growth and prosperity, and to raise the quality of human life everywhere.
[0:33:01] While the promise of AI applied to some domains is still distant, AI is already being used to solve important challenges. In healthcare, already mentioned, AI systems are increasingly able to recognize patterns in the medical field, helping human experts interpret and scan and detect cancers. These methods will only become more effective as large data sets become more widely available.
[0:33:23] And beyond healthcare, AI has important applications in environmental conservation, education, economic inclusion, accessibility, and mobility, among other areas. As AI continues to develop, researchers and practitioners must ensure that AI-enabled systems are safe, that they can work effectively with people and benefit all parts of society, and that their operation will remain consistent and aligned with human values and aspirations. World-changing technologies
[0:33:50] Need to be applied and ushered in with corresponding social responsibility, including attention paid to the impacts that it has on people's lives. For example, as technologies are applied in areas like criminal justice, it is critical for the partnership to raise and address concerns related to the inevitable bias in data sets used to train algorithms. It's also critical for us to engage with those using such algorithms in the justice system so that they understand the limits of these technologies and how they work.
[0:34:17] Good intentions, too, are not enough to ensure positive outcomes. We need to ensure that ethics are put into practice when AI technologies are applied in the real world and that they reflect the priorities and needs of the communities that they serve. This won't happen by accident. It requires a commitment from developers and other stakeholders who create and influence technology to engage with broader society, working together to predict and direct AI's benefits and to mitigate potential harms.
[0:34:44] Identifying and taking action on high priority questions for AI research, development, and governance will require the diverse perspectives and resources of a range of different stakeholders, both inside and outside of the partnership on AI community. There are several ways in which we are delivering this. A key aspect of this work at the partnership has so far taken the form of a series of working groups, which we have established to approach three of our six thematic pillars, with the other three to follow soon.
[0:35:11] These first working groups are on safety critical artificial intelligence, fairness, transparency, and accountability in AI, and also AI labor in the economy. The partnership will also tackle questions that we think need to be addressed urgently in the field and are ripe for collective action by a group of interests and expertise as widespread and diverse as ours. Our work will take different forms and could include research, standards development, policy recommendations, best practice guidelines, or codes of conduct.
[0:35:39] Most of all, we hope to provide policymakers and the general public with the information they need to be agile, adaptive, and aware of technology development so that they can hold technologists accountable for upholding ethical standards in research and development and better understand how these technologies affect them. We are encouraged by these hearings and the interest in policymakers in the U.S. and worldwide, both toward understanding the current state of AI and the future impacts it may have. I thank you for your time, and I look forward to questions.
[0:36:09] Appreciate you, Ms. Lyons. Dr. Buchanan, you're now recognized for five minutes for your opening remarks. Thank you, Chairman Hurd and Ranking Member Kelly, for holding this important hearing and for inviting me to testify. As you mentioned, I'm a fellow at Harvard University's Belfer Center for Science and International Affairs, and my research focus is on how nations deploy technology, in particular cybersecurity, including offensive cyber capabilities and artificial intelligence.
[0:36:39] Recently, with my friend and colleague Taylor Miller of the Icahn School of Medicine at Mount Sinai, we published a report entitled Machine Learning for Policymakers. And to help open today's hearing, I'd like to make three points, one on privacy, one on cybersecurity, and one on economic impact. And I'll try to tailor this to not be repetitive. I think we're in agreement on a lot of these areas.
[0:37:04] To simplify a little bit, we can think about modern artificial intelligence as relying on a triad of parts. Some data, some computing power, and some machine learning algorithms. And while we've seen remarkable advances on the computing and learning algorithm side, I think for policymakers such as yourselves, it's data that's most important to understand. And data is the fuel of machine learning systems.
[0:37:31] Without this data, these systems sometimes produce results that are embarrassingly wrong and incorrect. Gathering relevant and representative data for training, development, and testing purposes is a key part of building modern artificial intelligence technology. On balance, the more data that is fed into a machine learning system, the more effective it will be.
[0:37:52] It is no exaggeration to say that there are probably many economic, scientific, and technological breakthroughs that have not yet occurred because we have not assembled the right data sources and right data sets. However, there is a catch and a substantial one. Much of that data that might, and I emphasize might, be useful for future machine learning systems is intensely personal, revealing, and appropriately private.
[0:38:17] Too frequently, the allure of gathering more data to feed a machine learning system distracts from the harms that collecting that data brings. There is a risk of breaches by hackers, of misuse by those who collect or store the data, and of secondary use, in which data is collected for one purpose and later reappropriated for another. Frequently, attempts at anonymization do not work nearly as well as promised. It suffices to say that, in my view, any company or government agency
[0:38:46] Collecting large amounts of data is assuming an enormous responsibility. Too often, these collectors fall far short of meeting that responsibility. And yet in an era of increased artificial intelligence, the incentive to collect ever more data is only going to grow.
[0:39:01] And technology cannot replace policy, but some important technological innovations can offer mitigation to this problem. Technologies such as differential privacy, that approach can ensure that large data sets retain a great deal of their value, but protecting the privacy of any one individual member. On-device processing can reduce the aggregation of data in the first place. This is an area in which much remains to be done. Second, AI is poised to make a significant impact in cybersecurity.
[0:39:29] Potentially redefining key parts of the entire industry. Automation on offense and on defense is an area of enormous significance. We already heard about the DARPA Grand Cyber Challenge, which I agree was a significant seminal event. And we've certainly seen what I would describe as the beginnings of significant automations of cyber attacks in the wild. In the long run, it's uncertain whether increased automation
[0:39:54] will give a decisive cybersecurity advantage to hackers or to network defenders, but there's no doubt of its immediate and growing relevance. AI systems also pose new kinds of cybersecurity challenges. Most significant among these is the field of adversarial learning, in which the learning systems themselves can be manipulated, oftentimes by what we call poisoned data sets, to produce results that are inaccurate and sometimes very dangerous. And that's another area which is very nascent
[0:40:24] Not nearly as developed as mainstream cybersecurity literature. Again, much more remains to be done. A more general concern is AI safety. And this conjures up notions of Terminator and AI systems that will take over the world. In practice, it is often far more nuanced and far more subtle than that, though the risk is still quite severe. I think it is fair to say that we have barely scratched the surface.
[0:40:48] And this is an area, as my fellow witnesses suggest, in which the United States should be a leader. Third, AI will have significant economic effects. My colleagues here have discussed many of them already. The ranking member mentioned two notable studies. I would point you to two other studies, both, I believe, by MIT economists, which show that while theory often predicts a job's loss will be quickly replaced, in practice, at least in that one instance, that did not immediately occur.
[0:41:18] With that, I will leave it there, and I look forward to your questions. Thank you. Thank you, Dr. Buchanan. I now recognize a ranking member for five minutes or so for your first round of questions. Thank you, Mr. Chair, and thank you to the witnesses again. The recent news that Cambridge Analytica had improperly obtained the personal data of up to 87 million Facebook users highlights the challenges to privacy when companies collect large amounts of personal information.
[0:41:46] Dr. Buchanan, in your written testimony, you state, and I quote, that much of the data that might, and I emphasize might, be useful for future machine learning systems is intensely personal, revealing, and appropriately private. Is that right? You just said that. That's correct, Congresswoman. And can you explain for us what types of risks and threats consumers are exposed to when their personal information is collected and used in AI systems?
[0:42:15] Sure. As you'd expect, Congresswoman, it would depend on the data. Certainly some financial data, if it were to be part of a breach, would lead to potential identity theft. There's also data revealed in terms of preferences and interests that many members of society might want to keep appropriately private. We've heard a lot about AI in medical systems. Many people want to keep their medical data private. So I think it depends on the data.
[0:42:42] But there's no doubt that, in my view, if a company or a government organization cannot protect the data, it should not collect the data. In light of these risks and your assessment of the majority of companies that do collect and use personal data for the AI systems, are they taking adequate steps to protect the privacy of citizens? Speaking as a generalization, I think we have a long way to go. Certainly the number of breaches that we've seen in recent years.
[0:43:11] Including a very large data set such as Equifax suggests to me that there is a lot more work that needs to be done in general for cybersecurity and data protection. And also in your written testimony, you also outlined different types of safeguards that
[0:43:27] Could improve the level of protection of consumers' privacy when their data is collected and stored in AI systems. One of those safeguards is the use of a technical approach you refer to as differential privacy. Can you explain that in layman's terms? Sure. Simplifying a fair amount here, differential privacy is the notion that before we put data into a big database from any individual person, we add a little bit of statistical noise to that data.
[0:43:56] And that obscures what data comes from which person. And in fact, it obscures the records of any individual person. But it preserves the validity of the data in the aggregate. So you can imagine if we asked every member of Congress, have you committed a crime? Most Congress people and most people don't want to answer that question. But if we said to them, flip a coin before you answer, if its heads answer truthfully, if its tails don't answer truthfully, flip another coin.
[0:44:25] And use a second coin flip to determine your made-up answer. We're adding a little bit of noise when we collect the answers at the end. And using a little bit of math at the back end, we can subtract that noise and get a very good aggregate picture without knowing which members of Congress committed crimes. So the broader principle certainly holds, again, with the fair more math involved, that we can get big picture views without sacrificing the privacy or criminal records of individual members of the data set.
[0:44:53] I have not committed a crime, by the way. But do you feel like if more businesses adopted this differential privacy, this type of security measure would be more effective in mitigating the risk to personal privacy? With something like differential privacy, the devil's in the details. It has to be implemented well. But as a general principle, yes, I think it's a very positive technical development and one that is fairly recent. So we have a lot of work to do.
[0:45:20] It shows enormous promise, in my view. Thank you. And in addition to this, you also identify in your written testimony another type of security control known as on-device processing. Can you again, in layman's terms, explain on-device processing and how it operates to protect sensitive and personal data? Sure. This one is much more straightforward. It is essentially the notion that
[0:45:43] If we're going to have a user interact with an AI system, it is better to bring the AI system to them rather than bring their data to some central repository. So if an AI system is going to be on your telephone, your cell phone rather, you can interact with the system and do the processing on your own device rather than at a central server where the data is aggregated. Again, as a general principle, I think that increases privacy.
[0:46:09] What are the reasons why more companies in general are not deploying these types of security controls? Certainly, as a matter of practice, they require enormous technical skill to implement. Frankly, I think some companies want to have the data, want to aggregate the data and see the data, and that's part of their business model. And that's an incentive for those companies not to pursue these approaches. What recommendations would you have for ways in which...
[0:46:38] Congress can encourage AI companies to adopt more stringent safeguards for protecting personal data on consumers. I think Mr. Clark has made excellent points about the importance of measurement, and I think this is an area that I would like to know more about and measure better of how are American companies storing, securing, and processing the data on Americans. So that would be Chairman Hurd mentioned measurements, a topic of interest to this committee, and I think that would be one place to start.
[0:47:07] And just lastly, a part of the struggles that companies have, is it because they don't have enough of the expertise because it's not in the workforce? Yes, Congresswoman, I think that's right. There's an enormous demand that has not yet been met for folks with the skills required to build and secure these systems. That's true in AI, and that's true in cybersecurity generally. And would the rest of the witnesses agree with that also? Yes, the last comment.
[0:47:38] Yeah, we need 10, 100 times more people with these skills. I would agree with Dr. Buchanan. Thank you, and thank you, Mr. Chair. Thank you, Ranking Member. Mr. Shapiro, I had the fortune of attending CES, the Consumer Electronics Show, this recent January. Thanks for putting on such a good show. I learned a lot about artificial intelligence and how...
[0:48:08] How important data is in training the actual, training the algorithm. And one of the questions that we have come up, or we have heard, one of the issues we have heard is the importance of data. And we've learned about bias and preventing that. We learned about being auditable.
[0:48:35] We know we have to invest more money in AI. We also know we need to train people better. Who should be taking the lead? Who is the person, who should be driving kind of this conversation? Or maybe let me narrow the question. Who in government should be driving kind of the investment dollars in this? And I know you have...
[0:49:05] Peer research at universities. You have the national labs. Who should be coming up with that with our investment plan in AI? Well, I think, first of all, we have to agree on the goals. I'd like the idea of measurement as well. And I think the goals are, number one, we would like the U.S. to be the leader. Two, we want to solve some fundamental human problems involving health care, safety, cybersecurity, and others. And we could define goals with those.
[0:49:34] Third, we want to respect people's privacy. And I think there has to be a national discussion on some of these issues. Because let's take the issue of privacy, for example. And we've heard a lot about that today. The reality is culturally we're different on privacy than other parts of the world. In China, the concept of privacy is, especially in this area, is that the citizens really don't have any. They're getting social scores. Their government's monitoring what they do socially. And certainly there doesn't seem to be much legal.
[0:50:03] They have the right to be forgotten. They have the right to erase history, something that seems anathema to us. How could you change the facts and take them off the Internet? And they've really clamped down, and they're going forward in a very...
[0:50:22] Arguably bold and unfortunate way on this GDPR, which is really, you could argue it's for privacy, or you could argue it's to shut Europe out in a competitive fashion. When I look at the U.S. and our focus on innovation and our success, and I compare it to Europe, I see they have maybe a handful of unicorns, billion-dollar valuation companies, and the U.S. has most of them in the world, over 150. And why is that? There's many answers.
[0:50:49] It's our First Amendment. It's our diversity. It's our innovation. It's the culture we have of questioning. There's many things go to it. But part of it, we're a little more willing to take some risks in areas like exchange of information. Europe's going forward with GDPR. And frankly, it's going to hurt American companies. I was just in France last week. It's going to hurt European companies. They're terrified of it. They're talking about trying to delay it. But it's also going to kill people.
[0:51:16] Because if you can't transfer, for example, medical information from one hospital to another in the same region, that has life consequences. So when we talk about the issue of privacy and who should lead on it, I think we should do it in a common sense way. And we shouldn't let HIPAA, for example, be our model. The model should be what is going to be...
[0:51:34] What kind of information is and what's the situation? We've done a lot of research, and we have found, for example, that Americans are willing to give up personal information for a greater good, as they've done with health information on our Apple Watches. They're willing to do it for safety. They're willing to do it for the security of their children. They're willing to do it for their own safety, involving, for example,
[0:51:52] Where your car is, if it hits a car in front of them. So in the privacy area, I think we have a pretty good cultural sense. I think the Federal Trade Commission has the broad mandate to do a pretty good job in that area. And then I don't want to take all the time, but those other two areas I talked about in terms of the measurements and artificial intelligence and what you should do, it goes into how you get the skilled people, what you do, how you change your educational system, how you retrain for jobs. There's a lot of things that government can do and Congress can do, and I applaud you and this committee for taking the first big step in having
[0:52:22] What I would expect Congress to do in the future is rather than come up with immediate solutions, is instead to focus on what the goals are and how we could do there. And I would look at two examples that I was both personally involved with, which was government setting big goals, but working with industry who came up with private things. Actually, I'll give three quickly. One is, and Congressman Issa is very aware of this because he was part of it, the transition to high-definition television. That was, we set the goal. We wanted to have the best system in the world.
[0:52:49] Private industry, no spending of government money, we did it. Second is commercialization of the internet, doing business over it. It was done in Virginia with bipartisan way, and the goals were there, and it worked. And the third is you talked about privacy for wearable devices, healthcare devices, which came up earlier. At CTA, we got everyone in the room that made those devices, and we agreed on a regimen of saying this is what we should voluntarily do. This is what we should follow. It should be transparent, clear language.
[0:53:15] Opt out, and you can't sell information or use it without any permission from your customers. And the Obama administration seemed pretty happy with that, and even they didn't act, because that was industry self-regulation. Got you. Thank you, Mr. Shapiro. I'm going to come back for another round of questions, but now I'd like to recognize my friend from the Commonwealth of Virginia for his first round of questions. Thank you, Mr. Chairman. And Gary, you were doing speed.
[0:53:42] Dating there. And welcome. Good to see you again. I want to give you a little bit more opportunity maybe to those three things you were just talking about if you want to elaborate a little bit more. Because this idea, let you catch your breath for a minute. This idea of the zone of privacy and some of its cultural bound, I think it's absolutely true. But I can remember going to Silicon Valley about nine years ago meeting with Facebook people. And their view about privacy was we...
[0:54:12] Americans need to get used to shrinking boundaries for privacy and that the younger generations were already there. Older generations needed to just learn to suck it up and accept it. And I think watching what happened to Mr. Zuckerberg here in the last couple of weeks, one needs to not be so facile.
[0:54:38] You're not being, but I mean, I think you're raising, though, those questions. Some of it's cultural bound, some of it, the rules of engagement aren't quite there yet. We debate, do we get involved? If so, what do we do? And so I think your thoughts are very helpful, given your experience and your position, in providing some guidance.
[0:55:01] I want to give you an opportunity to elaborate just a little bit. Well, thank you very much, Congressman. I appreciate it. I guess my view is that as a nation, we're not China, where we totally don't devalue privacy, and we're not Europe, where we use privacy as a competitive tool against other countries, frankly, but it also tamps down innovation in our own country. Our competitive strength is innovation. That's what we're really good at. It's the nature of who we are. So the question is, how do we foster innovation in the future in AI and other areas and also maintain our...
[0:55:31] Correct, our citizens view that they are entitled to certain things. Now, to a certain extent, it's educating. Everyone has an obligation. The obligation of business is to tell our customers what it is we're doing with our data in a clear and transparent way. And frankly, we haven't done a great job at it. I mean, if I had my way, I wouldn't want to have to click on those I agree just to get to the website I want to. I'd like to click on platinum, gold, or silver standard. If there's some standardization, it would probably help. And government could play a role in that.
[0:55:58] But we also want to make it sure that we could innovate. And consumers should understand that they're giving away something in return for free services. You give a tailor your information on your body size to get clothes that fit. You give your doctor information about your health, and you're always giving away something. And the truth is, if you're going to get a free service like Facebook or Google and you want to keep it free, they are using that data to get people to know you. But it's like I shop at Kroger's in Michigan, actually, because that's where...
[0:56:25] I commute to. And Kroger's knows a lot about me. They know everything I buy. And they give me coupons all the time. And I value those coupons. But they know what I buy. And I am willing to do that. It's the same thing with other frequent user programs. We're doing that all the time. We're giving up information about ourselves. We get discounts. We get deals. And we get better service. If we do it with our eyes open and we're educated about it, that's fine. Now, the role of citizens is to understand. By the way, we know you shop at Kroger's last Thursday.
[0:56:52] The fondness you've got for frozen rolls has frankly surprised us. So in terms of the role of government, I think the role of government is to start out by having hearings like this one, define the goals and the measurements culturally for the future, and the role, frankly, of the administration, in my view, is to set the big goals and to make sure that we buy in them.
[0:57:10] into them in a bipartisan way. And I love the idea of some big goals, as Mr. Clark suggested, because we need big goals in this area. For example, having self-driving cars by 2025, or not saying self, dropping the death rate from automobiles down by half by a certain date would be a very admirable goal that everyone in this country could rail around. Thank you so much, Gary. Mr. Clark, in the time I've got left, you said in the report on OpenAI,
[0:57:39] That artificial intelligence continues to grow, cyber attacks that will utilize AI and will be, and you said, quote, more effective, finely targeted, difficult to attribute, and likely to exploit vulnerabilities in AI systems. Wanted to give you an opportunity to expand a little bit on that. So how worried should we be? So you can think of AI as something that we're going to add to pretty much every aspect of technology, and it's going to make it more powerful and more capable.
[0:58:08] So this means that our defenses are also going to get substantially better. And as Dr. Buchanan said earlier, you weren't in the room, it's not clear yet whether this favors the defender or the attacker. And this is why I think that hosting competitions, having government measure these capabilities as they develop, will give us a kind of early warning system. You know, if there's something really bad that's about to happen as a consequence of an AI capability,
[0:58:34] I'd like to know about it, and I'd like an organization or an agency to be telling us about that. So you can think about that and take that and view it as an opportunity, because it's an opportunity for us to learn in an unprecedented way about the future before it happens and make the appropriate regulations before harm occurs. If the chair will allow, I don't know if Dr. Buchanan or Ms. Lyons want to add to that, and my time is up. I have nothing more to add. Thank you.
[0:59:07] I think we probably can return to the subject later, but I would suggest we've seen some indications already of increased autonomy in cyber attack capabilities. There's no doubt in my mind we will see more of that in the future. The distinguished gentleman from California is now recognized for his round of questions. You know, this is what happens when you announce your retirement. You become distinguished. You know, I...
[0:59:32] I know in these hearings that there's sort of an exhaustive repeat of a lot of things, but let me skip to something I think hasn't happened, and I'll share it with each of you, but I'll start with Mr. Clark. The weaponization of artificial intelligence, there's been some discussion about how far it's gone, but it's inevitable. The tools of artificial intelligence disproportionately favor U.S. companies. Now, when that happened in satellites, nuclear capability,
[1:00:02] And a myriad data processing. We put stringent export control procedures on those things which may have a dual use. We've done no such thing in artificial intelligence. Would you say today that that is an area in which the Commerce Department's Export Assistant Secretary doesn't have specific authority but needs it? Thank you. I think this is a question of existential importance to basically the world. The issue of AI is that...
[1:00:32] It runs on consumer hardware. It's embodied in software. It's based on math that you can learn in high school. You can't really regulate a lot of aspects of the fundamental AI development because it comes from technology which 17-year-olds are taught in every country of the world. And every country is developing this. So while the U.S. economy favors the development of AI here and we have certain advantages, other countries are working on this.
[1:00:59] So I think that export controls, arms controls do not really apply here. We're in a new kind of regime because you can't control a specific thing with this AI technology. Instead, you need to develop norms around what is acceptable. You need to develop shared norms around what we think of in AI as safety, which is about being able to offer guarantees about how the systems work and how they behave. And we need to track those capabilities.
[1:01:28] So I think that your question's a really important one, and I think it touches an area where much more work needs to be done because we don't have the right tool today to let us approach the problem. And let me follow up quickly. When we look at artificial intelligence, we look at those producing advanced algorithms, and I went to a different high school, apparently, than you did. Mine was in Caltech. So let's assume for a moment that it's slightly above high school level.
[1:01:55] The creators of those, and let's assume for a moment, hypothetically, they're all in the first world, and the first world defined as those who want to play nice in the sandbox, the US, Europe, and a number of other countries. Do you believe, if that's the case, the government has a role, though, in ensuring that when you make the tool that is that powerful, the tools that...
[1:02:17] If you will, allow it to be safely controlled are also part of the algorithm. In other words, the person who can make a powerful tool for artificial intelligence also can in fact design the safety mechanism to ensure that it couldn't be used clandestinely. Do you think that's a social responsibility of, let's say, the Facebooks and the Googles? I think we have a social responsibility.
[1:02:45] to ensure that our tools are safe and that we're developing technologies relating to safety and reliability in lockstep with capabilities. That's something that the organization I work for, OpenAI, does. We have a dedicated safety research team, as does Google's DeepMind. They do as well. So you need to develop that. But I think to your question, it's how do you, if you have those tools, make sure everyone uses it. I think there you're going to deal with...
[1:03:13] Well, you know, as we discovered today that we've sent our CIA director to meet with Kim Jong-un because he can't be trusted with the tools he's created that got to him. I might have a different view on the export controls, but Mr. Shapiro, since you've given me every possible look on your very creative face as these answers came through, let me allow you to answer that, but I want to shift to one other question. You mentioned a little bit HIPAA.
[1:03:42] Now, the history of HIPAA is pre-computer data. It is, in fact, a time in which basically pieces of paper were locked up at night and not left out on desks so that one patient didn't see another patient's records and that you didn't arbitrarily just answer anyone over the phone. The reality, though, today is that the tools that your industry, the industry you so well represent,
[1:04:07] You have tens of thousands of tools that are available that can gather information, and often they're limited by these requirements from really interacting with the healthcare community in an efficient way. Do we need to set up those tools to allow healthcare to prosper in an interactive, cloud-based computer generation? And I'll just mention, for example,
[1:04:33] The problem of interoperability between Department of Defense, the Veterans Administration, and private doctors. That has confounded our veterans, often leading to death, to overdose, for a lack of that capability. Do you have the tools, and what do we need to give you to use those tools? Well, it's probably fair to say that the promise of Obamacare, which was very positive of...
[1:04:59] Allowing easy transfer of electronic medical records has not been realized. I think even the American Medical Association, which urged that it be passed, has now acknowledged that, and it's been a great frustration to doctors, as I think you know. In terms of the tools that we have today to allow easy transfer, the administration has endorsed this Blue Button Initiative, which allows medical records, especially in emergency cases, to be transferred easily. I think we have a long way to go.
[1:05:25] as a country to make it easy to transfer your own health information. The old ways they did in the communist countries, you'd walk around with your own records, your paper records, were actually a simpler transaction than what we have today where everyone goes in and has to start from zero. Well, you know, the chairman is too young to know, but I walked around in the army with that brown folder with all my medical records, and it was very efficient.
[1:05:48] What's a folder? But the opportunity we see in this, and we're concerned as an organization about the growing deficit and the impact that will have existentially on our country, frankly, and we see the opportunity there in providing health care and using technology in a multitude of ways to lower costs, to be more efficient, to cut down on doctor visits, and to just allow easy transfer of information. In terms of what the government can do, we're actively advocating for a number of things. We're working with the FDA. We're moving things along, and we've found...
[1:06:16] With this administration and the prior administration, a great willingness to adopt a technology system no matter how fast. Thank you. Mr. Chairman, are we going to have another round? Okay, I'll wait. Thank you. Ranking member, round two. Thank you. Given Facebook CEO Mark Zuckerberg's comments last week to Congress, how would you evaluate AI's ability to thwart crime online from stopping rogue pharmacies, sex trafficking, IP theft, identity theft?
[1:06:47] To cyber attacks and whoever wants to answer that question, I'm listening. I think speaking generally here, there's enormous promise from AI in a lot of those areas. But as I said in my opening remarks, we should recognize that technology will not replace policy. And I think it's almost become a cliche in certain circles to suggest that while we have this very thorny...
[1:07:16] Complex, interdisciplinary problem. So let's just throw machine learning at it and the problem will go away. And I think that's a little bit too reductive as a matter of policymaking. Anybody else? I would echo Dr. Buchanan's remarks insofar as I think part of the solution really needs to be in bringing multiple solutions together.
[1:07:44] Policy is certainly part of the answer. I think technology and further research in certain areas related to security, as you mentioned in this specific case, is the answer. And I think also that is sort of the project of the organization that I represent insofar as the interest of bringing different sectors together to discuss the means by which we do these things in the right ways. Thank you. Dr. Buchanan, what types of jobs do you see?
[1:08:12] That will be threatened in the short term by AI automation. And what about in the long term as well? Certainly in the near term, I think the jobs that are most at risk are jobs that involve repetitive tasks. And certainly this has always been the case with automation. But I think as you can imagine, as artificial intelligence systems become more capable, what they can do, what they consider repetitive, certainly would increase. And I think jobs that involve...
[1:08:41] Again, repetition of particular tasks that are somewhat by rote, even if they're jobs that still involve intellectual firepower, are on balance more likely to be under threat first. And in the long term, what do you see? As we move towards an era of things like self-driving cars, one could imagine that services like Uber and Lyft might not see a need for drivers, and taxis might not see a need for...
[1:09:09] Taxi companies, rather, might not see a need for drivers. There's some suggestion that if we had such a world, we would need fewer cars in general. Certainly members of Congress are acutely aware of how important the auto industry is to the United States. So when you look at a longer-term horizon, I think there's more uncertainty, but there's also a lot more potential for disruption, particularly with knock-on effects. If the auto industry is smaller, for example, what would the knock-on effects be on suppliers to companies?
[1:09:37] Even beyond just the narrow car companies themselves. And to everyone's answer, what type of investments do you feel that we should be making now for people that are going to probably lose their job? How do you see them transitioning to these type of jobs? Of course. So I would, the prior question about the jobs, the great news is the really unpleasant, unsafe jobs
[1:10:07] Most of them will go away. So for example, I was visiting a robotics company and they have something specialized which is very good at picking up and identifying and moving things around using AI and robotics. The jobs, one of the potential buyers was a major pizza company that delivers to homes. The way they do it is they make dough and the dough today is made by people in a very cold, sterile environment. They wear all this equipment to be warm and also to be sterile. They can only work.
[1:10:36] It's very ineffective. No one wants to do the job at all. And this solves that problem. There's also thousands of other conditions where jobs are really difficult. It could be picking agriculturally, where now there's increasingly there's devices which do that. And they do have to be fairly smart to identify the good versus the bad and what to pick and what not to pick.
[1:10:57] In terms of what investment has to make, in terms of retraining, I think we have to look at the best practices in retraining and figure out what you could do. I mean, we do have millions of unemployed people today, but we have more millions of jobs that are open and not filled. Some of it is geographic. We should be honest about that. Maybe we need to offer incentives for people to...
[1:11:18] Move elsewhere. But some of it is skills. And the question is, what has worked before? What skills that you can train someone for? Whether it's a customer service center or whether it's something basic programming or helping out in other ways. I think we have to look at individual situations.
[1:11:34] Out there that's already worked and try some new things. Because a lot of what has worked in the past will not work in the future. And the longer term investment is obviously with our kids. We just have to start training them differently. And we also have to bring back respectability to work which is technical work. As Germany has done. And focus on apprentice programs and things like that. And not just assume that a four year degree is for every American. Because it's not a good investment of society. And there's a lot of unemployed people who went to college who don't have marketable skills.
[1:12:06] Mr. Clark. So I think that this touches on a pretty important question, which is where the jobs get created, because new jobs will be created, will be uneven. And where the jobs get taken away will also be uneven. So I want to refer to a couple of things I think I've already mentioned. One is measurement. It's very difficult for me to tell you today what happens if I drop a new industrial robot into a manufacturing region. I don't have a good economic model to tell you how many jobs get lost, but I have an intuition some do.
[1:12:36] Thank you very much. We'll be right back.
[1:12:57] Mr. Clark, you talked about a competition akin to Robotics One, akin to self-driving car. What is the competition for AI? What is the question that we send out and say, hey, do this thing show up on Wednesday, August 19th, and bring your best neural network and machine learning algorithm?
[1:13:22] So I have a suggestion. The suggestion is a multitude of competitions, this being the Oversight Committee. I'd like a competition on removing waste in government bureaucracy, which is something that I'm sure that everyone here has a feeling about. But I think that that actually applies to every committee and every agency. The Veterans Agency can do work on health care. They can do a health care moonshot within that system that they have to provide health care to a large number of our veterans.
[1:13:52] Thank you very much.
[1:14:17] Lots of experience with things that typical people don't. You have lots of awarenesses of things that are threats or opportunities that may not be obvious. And if you can galvanize kind of investment and galvanize competition there, it can be kind of fun and we can do good work. So along those lines, how would you declare a winner? In which aspect? Let's say we were able to get... We can take one... Let's take...
[1:14:47] HHS. Let's take Medicare. Medicare overpayment is a perfect example. And let's say we were able to get that data, protect it in a way that the contestants were able to have access to it. And you got 50 teams to come in and solve this problem.
[1:15:04] How would you grade them? So in AI, we have a term called the objective function. What it really means is just the goal. And whatever you optimize the goal of the thing for is what you'll get out. So doing that goal selection is important because you don't want to pick the wrong goal because then you'll kind of mindlessly work towards that. But a suggestion I'd have for you is the time it takes a person to flow through that system. And you can evaluate how the application of new technologies can reduce the time it takes for that person to be processed by the system.
[1:15:33] And then you can implement systems which dramatically reduce that amount of time. I think that's the sort of thing which people naturally approve of. And just thinking through it on the spot here, I can't think of anything too bad that would happen if you did that. But I would encourage you to measure and analyze it before you set the goal. Miss Lyons, what's the next milestone when it comes to artificial intelligence?
[1:15:56] From a technical perspective or otherwise? From a technical perspective. Well, I think what a lot of the AI research community is looking towards is AI platforms that can be applied to more generalized tasks than just the very narrow AI that we see applied in most of the circumstances that we've described today. So I would say that is the next sort of moonshot milestone for the technical community. Is that a decade? Is that nine months? Is it 20 years?
[1:16:25] You know, I have my own personal perspectives on this. The partnership hasn't really formulated one yet, but I think we have a lot of organizations involved in ours which have disagreeing viewpoints on this. And I'm sure, actually, if this committee was quizzed, we might all have different answers as well. But I think we're years and years away from that, and it's useful to be thinking about it right now, but I do think we're probably decades. And what are the elements that are preventing us from getting there?
[1:16:53] I actually don't think I'm the best person equipped on this panel to give you an answer to that. I'm pretty far away from the technical research from where I'm sitting right now, but they are technical impediments that are stopping us from achieving that at this moment. Good copy. Dr. Buchanan, how do we detect bias? I think one of the things that we have heard through these hearings is bias, and we know how you...
[1:17:22] Can the algorithm itself be biased? Is the only way to introduce bias is by the data set? And then how are we detecting whether or not the decisions that are being made by the algorithm show bias? I'm not convinced that we're looking as much as we should. So when you say how are we detecting, I think in many cases,
[1:17:53] We are not detecting bias systems. But speaking generally of how do you... Philosophically, how do we solve that problem? I think it's worth, again, as I said before, technology cannot replace policy. And we should first develop an understanding of what we mean by bias. Is the system biased, whether it's automated or not, if it disproportionately affects a particular racial group or gender or socioeconomic status? And I think...
[1:18:20] Most people would answer yes, and you would want to look at what the outcomes of that system were and how it treated individuals from certain groups. And there's a number of different values you can instantiate in the system that try to mitigate that bias. But bias is a concept that we intuitively all feel, but it's often quite difficult to define. And I think a lot of the work in detecting bias is first work in defining bias. Mr. Clark.
[1:18:49] So I have two suggestions. I think both are pretty simple and doable. One is whenever you deploy AI, you deploy it into a cohort of people. Let's say I'm deploying a public service speech recognition system to do a kind of better version of a 311 service for a city. Well, I know I have demographic data for that city, and I know that people that speak, perhaps not with my accent, but a more traditional American one, are going to be well represented. You have an accent? It's sometimes called Australian.
[1:19:18] But it's actually English. So I think that when you look at your city, you're going to see people who are represented in that city but are not the majority. So you test your system against the least represented people and see how it breaks. That will almost invariably surface areas where it can be improved. And the second aspect is you need more people in the room. This requires like a concerted effort on STEM education and on fixing the diversity in STEM because if you're not in the room, you probably just won't.
[1:19:48] Realize that a certain bias is going to be obvious, and we do need to fix that as well. So we've all... I might just chime in by summarizing, because I don't think I heard this clarified in the way that I might describe it, which is in the different ways in which bias is represented in systems. And I think that is through data inputs, which we've talked a little bit about. It's also in the algorithms themselves, and I think that gets to some of...
[1:20:16] Points that Mr. Clark has made around who is building the systems and how representative those developer communities are. And then I think further than that, it's also in the outcomes represented by those various inputs and the ways in which there might be adverse or outsized impacts on particularly at-risk communities who are not involved in technology development. So I just wanted to add that. That's incredibly helpful. We've all been talking about...
[1:20:43] What should the standards be or what are the equivalent of the three rules from iRobot? And one, that it seems that there's most agreement, and correct me if I'm wrong on this, is making sure the decisions of the algorithm are auditable, that you understand how...
[1:21:13] That decision was made by that algorithm. There's been so many examples of an AI system producing something in the
[1:21:28] People that design the algorithm have no clue how the algorithm produced that. Is that the first rule of artificial intelligence? What are some potential contenders for the rules of ethical AI? And Dr. Buchanan, maybe start with you, go down the line if anybody has opinions. I suggest that the first rule might generate more discussion than you'd expect on this panel.
[1:21:58] In general, there is oftentimes a trade-off because of the technology involved in AI systems between what we call the explainability or interpretability of an algorithm's decision and how effective the algorithm is or how scalable the algorithm is. So while I certainly think it's an excellent aspiration to have an explanation in all cases, and while I probably believe that more than many others, I could imagine cases in which we worry less about
[1:22:27] Anybody else have an opinion? Ms. Lyons?
[1:22:53] I think the question that you've raised is actually fairly central to ongoing dialogues happening in the field right now. And the easy answer is that there is no easy answer, I think. And I think Dr. Buchanan has demonstrated that with his remarks as well. But generally speaking, I do think that it has been a particular focus, especially in the last several years, of a certain subset of the AI machine learning.
[1:23:20] Technical community to consider questions associated with issues regarding fairness, accountability, transparency, explainability, and those are issues associated with auditability as you describe. And a keen interest, I think, in making sure that those conversations are multidisciplinary in nature as well, and including people from fields not necessarily traditionally associated with computer science and AI and machine learning communities, but also inclusive of the ethics community, law and policy.
[1:23:49] I recognize that this is a loaded question and there's not an agreement on this, but what are some of the other contenders that people say, hey, we should be doing this? Even if we recognize there's not agreement when it comes to what are the rules of ethical AI?
[1:24:12] Well, the partnership for its part has a set of tenets, which are essentially eight rules that govern the behavior at a very broad level of the organizations associated with us. And they're posted on our website. We included them in our written remarks or written testimony as well. But generally speaking, at a high level, we have, as a community, decided on certain sort of codes of conduct to which we ascribe as organizations involved in this endeavor.
[1:24:40] I think a central project of this organization moving forward from this point is in figuring out how to actually operationalize those tenants in such a way that they can be practiced on the ground by developers and by other associated organizations in the AI technical community. Mr. Clark? So I want to put a slightly different spin on this, and that's about making sure that the decisions an AI makes are sensible. And what I mean by sensible is...
[1:25:07] Why do we send people to school? Why do we do professional accreditation? Well, it's because we train people in a specific skill, and then they're going to be put into a situation that they may not have encountered before, but we trust that the training and education they had in school means that they'll take the correct action. And this is a very common thing, especially in areas like disaster response, where we train people to be able to improvise.
[1:25:33] Fully auditable, like you ask them, why did you do that in that situation? And they'll say, well, it seemed like the right thing to do. That's not a super auditable response, but it's because we're confident in the training they've had. And so I think some of it is about how do we make sure that the AI systems we're creating are trained or taught by appropriate people. And that way we can have them act autonomously in ways that may not be.
[1:25:58] Mr. Shapiro, you have 45 seconds if you'd like to respond. You've raised so many issues that I'll pass on this one. Mr. Issa, you're now recognized for round two. Thank you. While you were doing that, I was listening to the activation of 9000. I know I've made something. Well, it's not 2001 anymore, but it is.
[1:26:48] You know, this dialogue on AI, this portion of it, I think is particularly for people who saw that movie new, is important because Hal was able to correspond, able to have a dialogue, but it didn't have to answer honestly, and it didn't have to be, if you will, proofed. In other words, nobody put in the ability and the algorithm for it to be queried and to...
[1:27:17] So I think for all of us that are having this dialogue, and for those of you working in it, the question is, will we make sure that we have open algorithms, ones that can be queried and as a result can be diagnosed? And if we don't have them, then what you have to rely on, as was being said, is outcome. Outcome's not acceptable.
[1:27:39] Outcome is why the IRS shut down yesterday and wasn't taking your tax returns on the tax day, and all they knew was they had to fix it, but they didn't know why it happened, or at least it happened in a portion of their system. So that's something that I can't do, we can't mandate. But there are some things that hopefully we can work on jointly, and Mr. Shapiro, I...
[1:28:03] Nearly 100 years ago, the Radio Manufacturers Association formed, and one of the things it began to do was standard setting. Earlier today, you talked about we should have a standard, if you will, for what am I disclosing, platinum, gold, silver. You had a way of saying it. My question to you is, where are the responsible parties?
[1:28:27] As the radio manufacturers, now CTA, 100 years ago, who begin saying, if we're going to have the flourishing of technology, we're going to have to have standards. Privacy standards are complex.
[1:28:42] But how do you make them simple? Well, you make them simple by building standards that are predictable, that people can share a decision process with their friends. Yes, I always go for silver if it's my medical and gold if it's my financial. You alluded to it. How do we get there knowing it's not going to be mandated from this side of the dais? Or at least we certainly couldn't come up with the examples.
[1:29:06] Well, I mean, that's not the only choice. I mean, there's the executive branch, the FTC is comfortable in that area, and sometimes... Wait a second. I know the FTC. They're very comfortable after something goes bad telling us it went wrong. How often do they actually predictively able to say what the, quote, industry standard is before something? They certainly haven't done it in data intrusions. Well, there are...
[1:29:33] They do have a history of providing guidelines and standards, and I'm not advocating that. What I'm saying is on the issue of privacy and click-on, there are so many different systems out there that I am not personally convinced that the industry could come forward together without some concern that government would instead. I think it's always preferable for government and industry to work.
[1:29:55] But sometimes the concern that government will act does drive industry to act. That's just the reality. In this area, that cat's out of the bag a long time ago, and we're all clicking on stuff we don't understand. And that may have been one of the issues even in the Facebook disclosures and things like that, which I think caused some concern, is that we're agreeing on things that we don't understand. I mean, I used to read that stuff. I've stopped a long time ago. It's just you can't read it or understand it.
[1:30:24] Gary, back to what we were talking about in the last round. When we look at the health care and at personal information related to your health care, your drugs, your body weight, whatever it is, those are not such a large and complex hypothetical. Those are fairly definable. If we want to have the benefits of group data,
[1:30:52] such as a Fitbit gives and other data, and yet protect individual privacy.
[1:31:00] We should be able to demand, be produced, and then codified, hopefully, with some part. I mean, the FTC is very good if you produce a standard that's saying that's now the industry standard. They're less good at defining it proactively. Well, thank you for raising that specific case. We have done that as an industry. We've come up with standards. They are voluntary. And we haven't heard about...
[1:31:22] Any database that I'm aware of in the personal wearable area because I think that was a model that came together. The automobile industry is doing something similar and other industries are doing it. It's not too late. I was just talking about the online click agreements. There's opportunity in other areas and I think to move forward and to move forward quickly, it's an opportunity. The advantage for the companies is they're kind of safe in the herd if they follow the herd. And I'm going to cut you off but not finish this.
[1:31:50] On one door down there in judiciary, we control the question of limiting liability for certain behavior or not limiting it. Where, in your opinion, and I can take the rest of you if the chairman will allow, where is it we need to act to show that if you live by those best practices, knowing that just like that thing I played, it will not be 100%, but...
[1:32:14] If you live by those practices, your liability is in some way limited. In other words, non-punitive if you're doing the right thing. Because right now Congress has not acted fully to protect those who would like to enter this industry. Well, you've asked the question and answered it at the same time. Obviously, being in business yourself, you understand that...
[1:32:34] Risk and uncertainty are factors. We're seeing that in the trade problems we face today, the potential tariffs, the uncertainty. I did have that told to me just last night by somebody who knows about the question of not setting prices for their customers for Christmas because they don't yet know what the tariff will be.
[1:32:52] So uncertainty in the business environment, we're seeing it increasingly reflect in the stock market, but in terms of potential liability, our companies welcome certainty. And one thing, for example, Congress did when credit cards were introduced, they said your liability as an individual is limited to $50, and it all of a sudden allowed people to get over that uncertainty of going from cash to credit cards, and it helped.
[1:33:13] Grow our economy enormously and take a lot of friction out. We're facing some of the same things now as we go forward in so many different areas because of AI. And we do have an opportunity for Congress to address and say, if you follow these practices, you have a safe harbor here. But that's a very difficult thing to do, especially when it gets to debates about privacy and leaks and things like that. But everyone's looking for best practices, and the issues we were discussing earlier having to do with cyber and how do you protect. I mean, this is...
[1:33:39] Game will never end. You build a better mousetrap, you get smarter mice. So we're going to keep raising that bar, and that's the challenge that Congress will face. But some safe harbors would be certainly welcome in this area as it grows rapidly, and I think there's a role to play. And I think this is a great, amazing set of first three hearings to start on what will be a process with government and industry and consumers. Sorry, I hear that from all of you. I saw a lot of heads nodding.
[1:34:06] Safe harbors should exist if we're going to promote the advancement and use of data and artificial intelligence. Any no's? There's always a caveat, but any no's? I actually don't really have any comments about safe harbors specifically, but I think in general the issue of...
[1:34:28] Generating best practices is one which is really important to be considered in this field. And that, again, was sort of the reason why the partnership on artificial intelligence was created because there is a sort of understanding, I think, that's been come to in a collective sense about the necessity of determining what these guardrails should be to a certain extent. And I think that project can't really be undertaken without the policy community as well as other stakeholders who just necessarily need to be involved. Thank you, Mr. Chairman.
[1:34:57] I would also put myself down as embracing a caveat here, Congressman. I think one of the dangers is that we agree on a set of best practices that are not, in fact, anywhere near best practices, and we think our work is done. So while I support safe harbors if they align to practices that do protect privacy and advance security, I would suggest we are a long way from those practices in place today. So we should not lock in the status quo and think our work is done. Well, you know, I've owned Model Ts. I've owned...
[1:35:26] Cars from the 50s, 60s, 70s, 80s, and so on. I don't think we lock in best practices. We only lock them in for a product at the time that the product is new and innovative, and we have an expectation for the manufacturer that that product will become obsolete. Nobody assumes that a Model T is the safest vehicle or even a 66 Mustang, but we did...
[1:35:52] We do make expectations at the time of manufacturing. There was a time years ago when a man lost a limb on a lathe, and he sued the company, even though the lathe had been made in 1932, and it was already 50 years later. And we had to create a law that prohibited you from going back and using today's standards against the manufacturer. You could use it against...
[1:36:21] The company, if they hadn't updated, but you couldn't use it against the manufacturer, that's an example of safe harbor where if you make to the standards of the day, you are not held for the standards of change on a product that is a fire and forget. You don't own it or control it. And so that's what I was referring to. Your expectation that, yes, there has to be continuous innovation and that people have to stay up with the standards, of course, we're not expecting that. But then the question is...
[1:36:49] Will we see it from your side, or will we try to have the same people who have the system that shut down on the last tax filing day be the ones determining best practices? Thank you, Mr. Chairman. Would the gentleman engage in a colloquy? Of course. What is a Model T? No, I am joking. Well, you know, I just want you to know that when the big bang comes, the Model T is one of the vehicles that will still crank up and run.
[1:37:18] Well, I'm coming to your house, Congressman Isom. I have two final questions. The last question is actually a simple question. But the first question is, I recognize we can have a whole hearing on the topic. And I lump it generally in pre-crime. You have jails that are making decisions on whether someone should be released
[1:37:51] Decisions based on algorithms. We have people making decisions about whether they believe someone's going to potentially commit a crime in the future. And I would lump this in pre-crime. And the question is, should that be allowed? Gary? I'll foolishly take a shot at that. It depends on the risk involved. For example, in an airplane...
[1:38:17] Security situation, I think it makes sense to use biometrics and predictive technology and gait analysis and voice analysis and all the other tools that are increasingly available to predict whether someone's at risk on a flight. Israel does it increasingly, and it makes sense.
[1:38:36] We have a robot at OpenAI, and we trained it to try and reach towards this water bottle. And so we obviously expected that the robot would eventually grab the water bottle and pick it up.
[1:39:06] But what we discovered the robot had learned to do was to take the table the water bottle was on and just bring it towards itself, fulfilling the objective but not really in the way that we wanted it to. So I think I'd agree with Gary that maybe there are some specific areas where we're comfortable with certain levels of classification because the risk of getting it wrong, like a plane, is so high. But I think we should be incredibly cautious because this is a road where once you go down it, you're dealing with people's...
[1:39:35] And you can't, in the case of pre-crime, really inspect whether it's pulling that table towards it. It may be making completely bananas decisions and you're not going to have an easy way to find out and you've dealt with someone's life in the process. So I'd urge caution here.
[1:39:53] I'll say this with a caveat that I provided previously on other answers, which is that the partnership hasn't yet had a chance to formulate a position on this formally. But I think that this question speaks to a lot of the challenges associated with bias in the field right now, which we discussed a little bit earlier. And I think also the challenges of what happens...
[1:40:12] As a result of the decontextualization of technology and the application of it in areas where it may or may not be appropriate to have it be applied. So I think it's really important to consider the impacted communities, especially in the case of criminal justice applications. And I think that needs to be a required aspect of conversation about these issues. Dr. Buchanan. I'd echo Ms. Lyon's points and Mr. Clark's points. I would make three other points here. The first is that not only is there a risk of bias.
[1:40:41] But there's a risk. Sometimes machine learning is said to be money laundering for bias in that it takes a system that something is dirty in it and outputs it in this veneer of impartiality that comes from a computer. And we don't interrogate that system as much as we should. It's a major risk, I think, in this area, but in many areas. Secondly, I think you pose the question somewhat of a hypothetical. Mr. Clark is a measurer here, but I would encourage you and Mr. Clark to investigate.
[1:41:09] How much those systems are already in place. I think ProPublica did an excellent bit of reporting on bias in sentencing in the criminal justice system already in place today. And that would certainly deserve more attention, in my view. And the third is that we should make sure that the inputs to the system are transparent and the system itself is transparent. And one of my concerns, speaking generally here, is that the systems used for sentencing
[1:41:38] Now and in the future, often are held in a proprietary fashion. So it's very hard to interrogate them and understand how they work. And of course, hard in general to understand the outputs of such a system. And I think while that causes me concern in general, it should cause extreme concern in this case, if we're sentencing people on the basis of proprietary closed systems that we do not fully understand in public view.
[1:42:04] Thank you, Dr. Buchanan. And my last question is for the entire panel, and maybe, Dr. Buchanan, we start with you, and we'll work our way down. And it's real simple. Take 30 seconds to answer it. What would you all like to see from this committee in Congress when it comes to artificial intelligence in the future? Mr. Chairman, I think you've done a great job by holding this series of hearings, and I was encouraged...
[1:42:33] I think that the more you can do to force conversations like this out in the open and elevate them as a matter of policy discourse is important. I would suggest, as an academic, I view my job to think about topics that are important but are not urgent, that are coming but are not here in the next month or two. I would suggest that many committees in Congress should take that as a mandate as well.
[1:42:59] And I would encourage you to adopt that mindset as you approach AI. There are a lot of very important subjects in this field that will never reach the urgency of the next week or the next month, but will very quickly arrive and are still fundamentally important to virtually all of our society. At the risk of redundancy, I also want to say thank you for the engagement, Chairman. I think that...
[1:43:25] Having more of these types of conversations and more knowledge transfer between those working on technology and those governing it in fora like this is deeply important. And I think, again, I'd like to offer myself and the rest of the organizations in the partnership as a resource to whatever extent is possible in that project of education and further understanding.
[1:43:49] And think that it's deeply important for policymakers as well to consider the unique impact and role that they might have in technology governance, especially within the context of a multi-stakeholder setting, which is especially characteristic, I think, of the AI field right now. Thank you. Well, before we get to you, Mr. Clark and Mr. Shapiro, y'all aren't allowed to thank us, because I want to thank y'all. We have to prevent, as we've learned in the last couple of weeks,
[1:44:19] And many of our colleagues in both chambers are unfamiliar with basic things like social media. And so we have to elevate the common body of understanding on some of these topics. And so y'all's participation today, y'all's written statements, y'all's oral arguments help inform many of us on a topic that...
[1:44:44] When I went around the streets here in the Capitol and asked everybody, what is AI, most people, if they were older than me, described how. That's why I was laughing when Isa brought that in. And people that were younger than me referred to Ava from Ex Machina. And so y'all are helping educate us. So Mr. Clark, Mr. Shapiro, what should this committee in Congress be doing on AI?
[1:45:11] Until the first time I tried to build a table, I was a measure once, cut twice type of person. And then after I built that really terrible broken table, I became a measure twice, cut once person. The reason why I say that is that I think that if Congress and the agencies start to participate in more discussions like this, and we actually come to specific things that we need to measure, that we want to build around, like competitions, it will...
[1:45:39] Further understanding in both groups. There's a lot that the AI community can learn from these discussions, and I think the inverse is true as well. So I'd welcome that, and I think that's probably the best next step we can take. Mr. Shapiro, last word. I'm happy to embrace my colleagues' offers and views and appreciation. I have three quick suggestions. One, I think you should continue this, but go to field hearings at great places where there is technology, like Massachusetts or...
[1:46:08] Las Vegas in January, CES. Second, I think government plays a major role because government's a big buyer. In terms of procurement, I think you should focus on where AI could be used in procurement and set the goals and the results rather than focus on the very technical aspects of it. Third, in terms of, I think also that while Congress may not easily get legislation, it could have a sense of Congress. It could have a sense of Congress that it's important national goal that we cut...
[1:46:34] Automobile deaths or we do certain things by a certain date. And setting a national goal with or without the administration could be very valuable in terms of gathering the nation and moving us forward in a way which benefits everyone and really keeps our national lead in AI. That's a great way to end.
[1:46:54] Our series, I want to thank our witnesses for appearing before us today. The record is going to remain open for two weeks for any member to submit a written opening statement or questions for the record. If there's no further business, without objection, the subcommittee stands adjourned.