OpenAI CEO Sam Altman testifies at Senate artificial intelligence hearing — full video

[0:00] is on the oversight of artificial intelligence, the first in a series of hearings intended to [0:07] write the rules of AI. Our goal is to demystify and hold accountable those new technologies to [0:17] avoid some of the mistakes of the past. And now for some introductory remarks. [0:24] Too often we have seen what happens when technology outpaces regulation. The unbridled [0:33] exploitation of personal data, the proliferation of disinformation and the deepening of societal [0:41] inequalities. We have seen how algorithmic biases can perpetuate discrimination and prejudice [0:50] and how the lack of transparency can undermine public trust. This is not the future we want. [1:00] If you were listening from home, you might have thought that voice was mine and the words [1:06] from me. But in fact, that voice was not mine. The words were not mine. And the audio was [1:17] an AI voice cloning software trained on my floor speeches. The remarks were written by ChatGPT [1:26] when it was asked how I would open this hearing. And you heard just now the result. I asked ChatGPT, [1:40] why did you pick those themes and that content? And it answered, and I'm quoting, Blumenthal has a strong [1:49] record in advocating for consumer protection and civil rights. He has been vocal about issues such as data [1:55] privacy and the potential for discrimination in algorithmic decision making. Therefore, the statement [2:08] emphasizes these aspects. Mr. Altman, I appreciate ChatGPT's endorsement. In all seriousness, this apparent [2:19] reasoning is pretty impressive. I am sure that we'll look back in a decade and view ChatGPT [2:28] and GPT-4 like we do the first cell phone, those big clunky things that we used to carry around. But we [2:35] recognize that we are on the verge really of a new era. The audio and my playing it may strike you as [2:44] as curious or humorous. But what reverberated in my mind was, what if I had asked it and what if it had [2:55] provided an endorsement of Ukraine's surrendering or Vladimir Putin's leadership? That would have been [3:04] really frightening. And the prospect is more than a little scary, to use the word, Mr. Altman, you have used [3:11] yourself. And I think you have been very constructive in calling attention to the pitfalls, as well as the [3:18] promise. And that's the reason why we wanted you to be here today. And we thank you and our other witnesses for [3:24] joining us. For several months now, the public has been fascinated with GPT, DALI, and other AI tools. These [3:35] examples like the homework done by ChatGPT are the articles and op-eds that it can write feel like [3:43] novelties. But the underlying advancement of this era are more than just research experiments. They are [3:52] no longer fantasies of science fiction. They are real and present. The promises of curing cancer or [4:01] developing new understandings of physics and biology or modeling climate and weather, all very [4:09] encouraging and hopeful. But we also know the potential harms. And we've seen them already. [4:16] Weaponized disinformation, housing discrimination, harassment of women and impersonation fraud, [4:23] voice cloning, deep fakes. These are the potential risks, despite the other rewards. And for me, [4:36] perhaps the biggest nightmare is the looming new industrial revolution, the displacement of [4:42] millions of millions of workers, the loss of huge numbers of jobs, the need to prepare for this [4:51] new industrial revolution in skill training and relocation that may be required. And already, [4:58] industry leaders are calling attention to those challenges. To quote ChatGPT, [5:09] this is not necessarily the future that we want. We need to maximize the good over the bad. Congress [5:20] has a choice now. We had the same choice when we faced social media. We failed to seize that moment. [5:26] The result is predators on the internet, toxic content, exploiting children, creating dangers for them. [5:35] And Senator Blackburn and I and others like Senator Durbin on the Judiciary Committee are trying to deal [5:43] with it, Kids Online Safety Act. But Congress failed to meet the moment on social media. Now we have the [5:49] obligation to do it on AI before the threats and the risks become real. Sensible safeguards are not in [6:00] opposition to innovation. Accountability is not a burden. Far from it. They are the foundation of how [6:07] we can move ahead while protecting public trust. They are how we can lead the world in technology and [6:14] science, but also in promoting our democratic values. Otherwise, in the absence of that trust, I think we may [6:23] well lose both. These are sophisticated technology, but there are basic expectations common in our law. [6:31] We can start with transparency. AI companies ought to be required to test their systems, [6:37] disclose known risks, and allow independent researcher access. We can establish scorecards [6:43] and nutrition labels to encourage competition based on safety and trustworthiness. [6:51] Limitations on use. There are places where the risk of AI is so extreme that we ought to impose [6:57] restriction or even ban their use, especially when it comes to commercial invasions of privacy [7:04] for profit and decisions that affect people's livelihoods. And of course, accountability or [7:10] liability. When AI companies and their clients cause harm, they should be held liable. We should [7:17] not repeat our past mistakes. For example, Section 230. Forcing companies to think ahead and be [7:25] responsible for the ramifications of their business decisions can be the most powerful tool of all. [7:33] Garbage in, garbage out. The principle still applies. We ought to be aware of the garbage, [7:41] whether it's going into these platforms or coming out of them. And the ideas that we develop in this [7:51] hearing, I think, will provide a solid path forward. I look forward to discussing them with you today. [7:58] And I will just finish on this note. The AI industry doesn't have to wait for Congress. [8:05] I hope there are ideas and feedback from this discussion and from the industry and voluntary action, [8:11] such as we've seen lacking in many social media platforms. And the consequences have been huge. [8:20] So I'm hoping that we will elevate rather than have a race to the bottom. And I think these hearings will [8:29] be an important part of this conversation. This one is only the first. The ranking member and I have [8:36] agreed there should be more. And we're going to invite other industry leaders, some have committed [8:42] to come, experts, academics, and the public, we hope, will participate. And with that, [8:49] I will turn to the ranking member, Senator Hawley. Thank you very much, Mr. Chairman. [8:54] Thanks to the witnesses for being here. I appreciate that several of you had long journeys [8:59] to make in order to be here. I appreciate you making the time. I look forward to your testimony. [9:03] I want to thank Senator Blumenthal for convening this hearing, for being a leader on this topic. [9:07] You know, a year ago, we couldn't have had this hearing because the technology that we're talking [9:13] about had not burst into public consciousness. That gives us a sense, I think, of just how rapidly [9:19] this technology that we're talking about today is changing and evolving and transforming our world [9:26] right before our very eyes. I was talking with someone just last night, a researcher in the field [9:32] of psychiatry, who was pointing out to me that the chat GPT and generative AI, these large language models, [9:39] it's really like the invention of the internet in scale, at least, at least, and potentially far, [9:46] far more significant than that. We could be looking at one of the most significant technological [9:51] innovations in human history. And I think my question is, what kind of an innovation is it [9:57] going to be? Is it going to be like the printing press that diffused knowledge and power and learning [10:05] widely across the landscape that empowered ordinary, everyday individuals that led to greater flourishing, [10:13] that led above all to greater liberty? Or is it going to be more like the atom bomb? Huge technological [10:19] breakthrough, but the consequences, severe, terrible, continue to haunt us to this day. [10:27] I don't know the answer to that question. I don't think any of us in the room know the answer to that [10:31] question because I think the answer has not yet been written. And to a certain extent, it's up to us [10:36] here and to us as the American people to write the answer. What kind of technology will this be? How [10:43] will we use it to better our lives? How will we use it to actually harness the power of technological [10:49] innovation for the good of the American people, for the liberty of the American people, not for the [10:55] power of the few? You know, I was reminded of the psychologist and writer Carl Jung who said at the [11:03] beginning of the last century that our ability for technological innovation, our capacity for [11:09] technological revolution had far outpaced our ethical and moral ability to apply and harness [11:15] the technology we developed. That was a century ago. I think the story of the 20th century largely bore [11:21] him out. And I just wonder what will we say as we look back at this moment about these new technologies, [11:28] about generative AI, about these language models, and about the hosts of other AI capacities that are [11:34] even right now under development, not just in this country, but in China, the countries of our [11:39] adversaries and all around the world. And I think that the question that Jung posed is really the [11:44] question that faces us. Will we strike that balance between technological innovation and our ethical and [11:49] moral responsibility to humanity, to liberty, to the freedom of this country? And I hope that today's [11:57] hearing will take us a step closer to that answer. Thank you, Mr. Chairman. Thanks. Thanks, Senator Hawley. [12:03] I'm going to turn to the chairman of the Judiciary Committee and the ranking member, [12:08] Senator Graham, if they have opening remarks as well. Yes, Mr. Chairman. Thank you very much, [12:13] and Senator Hawley as well. Last week in this committee, full committee, Senate Judiciary Committee, [12:19] we dealt with an issue that had been waiting for attention for almost two decades, and that is [12:25] what to do with the social media when it comes to the abuse of children. [12:29] We had four bills initially that were considered by this committee, and what may be history in the [12:36] making, we passed all four bills with unanimous roll calls, unanimous roll calls. I can't remember [12:43] another time when we've done that and an issue that important. It's an indication, I think, of the [12:49] important position of this committee in the national debate on issues that affect every single family and [12:55] affect our future in a profound way. 1989 was a historic watershed year in America, because that's [13:03] when Seinfeld arrived, and we had a sitcom which was supposedly about little or nothing, which turned out to [13:10] be enduring. I like to watch it, obviously, and I'm always marvel when they show the phones that he used in 1989, [13:19] and I think about those in comparison to what we carry around in our pockets today. [13:24] It's a dramatic change. And I guess the question as I look at that is, does this change in phone [13:29] technology that we've witnessed through the sitcom really exemplify a profound change in America, [13:38] still unanswered? But the very basic question we face is whether or not this issue of AI [13:46] is a quantitative change in technology or a qualitative change. The suggestions that I've [13:53] heard from experts in the field suggest it's qualitative. Is it AI fundamentally different? [13:59] Is it a game changer? Is it so disruptive that we need to treat it differently than other forms of [14:05] innovation? That's the starting point. And the second starting point is one that's humbling, [14:10] and that is a fact when you look at the record of Congress in dealing with innovation, technology, [14:16] and rapid change were not designed for that. In fact, the Senate was not created for that purpose, [14:22] but just the opposite. Slow things down. Take a harder look at it. Don't react to public sentiment. [14:28] Make sure you're doing the right thing. Well, I've heard of the potential, the positive potential of [14:33] AI, and it is enormous. You can go through lists of the deployment of technology that would say that [14:40] an idea you can sketch for a website on a napkin can generate functioning code. Pharmaceutical companies [14:48] could use the technology to identify new candidates to treat disease. The list goes on and on. And then, of [14:55] course, the danger, and it's profound as well. So I'm glad that this hearing is taking place. I think it's important [15:02] for all of us to participate. I'm glad that it's a bipartisan approach. We're going to have to scramble to keep up with the pace of [15:08] innovation in terms of our government public response to it. But this is a great start. Thank you, Mr. Chairman. [15:15] Thanks. Thanks, Senator Urban. It is very much a bipartisan approach, very deeply and broadly [15:21] bipartisan. And in that spirit, I'm going to turn to my friend Senator Graham. Thank you. That was not written [15:34] by AI for sure. Let me introduce now the witnesses. We're very grateful to you for being here. Sam Altman is [15:44] the co-founder and CEO of OpenAI, the AI research and deployment company behind chat GPT and DALI. [15:55] Mr. Altman was president of the early stage startup accelerator Y Combinator from 1914, I'm sorry, 2014 [16:04] to 2019. OpenAI was founded in 2015. Christina Montgomery is IBM's vice president and chief privacy and trust [16:15] officer overseeing the company's global privacy program policies, compliance and strategy. She also [16:22] chairs IBM's AI ethics board, a multidisciplinary team responsible for the governance of AI and emerging [16:29] technologies. Christina has served in various roles at IBM, including corporate secretary to the company's [16:37] board of directors. She is a global leader in AI ethics and governments. And Ms. Montgomery also is a [16:45] member of the United States Chamber of Commerce AI Commission and the United States National AI [16:51] Advisory Committee, which was established in 2022 to advise the president and the national AI initiative [16:58] office on a range of topics related to AI. Gary Marcus is a leading voice in artificial intelligence. He's a [17:07] scientist, best-selling author and entrepreneur, founder of the robust AI and geometric AI acquired by Uber, [17:17] if I'm not mistaken, an emeritus professor of psychology and neuroscience at NYU. Mr. Marcus is well known for [17:26] his challenges to contemporary AI, anticipating many of the current limitations decades in advance and for his [17:34] research in human language development and cognitive neuroscience. Thank you for being here. And as you [17:43] may know, our custom on the Judiciary Committee is to swear in our witnesses before they testify. So if you [17:51] would all please rise and raise your right hand. You solemnly swear that the testimony that you are going to [18:00] give is the truth, the whole truth and nothing but the truth. So help with God. Thank you. Mr. Altman, [18:08] we're going to begin with you, if that's okay. Thank you. Thank you, Chairman Blumenthal, [18:14] Ranking Member Hawley, members of the Judiciary Committee, thank you for the opportunity to speak [18:18] to you today about large neural networks. It's really an honor to be here, even more so in the moment [18:22] than I expected. My name is Sam Altman. I'm the Chief Executive Officer of OpenAI. OpenAI was founded on the [18:29] belief that artificial intelligence has the potential to improve nearly every aspect of our [18:34] lives, but also that it creates serious risks we have to work together to manage. We're here because [18:41] people love this technology. We think it can be a printing press moment. We have to work together to [18:46] make it so. OpenAI is an unusual company, and we set it up that way because AI is an unusual technology. [18:53] We are governed by a non-profit and our activities are driven by our mission and our charter, which commit [18:59] us to working to ensure that the broad distribution of the benefits of AI and to maximizing the safe [19:04] of AI systems. We are working to build tools that one day can help us make new discoveries and address [19:10] some of humanity's biggest challenges, like climate change and curing cancer. Our current systems aren't [19:16] yet capable of doing these things, but it has been immensely gratifying to watch many people around [19:22] the world get so much value from what these systems can already do today. We love seeing people use our [19:28] tools to create, to learn, to be more productive. We're very optimistic that there are going to be [19:33] fantastic jobs in the future and that current jobs can get much better. We also love seeing what [19:39] developers are doing to improve lives. For example, Be My Eyes used our new multimodal technology in GPT-4 [19:46] to help visually impaired individuals navigate their environment. We believe that the benefits of the [19:53] tools we have deployed so far vastly outweigh the risks, but ensuring their safety is vital to our work, [19:58] and we make significant efforts to ensure that safety is built into our systems at all levels. [20:04] Before releasing any new system, OpenAI conducts extensive testing, engages external experts for [20:11] detailed reviews and independent audits, improves the model's behavior, and implements robust safety [20:17] and monitoring systems. Before we released GPT-4, our latest model, we spent over six months conducting [20:25] extensive evaluations, external red teaming, and dangerous capability testing. We are proud of the [20:31] progress that we made. GPT-4 is more likely to respond helpfully and truthfully, and refuse harmful requests, [20:38] than any other widely deployed model of similar capability. However, we think that regulatory [20:44] intervention by governments will be critical to mitigate the risks of increasingly powerful models. [20:49] For example, the US government might consider a combination of licensing and testing requirements [20:56] for development and release of AI models above a threshold of capabilities. There are several other [21:02] areas I mentioned in my written testimony where I believe that companies like ours can partner with [21:07] governments, including ensuring that the most powerful AI models adhere to a set of safety [21:12] requirements, facilitating processes to develop and update safety measures, and examining opportunities [21:18] for global coordination. And as you mentioned, I think it's important that companies have their own [21:23] responsibility here, no matter what Congress does. This is a remarkable time to be working on [21:28] artificial intelligence. But as this technology advances, we understand that people are anxious about how it [21:35] could change the way we live. We are too. But we believe that we can and must work together to identify [21:41] and manage the potential downsides so that we can all enjoy the tremendous upsides. It is essential that [21:48] powerful AI is developed with democratic values in mind, and this means that US leadership is critical. [21:54] I believe that we will be able to mitigate the risks in front of us and really capitalize on this [21:59] technology's potential to grow the US economy and the worlds. And I look forward to working with you all to meet [22:04] this moment. And I look forward to answering your questions. Thank you. Thank you, Mr. Altman. Ms. Montgomery. [22:11] Chairman Blumenthal, Ranking Member Hawley, and members of the subcommittee, thank you for today's [22:18] opportunity to present. AI is not new, but it's certainly having a moment. Recent breakthroughs in [22:26] generative AI and the technology's dramatic surge in the public attention has rightfully raised serious [22:33] questions at the heart of today's hearing. What are AI's potential impacts on society? What do we do about [22:41] bias? What about misinformation, misuse, or harmful content generated by AI systems? Senators, these are [22:49] the right questions, and I applaud you for convening today's hearing to address them head on. While AI may [22:55] be having its moment, the moment for government to play a role has not passed us by. This period of [23:02] focused public attention on AI is precisely the time to define and build the right guardrails to protect [23:09] people and their interests. But at its core, AI is just a tool, and tools can serve different purposes. [23:16] To that end, IBM urges Congress to adopt a precision regulation approach to AI. This means establishing [23:24] rules to govern the deployment of AI in specific use cases, not regulating the technology itself. [23:31] Such an approach would involve four things. First, different rules for different risks. [23:38] The strongest regulation should be applied to use cases with the greatest risks to people and society. [23:46] Second, clearly defining risks. There must be clear guidance on AI uses or categories of AI [23:52] supported activity that are inherently high risk. This common definition is key to enabling a clear [23:57] understanding of what regulatory requirements will apply in different use cases and contexts. [24:03] Third, be transparent. So AI shouldn't be hidden. Consumers should know when they're interacting with [24:08] an AI system and that they have recourse to engage with a real person should they so desire. No person [24:14] anywhere should be tricked into interacting with an AI system. And finally, showing the impact. [24:20] For higher risk use cases, companies should be required to conduct impact assessments that show how [24:26] their systems perform against tests for bias and other ways that they could potentially impact the public [24:31] and to attest that they've done so. By following risk-based, use-case-specific approach at the core of [24:38] precision regulation, Congress can mitigate the potential risks of AI without hindering innovation. [24:44] But businesses also play a critical role in ensuring the responsible deployment of AI. [24:50] Companies active in developing or using AI must have strong internal governance, including, among other things, [24:57] designating a lead AI ethics official responsible for an organization's trustworthy AI strategy, [25:04] standing up an ethics board or a similar function as a centralized clearinghouse for resources to help [25:10] guide implementation of that strategy. IBM has taken both of these steps and we continue calling on our [25:17] industry peers to follow suit. Our AI ethics board plays a critical role in overseeing internal AI governance [25:24] processes, creating reasonable guardrails to ensure we introduce technology into the world in a [25:29] responsible and safe manner. It provides centralized governance and accountability while still being [25:34] flexible enough to support decentralized initiatives across IBM's global operations. We do this because [25:41] we recognize that society grants our license to operate. And with AI, the stakes are simply too high. [25:48] We must build, not undermine the public trust. The era of AI cannot be another era of move fast and [25:55] break things. But we don't have to slam the brakes on innovation either. These systems are within our [26:01] control today, as are the solutions. What we need at this pivotal moment is clear, reasonable policy and [26:08] sound guardrails. These guardrails should be matched with meaningful steps by the business community to do [26:14] their part. Congress and the business community must work together to get this right. The American people [26:20] deserve no less. Thank you for your time and I look forward to your questions. [26:25] Thank you. Professor Marcus. [26:32] Thank you senators. Today's meeting is historic. I'm profoundly grateful to be here. I come as a [26:38] scientist, someone who's founded AI companies, and as someone who genuinely loves AI, but who is [26:42] increasingly worried. There are benefits, but we don't yet know whether they will outweigh the risks. [26:48] Fundamentally, these new systems are going to be destabilizing. They can and will create persuasive [26:54] lies at a scale humanity has never seen before. Outsiders will use them to affect our elections, [27:00] insiders to manipulate our markets and our political systems. Democracy itself is threatened. [27:05] Chatbots will also clandestinely shape our opinions, potentially exceeding what social media can do. [27:11] Choices about data sets that AI companies use will have enormous unseen influence. Those who choose the [27:17] data will make the rules, shaping society in subtle but powerful ways. There are other risks too, [27:23] many stemming from the inherent unreliability of current systems. A law professor, for example, [27:29] was accused by a chatbot of sexual harassment. Untrue. And it pointed to a Washington Post article [27:35] that didn't even exist. The more that that happens, the more that anybody can deny anything. As one [27:41] prominent lawyer told me on Friday, defendants are starting to claim that plaintiffs are making up [27:46] legitimate evidence. These sorts of allegations undermine the abilities of juries to decide what or who to [27:51] believe and contribute to the undermining of democracy. Poor medical advice could have serious consequences [27:57] too. An open source large language model recently seems to have played a role in a person's decision [28:02] to take their own life. The large language model asked the human, if you wanted to die, [28:07] why didn't you do it earlier? And then followed up with, were you thinking of me when you overdosed? [28:11] Without ever referring the patient to the human health that was obviously needed. Another system [28:16] rushed out and made available to millions of children told a person posing as a 13-year-old [28:22] how to lie to her parents about a trip with a 31-year-old man. Further threats continue to emerge [28:28] regularly. A month after GPT-4 was released, OpenAI released chat GPT plugins which quickly led others [28:34] to develop something called auto GPT with direct access to the internet, the ability to write source code, [28:40] and increased powers of automation. This may well have drastic and difficult to predict security [28:45] consequences. What criminals are going to do here is to create counterfeit people. It's hard to even [28:51] envision the consequences of that. We have built machines that are like bulls in a china shop, [28:56] powerful, reckless, and difficult to control. We all more or less agree on the values we would like [29:03] for our AI systems to honor. We want, for example, for our systems to be transparent, to protect our [29:08] privacy, to be free of bias, and above all else, to be safe. But current systems are not in line with [29:14] these values. Current systems are not transparent. They do not adequately protect our privacy. And [29:19] they continue to perpetuate bias. And even their makers don't entirely understand how they work. [29:24] Most of all, we cannot remotely guarantee that they're safe. And hope here is not enough. The big [29:30] tech company's preferred plan boils down to trust us. But why should we? The sums of money at stake are [29:35] mind-boggling. Emissions drift. OpenAI's original mission statement proclaimed, [29:39] our goal is to advance AI in the way that is most likely to benefit humanity as a whole, [29:44] unconstrained by a need to generate financial return. Seven years later, they're largely beholden [29:50] to Microsoft, embroiled in part in an epic battle of search engines that routinely make things up. [29:55] And that's forced Alphabet to rush out products and de-emphasize safety. Humanity has taken a back seat. [30:02] AI is moving incredibly fast, with lots of potential, but also lots of risks. We obviously need government [30:07] involved. And we need the tech companies involved, both big and small. But we also need independent [30:13] scientists. Not just so that we scientists can have a voice, but so that we can participate directly [30:18] in addressing the problems and evaluating solutions. And not just after products are released, but before. [30:23] And I'm glad that Sam mentioned that. We need tight collaboration between independent [30:26] scientists and governments in order to hold the company's feet to the fire. Allowing independent [30:31] scientists access to these systems before they are widely released, as part of a clinical trial like [30:40] safety evaluation, is a vital first step. Ultimately, we may need something like CERN, global, international, [30:47] and neutral, but focused on AI safety rather than high energy physics. We have unprecedented [30:53] opportunities here, but we are also facing a perfect storm of corporate irresponsibility, widespread [30:59] deployment, lack of adequate regulation, and inherent unreliability. AI is among the most world-changing [31:05] technologies ever, already changing things more rapidly than almost any technology in history. [31:10] We acted too slowly with social media. Many unfortunate decisions got locked in [31:15] with lasting consequence. The choices we make now will have lasting effects for decades, maybe even [31:21] centuries. The very fact that we are here today in bipartisan fashion to discuss these matters [31:27] gives me some hope. Thank you, Mr. Chairman. [31:29] Thanks very much, Professor Marcus. We're going to have seven-minute rounds of questioning, [31:34] and I will begin. First of all, Professor Marcus, we are here today because we do face that perfect [31:41] storm. Some of us might characterize it more like a bomb in a China shop, not a bull. And as Senator Hawley [31:51] indicated, there are precedents here, not only the atomic warfare era, but also the genome project, the research [32:01] on genetics, where there was international cooperation as a result. And we want to avoid those past mistakes, [32:09] as I indicated in my opening statement, that we're committed on social media. That is precisely the reason [32:15] we are here today. ChatGPT makes mistakes. All AI does. And it can be a convincing liar, what people [32:25] call hallucinations. That might be an innocent problem in the opening of a judiciary subcommittee [32:33] hearing where a voice is impersonated, mine in this instance, or quotes from research papers that [32:41] don't exist. But ChatGPT and BARD are willing to answer questions about life or death matters, [32:48] for example, drug interactions. And those kinds of mistakes can be deeply damaging. I'm interested [32:55] in how we can have reliable information about the accuracy and trustworthiness of these models, [33:03] and how we can create competition and consumer disclosures that reward greater accuracy. [33:10] The National Institutes of Standards and Technology actually already has an AI accuracy test, [33:18] the face recognition vendor test. It doesn't solve for all the issues with facial recognition, [33:26] but the scorecard does provide useful information about the capabilities and flaws of these systems. So [33:33] there's work on models to assure accuracy and integrity. My question, let me begin with you, [33:42] Mr. Altman is, should we consider independent testing labs to provide scorecards and nutrition labels, [33:50] or the equivalent of nutrition labels, packaging that indicates to people whether or not the content [33:58] can be trusted, what the ingredients are, and what the garbage going in may be, because it could [34:05] result in garbage going out? Yeah, I think that's a great idea. I think that companies should put their own, [34:13] sort of, you know, here are the results of our test of our model before we release it. Here's where [34:18] it has weaknesses, here's where it has strengths, but also independent audits for that are very [34:23] important. These models are getting more accurate over time. You know, this is, this is, as we have, [34:30] I think, said as loudly as anyone, this technology is in its early stages. It definitely still makes mistakes. [34:36] We find that people, that users are pretty sophisticated and understand where the mistakes [34:42] are that they need, are likely to be, that they need to be responsible for verifying what the models [34:48] say, that they go off and check it. I worry that as the models get better and better, the users can [34:56] have sort of less and less of their own discriminating thought process around it. But, but I think users are more [35:02] capable than we give, often give them credit for in, in conversations like this. I think a lot of [35:07] disclosures, which if you've used ChatGBT, you'll see about the inaccuracies of the model are also [35:13] important. And I am, I'm excited for a world where companies publish with the models information about [35:23] how they behave, where the inaccuracies are, and independent agencies or companies provide that as well. [35:28] I think it's a great idea. I alluded in my opening remarks to the, the jobs issue, the economic [35:35] effects on employment. I think you have said, in fact, and I'm going to quote, development of [35:44] superhuman machine intelligence is probably the greatest threat to the continued existence of [35:50] humanity, end quote. You may have had in mind the effect on, on jobs, which is really my biggest [35:58] nightmare in the long term. Let me ask you what your biggest nightmare is and whether you share that [36:09] concern. Like with all technological revolutions, I expect there to be significant impact on jobs, [36:17] but exactly what that impact looks like is very difficult to predict. If we went back to the, [36:22] the other side of a previous technological revolution, talking about the jobs that exist on the other side, [36:27] you know, you can go back and read books of this. It's what people said at the time. It's difficult. [36:33] I believe that there will be far greater jobs on the other side of this. And the jobs of today will get better. [36:39] I think it's important. First of all, I think it's important to understand and think about GPT-4 as a tool, [36:45] not a creature, which is easy to get confused. And it's a tool that people have a great deal of control [36:51] over and how they use it. And second, GPT-4 and things, other systems like it, are good at doing tasks, [37:00] not jobs. And so you see already people that are using GPT-4 to do their job much more efficiently [37:08] by helping them with tasks. Now, GPT-4 will, I think, entirely automate away some jobs, [37:15] and it will create new ones that we believe will be much better. This happens, again, my understanding of [37:21] the history of technology is one long technological revolution, not a bunch of different ones put [37:25] together. But this has been continually happening. We, as our quality of life raises and as machines [37:32] and tools that we create can help us live better lives, the bar raises for what we do. And our human [37:38] ability and what we spend our time going after goes after more ambitious, more satisfying projects. [37:43] So there will be an impact on jobs. We try to be very clear about that. And I think it will require [37:49] partnership between the industry and government, but mostly action by government to figure out how [37:54] we want to mitigate that. But I'm very optimistic about how great the jobs of the future will be. [38:00] Thank you. Let me ask Ms. Montgomery and Professor Marcus for your [38:05] reaction to those questions as well. Ms. Montgomery. [38:08] On the jobs point, yeah. I mean, well, it's a hugely important question. And it's one that we've been [38:15] talking about for a really long time at IBM. You know, we do believe that AI, and we've said it [38:21] for a long time, is going to change every job. New jobs will be created. Many more jobs will be [38:27] transformed. And some jobs will transition away. I'm a personal example of a job that didn't exist when [38:35] I joined IBM. And I have a team of AI governance professionals who are in new roles that we created, [38:41] you know, as early as three years ago. I mean, they're new and they're growing. So I think the [38:46] most important thing that we could be doing and can and should be doing now is to prepare the workforce [38:53] of today and the workforce of tomorrow for partnering with AI technologies and using them. And we've been [39:00] very involved for years now in doing that, in focusing on skills-based hiring, in educating [39:09] for the skills of the future. Our skills build platform has seven million learners and over a [39:15] thousand courses worldwide focused on skills. And we've pledged to train 30 million individuals [39:22] by 2030 in the skills that are needed for society today. Thank you. Professor Marcus. May I go back to [39:29] the first question as well? Absolutely. On the subject of nutrition labels, I think we absolutely need to [39:35] do that. I think that there's some technical challenges and that building proper nutrition [39:39] labels goes hand in hand with transparency. The biggest scientific challenge in understanding [39:44] these models is how they generalize. What do they memorize and what new things do they do? [39:49] The more that there's in the data set, for example, the thing that you want to test accuracy on, [39:53] the less you can get a proper read on that. So it's important, first of all, that scientists be part of [39:58] that process. And second, that we have much greater transparency about what actually goes into [40:02] these systems. If we don't know what's in them, then we don't know exactly how well they're doing [40:06] when we give something new and we don't know how good a benchmark that will be for something that's [40:10] entirely novel. So I could go into that more, but I want to flag that. Second is on jobs, past performance [40:18] history is not a guarantee of the future. It has always been the case in the past that we have had more [40:24] jobs, that new jobs, new professions come in as new technologies come in. I think this one's going to [40:29] be different. And the real question is, over what time scale? Is it going to be 10 years? Is it going [40:34] to be 100 years? And I don't think anybody knows the answer to that question. I think in the long run, [40:39] so-called artificial general intelligence really will replace a large fraction of human jobs. [40:44] We're not that close to artificial general intelligence. Despite all of the media hype and so [40:48] forth, I would say that what we have right now is just a small sampling of the AI that we will build. [40:53] In 20 years, people will laugh at this, as I think it was Senator Hawley made the, [40:58] but maybe Senator Durbin made the example about this. It was Senator Durbin made the example about [41:02] cell phones. When we look back at the AI of today, 20 years ago, we'll be like, wow, that stuff was [41:08] really unreliable. It couldn't really do planning, which is an important technical aspect. Its reasoning [41:13] was ability, and reasoning abilities were limited. But when we get to AGI, artificial general intelligence, [41:19] maybe let's say it's 50 years, that really is going to have, I think, profound effects on labor. [41:25] And there's just no way around that. And last, I don't know if I'm allowed to do this, but I will [41:28] note that Sam's worst fear I do not think is employment, and he never told us what his worst [41:33] fear actually is, and I think it's germane to find out. Thank you. I'm going to ask Mr. Altman if he [41:43] cares to respond. Yeah. Look, we have tried to be very clear about the magnitude of the risks here. [41:49] I think jobs and employment and what we're all going to do with our time really matters. I agree [41:57] that when we get to very powerful systems, the landscape will change. I think I'm just more [42:01] optimistic that we are incredibly creative, and we find new things to do with better tools, [42:07] and that will keep happening. My worst fears are that we cause significant, we, the field, [42:13] the technology, the industry, cause significant harm to the world. I think that could happen in a lot [42:19] of different ways. It's why we started the company. It's a big part of why I'm here today, [42:24] and why we've been here in the past, and we've been able to spend some time with you. [42:28] I think if this technology goes wrong, it can go quite wrong, and we want to be vocal about that. [42:34] We want to work with the government to prevent that from happening, but we try to be very clear-eyed [42:40] about what the downside case is and the work that we have to do to mitigate that. [42:44] Thank you. And our hope is that the rest of the industry will follow the example [42:50] that you and IBM, Ms. Montgomery, have set by coming today and meeting with us, [42:58] as you have done privately, in helping to guide what we're going to do so that we can target [43:04] the harms and avoid unintended consequences to the good. Thank you. Senator Hawley. [43:13] Thank you again, Mr. Chairman. Thanks to the witnesses for being here. [43:15] Mr. Altman, I think you grew up in St. Louis, if I'm not mistaken. [43:18] It's great to see a fellow- Missouri is a great place. [43:20] It is. Thank you. I want that noted, especially underlined in the record. Missouri is a great place. [43:25] That is the takeaway from today's hearing. Maybe we just stop there, Mr. Chairman. [43:29] Let me ask you, Mr. Altman, I think I'll start with you, and I'll just preface this by saying my [43:34] questions here are an attempt to get my head around and to ask all of you to help us to get our heads [43:38] around what this generative AI, particularly the large language models, what it can do. [43:44] So I'm trying to understand its capacities and then its significance. So I'm looking at a paper here [43:50] entitled Large Language Models Trained on Media Diets Can Predict Public Opinion. This is just posted [43:55] about a month ago. The authors are Chu, Andreas, Encelabere, and Roy. And their conclusion, this work [44:03] was done at MIT and then also at Google. The conclusion is that large language models can [44:08] indeed predict public opinion. And they go through and model why this is the case. And they [44:15] conclude ultimately that an AI system can predict human survey responses by adapting a pre-trained [44:20] language model to subpopulation-specific media diets. In other words, you can feed the model a [44:25] particular set of media inputs and it can, with remarkable accuracy, and the paper goes into this, [44:30] predict then what people's opinions will be. I want to think about this in the context of elections. [44:37] If these large language models can, even now, based on the information we put into them, quite [44:44] accurately predict public opinion, you know, ahead of time. I mean, predict. It's before you even ask [44:50] the public these questions. What will happen when entities, whether it's corporate entities or whether [44:57] it's governmental entities or whether it's campaigns or whether it's foreign actors, take this [45:03] survey information, these predictions about public opinion, and then fine-tuned strategies [45:08] to elicit certain responses, certain behavioral responses. I mean, we already know this committee [45:12] is her testimony, I think three years ago now, about the effect of something as prosaic, it now seems, [45:19] as Google search. The effect that this has on voters in an election, particularly undecided voters [45:25] in the final days of an election who may try to get information from Google search. And what an enormous [45:30] effect, the ranking of the Google search, the articles that it returns has become an enormous [45:35] effect on an undecided voter. This, of course, is orders of magnitude, far more powerful, far more [45:40] significant, far more directive, if you like. So, Mr. Altman, maybe you can help me understand here what [45:48] some of the significance of this is. Should we be concerned about models that can, large language models, [45:53] that can predict survey opinion, and then can help organizations, entities, fine-tuned strategies to [45:59] elicit behaviors from voters? Should we be worried about this for our elections? [46:02] MR. Yeah. Thank you, Senator Hawley, for the question. It's one of my areas of greatest concern. [46:08] The more general ability of these models to manipulate, to persuade, to provide sort of one-on-one, [46:15] you know, interactive disinformation. I think that's like a broader version of what you're talking about, but [46:20] given that we're going to face an election next year and these models are getting better. [46:24] I think this is a significant area of concern. I think there's a lot of policies that companies [46:30] can voluntarily adopt, and I'm happy to talk about what we do there. I do think some regulation would [46:37] be quite wise on this topic. Someone mentioned earlier, it's something we really agree with, [46:42] people need to know if they're talking to an AI, if content that they're looking at might be generated [46:47] or might not. I think it's a great thing to do, is to make that clear. I think we also will need [46:53] rules, guidelines about what's expected in terms of disclosure from a company providing a model [47:04] that could have these sorts of abilities that you talk about. So I'm nervous about it. I think people [47:12] are able to adapt quite quickly. When Photoshop came onto the scene a long time ago, you know, [47:18] for a while people were really quite fooled by photoshopped images and then pretty quickly [47:24] developed an understanding that images might be photoshopped. This will be like that but on steroids [47:30] and the interactivity, the ability to really model, predict humans well as you talked about, [47:37] I think is going to require a combination of companies doing the right thing, regulation and public [47:45] education. Professor Marcus, do you want to address this? Yeah, I'd like to add two things. One is in the [47:50] appendix to my remarks. I have two papers to make you even more concerned. One is in the Wall Street [47:57] Journal just a couple days ago called Help, My Political Beliefs Were Altered by a Chatbot. And I [48:02] think the scenario you raised was that we might basically observe people and use surveys to figure [48:08] out what they're saying. But as Sam just acknowledged, the risk is actually worse, that the systems will [48:13] directly, maybe not even intentionally, manipulate people. And that was the thrust of the Wall Street Journal [48:18] article. And it links to an article that I've also linked to called Interacting, and it's not yet [48:23] published, not yet peer reviewed. Interacting with opinionated language models changes users views. [48:28] And this comes back ultimately to data. One of the things that I'm most concerned about with GPT-4 is [48:34] that we don't know what it's trained on. I guess Sam knows, but the rest of us do not. And what it is [48:39] trained on has consequences for essentially the biases of the system. We could talk about that in technical [48:44] terms. But how these systems might lead people about depends very heavily on what data is trained [48:50] on them. And so we need transparency about that. And we probably need scientists in there doing [48:56] analysis in order to understand what the political influences, for example, of these systems might be. [49:01] And it's not just about politics. It can be about health. It could be about anything. These systems [49:06] absorb a lot of data. And then what they say reflects that data. And they're going to do it differently [49:12] depending on what's in that data. So it makes a difference if they're trained on the Wall Street [49:16] Journal as opposed to the New York Times or Reddit. I mean, actually, they're largely trained on all [49:21] of this stuff. But we don't really understand the composition of that. And so we have this issue [49:25] of potential manipulation. And it's even more complex than that because it's subtle manipulation. [49:31] People may not be aware of what's going on. That was the point of both the Wall Street Journal [49:35] article and the other article that I called your attention to. [49:38] Let me ask you about AI systems trained on personal data, the kind of data that, for instance, [49:44] the social media companies, the major platforms, Google, Meta, etc., collect on all of us routinely. [49:51] We've had many a chat about this in this committee over many a year now. But the massive amounts of [49:56] data, personal data that the companies have on each one of us, an AI system that is trained on [50:03] that individual data that knows each of us better than ourselves and also knows the billions of data [50:09] points about human behavior, human language interaction generally, wouldn't we be able, [50:14] wouldn't we, can't we foresee an AI system that is extraordinarily good at determining what will grab human [50:22] attention and what will keep an individual's attention? And so for the war for attention, the war for [50:29] cliques that is currently going on on all of these platforms, that's how they make their money. [50:34] I'm just imagining an AI system, these AI models supercharging that war for attention such that we [50:40] now have technology that will allow individual targeting of a kind we have never even imagined [50:46] before, where the AI will know exactly what Sam Altman finds attention grabbing, will know exactly what [50:51] Josh Hawley finds attention grabbing, will be able to elicit, to grab our attention and then elicit [50:56] responses from us in a way that we have heretofore not even been able to imagine. Should we be [51:01] concerned about that for its corporate applications, for the monetary applications, for the manipulation [51:06] that could come from that? Mr. Altman. [51:08] MR. Yes, we should be concerned about that. To be clear, OpenAI does not, we're not off, you know, [51:16] we wouldn't have an ad-based business model, so we're not trying to build up these profiles of our [51:19] users. We're not trying to get them to use it more. Actually, we'd love it if they use it less, [51:24] because we don't have enough GPUs. But I think other companies are already, and certainly will in [51:30] the future, use AI models to create, you know, very good ad predictions of what a user will like. [51:36] I think that's already happening in many ways. MR. Mr. Marcus, anything you want to add? [51:40] MR. Yes, and perhaps Ms. Montgomery will want to, too, as well. I don't know. But [51:45] hyper-targeting of advertising is definitely going to come. I agree that that's not been OpenAI's [51:51] business model. Of course, now they're working for Microsoft, and I don't know what's in Microsoft's [51:56] thoughts. But we will definitely see it. Maybe it will be with open source language models. I don't [52:02] know. But the technology there is, let's say, part way there to being able to do that, and we'll [52:07] certainly get there. MR. So we're an enterprise technology company, not consumer-focused, so the space [52:17] isn't one that we necessarily operate in, in terms of, but these issues are hugely important issues. [52:23] And it's why we've been out ahead in developing the technology that will help to ensure that you can [52:34] do things like produce a fact sheet that has the ingredients of what your data is trained on, [52:40] data sheets, model cards, all those types of things, and calling for, as I've mentioned today, [52:46] transparency. So you know what the algorithm was trained on. And then you also know and can manage [52:53] and monitor continuously over the life cycle of an AI model, the behavior and the performance of [52:59] that model. MR. Senator Durbin. [53:03] Thank you. I think what's happening today in this hearing room is historic. I can't recall when we've had [53:11] people representing large corporations or private sector entities come before us and plead with us to [53:19] regulate them. In fact, many people in the Senate have based their careers on the opposite, that the [53:28] economy will thrive if government gets the hell out of the way. And what I'm hearing instead today is the [53:35] that stop me before I innovate again message. And I'm just curious as to how we're going to achieve [53:44] this. As I mentioned Section 230 in my opening remarks, we learned something there. We decided [53:52] that in Section 230 that we were basically going to absolve the industry from liability for a period of [54:00] time as it came into being. Well, Mr. Altman, on the podcast earlier this year, you agreed with host [54:08] Kara Swisher that Section 230 doesn't apply to generative AI and that developers like OpenAI should [54:16] not be entitled to full immunity for harms caused by their products. So what have we learned from 230 [54:23] that applies to your situation with AI? Thank you for the question, Senator. I don't know yet [54:30] exactly what the right answer here is. I'd love to collaborate with you to figure it out. [54:34] I do think for a very new technology, we need a new framework. Certainly companies like ours bear [54:41] a lot of responsibility for the tools that we put out in the world, but tool users do as well. [54:46] And how we want, and also people that will build on top of it between them and the end consumer, [54:52] and how we want to come up with a liability framework there is a super important question. And we'd love [54:59] to work together. The point I want to make is this. When it came to online platforms, [55:03] the inclination of the government was get out of the way. This is a new industry. Don't over-regulate [55:09] it. In fact, give them some breathing space and see what happens. I'm not sure I'm happy with the [55:15] outcome as I look at online platforms. Me either. And the harms that they have created, [55:21] problems that we've seen demonstrated in this committee, child exploitation, cyber bullying, [55:26] online drug sales, and more. I don't want to repeat that mistake again. And what I hear is the [55:32] opposite suggestion from the private sector. And that is, come in on the front end of this thing [55:37] and establish some liability standards, precision regulation. For a major company like IBM to come [55:44] before this committee and say to the government, please regulate us. Can you explain the difference [55:51] in thinking from the past and now? [55:53] Yeah, absolutely. So for us, this comes back to the issue of trust and trust in the technology. [56:01] Trust is our license to operate, as I mentioned in my remarks. And so we firmly believe, and we've [56:07] been calling for precision regulation of artificial intelligence for years now. This is not a new [56:12] position. We think that technology needs to be deployed in a responsible and clear way. That people, [56:20] we've taken principles around that, trust and transparency, we call them, are principles that [56:25] were articulated years ago and build them into practices. That's why we're here advocating for [56:30] precision regulatory approach. So we think that AI should be regulated at the point of risk, [56:36] essentially. And that's the point at which technology meets society. [56:40] Let's take a look at what that might appear to be. Members of Congress are a pretty smart lot of people, [56:47] maybe not as smart as we think we are many times. And government certainly has a capacity to do [56:53] amazing things. But when you talk about our ability to respond to the current challenge and perceived [57:00] challenge of the future, challenges which you all have described in terms which are hard to [57:07] forget. As you said, Mr. Altman, things can go quite wrong. As you said, Mr. Marcus, democracy is [57:14] threatened. I mean, the magnitude of the challenge you're giving us is substantial. I'm not sure that [57:21] we respond quickly and with enough expertise to deal with it. Professor Marcus, you made a reference to [57:28] CERN, the international arbiter of nuclear research, I suppose. I don't know if that's a fair characterization, [57:36] but it's a characterization I'll start with. What is it, what agency of this government do you think [57:42] exists that could respond to the challenge that you've laid down today? We have many agencies that [57:50] can respond in some ways. For example, the FTC, the FCC. There are many agencies that can, but my [57:59] view is that we probably need a cabinet level organization within the United States in order to [58:06] address this. And my reasoning for that is that the number of risks is large. The amount of information [58:15] to keep up on is so much. I think we need a lot of technical expertise. I think we need a lot of [58:21] coordination of these efforts. So there is one model here where we stick to only existing law and try to [58:28] shape all of what we need to do, and each agency does their own thing. But I think that AI is going to be [58:33] such a large part of our future and is so complicated and moving so fast. And this does not fully solve [58:40] your problem about a dynamic world. But it's a step in that direction to have an agency that's full-time job [58:47] is to do this. I personally have suggested, in fact, that we should want to do this in a global way. [58:52] I wrote an article in The Economist. I have a link in here, an invited essay for The Economist, [58:57] suggesting we might want an international agency for AI. [59:01] That's the point I wanted to go to next, and that is the fact that I'll get it aside from the CERN and [59:06] nuclear examples, because government was involved in that from day one, at least in the United States. [59:12] But now we're dealing with innovation which doesn't necessarily have a boundary. [59:17] That's right. [59:18] We may create a great U.S. agency, and I hope that we do, [59:22] that may have jurisdiction over U.S. corporations and U.S. activity, [59:26] but doesn't have a thing to do with what's going to bombard us from outside the United States. [59:31] How do you give this international authority the authority to regulate in a fair way for all [59:38] entities involved in AI? [59:40] I think that's probably over my pay grade. [59:43] I would like to see it happen, and I think it may be inevitable that we push there. [59:47] I mean, I think the politics behind it are obviously complicated. [59:51] I'm really heartened by the degree to which this room is bipartisan and supporting the same things, [59:57] and that makes me feel like it might be possible. [59:59] I would like to see the United States take leadership in such an organization. [1:00:03] It has to involve the whole world and not just the U.S. to work properly. [1:00:07] I think even from the perspective of the companies, it would be a good thing. [1:00:11] So the companies themselves do not want a situation where you take these models, [1:00:15] which are expensive to train, and you have to have 190 some of them, one for every country. [1:00:21] That wouldn't be a good way of operating. [1:00:23] When you think about the energy costs alone, just for training these systems, [1:00:26] it would not be a good model if every country has its own policies. [1:00:30] And for each jurisdiction, every company has to train another model. [1:00:35] And maybe different states are different. [1:00:37] So Missouri and California have different rules. [1:00:40] And so then that requires even more training of these expensive models with huge climate impact. [1:00:45] And it would be very difficult for the companies to operate if there was no global coordination. [1:00:50] And so I think that we might get the companies on board. [1:00:53] If there's bipartisan support here, and I think there's support around the world, [1:00:57] it is entirely possible that we could develop such a thing. [1:01:00] But obviously, there are many nuances here of diplomacy that are over my pay grade. [1:01:05] I would love to learn from you all to try to help make that happen. [1:01:08] Mr. Altman. [1:01:09] Can I weigh in just briefly? [1:01:11] Briefly, please. [1:01:12] I want to echo support for what Mr. Marcus said. [1:01:14] I think the U.S. should lead here and do things first. [1:01:17] But to be effective, we do need something global. [1:01:21] As you mentioned, this can happen everywhere. [1:01:23] There is precedent. [1:01:25] I know it sounds naive to call for something like this, and it sounds really hard. [1:01:28] There is precedent. [1:01:28] We've done it before with the IAEA. [1:01:31] We've talked about doing it for other technologies. [1:01:35] Given what it takes to make these models, the chip supply chain, [1:01:39] the sort of limited number of competitive GPUs, the power the U.S. has over these companies, [1:01:44] I think there are paths to the U.S. setting some international standards that other countries [1:01:50] would need to collaborate with and be part of that are actually workable, even though it sounds [1:01:54] on its face like an impractical idea. [1:01:57] And I think it would be great for the world. [1:01:59] MR. Thank you, Mr. Chairman. [1:02:00] Thanks, Senator Durbin. [1:02:02] And in fact, I think we're going to hear more about what Europe is doing. [1:02:06] The European Parliament already is acting on an AI act. [1:02:12] On social media, Europe is ahead of us. [1:02:16] We need to be in the lead. [1:02:18] I think your point is very well taken. [1:02:22] Let me turn to Senator Graham. [1:02:24] Senator Blackburn. [1:02:27] MS. Thank you, Mr. Chairman, and thank you all for being here with us today. [1:02:32] I put into my CHAT-GPT account, should Congress regulate AI CHAT-GPT, and it gave me four pros, [1:02:40] four cons, and says ultimately the decision rests with Congress and deserves careful consideration. [1:02:47] So on that you know it was very balanced. I recently visited with the [1:02:53] Nashville Technology Council. I represent Tennessee and of course you had people [1:02:59] there from health care, financial services, logistics, educational entities [1:03:04] and they're concerned about what they see happening with AI with the [1:03:10] utilizations for their companies. Ms. Montgomery you know some similar to you [1:03:14] they've got health care people are looking at disease analytics, they are [1:03:19] looking at predictive diagnosis, how this can better the outcomes for patients, [1:03:25] logistics industry looking at ways to save time and money and yield [1:03:30] efficiencies. You've got financial services that are saying how does this [1:03:36] work with quantum, how does it work with blockchain, how can we use this, but it [1:03:44] I think as we have talked with them Mr. Chairman one of the things that [1:03:48] continues to come up is yes Professor Marcus as you were saying the EU different [1:03:54] entities are ahead of us in this, but we have never established a federally preempt, [1:04:02] given preemption for online privacy, for data security and put some of those [1:04:07] foundational elements in place which is something that we need to do as we look [1:04:14] at this and it will require that Commerce Committee, Judiciary Committee decide how we [1:04:21] move forward so that people own their virtual you and Mr. Altman I was glad to see [1:04:31] last week that your open AI models are not going to be trained using consumer data I [1:04:39] think that that is important and if we have a second round I've got a host of [1:04:43] questions for you on data security and privacy but I think it's important to let [1:04:51] people control their virtual you their information in these settings and I want [1:04:57] to come to you on music and content creation because we've got a lot of [1:05:02] songwriters and artists and a I think we have the best creative community on the [1:05:08] face of the earth there in Tennessee and they should be able to decide if [1:05:15] their copyrighted songs and images are going to be used to train these models and I'm [1:05:23] concerned about open AI's jukebox it offers some re-renditions in the style of [1:05:31] Garth Brooks which suggests that open AI is trained on Garth Brooks songs I went in [1:05:39] this weekend and I said write me a song that sounds like Garth Brooks and it gave [1:05:44] me a different version of simple man so it's interesting that it would do that but [1:05:52] you're training it on these copyrighted songs these MIDI files these sound [1:05:57] technologies so as you do this who owns the right to that AI generated material [1:06:06] and using your technology could I remake a song insert content from my favorite [1:06:17] artist and then on the creative rights to that song thank you senator this is an [1:06:25] area of great interest to us I would say first of all we think that creators [1:06:30] deserve control over how their creations are used and what happens sort of beyond [1:06:38] the point of them releasing it into the world second I think that we need to [1:06:44] figure out new ways with this new technology that creators can win succeed [1:06:48] have a vibrant life and I'm optimistic that this will present it let me ask you [1:06:53] this how do you compensate the art the artist that's exactly what I was gonna say [1:06:58] okay um we'd like to we're working with artists now visual artists musicians to [1:07:03] figure out what people want there's a lot of different opinions unfortunately and [1:07:06] at some point we'll have let me ask you this do you favor something like sound [1:07:10] exchange that has worked in the area of radio I'm not familiar with some exchange [1:07:17] I'm sorry I mean okay you've got your team behind you get back to me on that that [1:07:22] would be a third party entity okay so let's discuss that let me move on can you [1:07:30] commit as you've done with consumer data not to train chat GPT open AI jukebox or [1:07:38] other AI models on artists and songwriters copyrighted works or use their voices and [1:07:47] their likenesses without first receiving their consent so first of all jukebox is not a product [1:07:54] we offer that was a research release but it's not you know unlike chat GPT or dollar but we've [1:07:58] lived through Napster yes but that was something that really cost a lot of artists a lot of money [1:08:08] oh I understand yeah for sure digital distribution era I don't I don't know [1:08:13] the numbers on jukebox on top of my head as a research release I can I can follow up with your [1:08:17] office but it's not jukebox is not something that gets much attention or usage it was put [1:08:21] out to show that something's possible well Senator Durbin just said you know and I think [1:08:26] it's a fair warning to you all if we're not involved in this from the get-go and you all already are a [1:08:35] long way down the path on this but if we don't step in then this gets away from you so are you [1:08:42] working with a copyright office are you considering protections for content generators and creators [1:08:51] in generative AI yes we are absolutely engaged on that again to reiterate my earlier point we think [1:08:56] that content creators content owners need to benefit from this technology exactly what the economic model is we're [1:09:03] still talking to artists and content owners about what they want I think there's a lot of ways this [1:09:07] can happen but very clearly no matter what the law is the right thing to do is to make sure people get [1:09:12] significant upside benefit from this new technology and we believe that it's really going to deliver [1:09:18] that but that content owners likenesses people totally deserve control over how that's used and to [1:09:24] benefit from it okay so on privacy then how do you plan to account for the collection of voice and other [1:09:33] user-specific data things that are copyrighted righted user-specific data through your AI applications [1:09:42] because if I can go in and say write me a song that sounds like Garth Brooks and it takes part of an [1:09:48] existing song there has to be a compensation to that artist for that utilization and that use if it was [1:09:58] radio play it would be there if it was streaming it would be there so if you're going to do that what is [1:10:06] your policy for making certain you're accounting for that and you're protecting that individual's right to [1:10:15] privacy and their right to secure that data and that created work so a few thoughts about this [1:10:24] number one we think that people should be able to say I don't want my personal data trained on [1:10:29] that's I think that's right that gets to a national privacy law which many of us here on the dais are [1:10:35] working toward getting something that we can use yeah I think strong privacy my time's expired [1:10:41] okay thank you mr. chair thanks senator blackburn senator clover thank you very much mr. chairman um [1:10:47] and uh uh senator blackburn I love nashville love tennessee love your music but i will say i use chat gpt [1:10:55] and just asked what are the top creative song artists of all time and two of the top three were from [1:11:01] minnesota that would be uh prince i'm sure they mentioned bob dylan okay all right so let us let us continue [1:11:09] one thing ai won't change and you're seeing it here all right so on a more serious note though [1:11:16] my staff and i um in my role as chair of the rules committee and leading a lot of the election bill [1:11:21] and we just introduced a bill that's uh representative yvette clark from new york introduced over the house [1:11:28] senator booker and bennett and i did um on political advertisements but that is just of course the tip of [1:11:34] the iceberg you know this from your discussions with senator holly and others about the images and my [1:11:39] own view um center grams of section 230 is that we just can't let people make stuff up and then not [1:11:47] have any consequence but i'm going to focus in on what my job one of my jobs will be on the rules [1:11:52] committee and that is election misinformation and we just asked chat gpt to do a tweet about a polling [1:12:01] location in bloomington minnesota and said there are long lines at this polling location um at atonement [1:12:07] lutheran church where should we go now albeit it's not an election right now but the answer the tweet [1:12:14] that was drafted was a completely fake thing go to one two three four elm street and so you can imagine [1:12:22] what i'm concerned about here with an election upon us with primary elections upon us that we're going to [1:12:29] have all kinds of misinformation and i just want to know what you're planning on doing it doing about [1:12:34] it i know we're going to have to do something soon not just for the images of the candidates but also [1:12:40] for misinformation about the actual polling places and election rules thank you senator that we we [1:12:49] talked about this a little bit earlier we are quite concerned about the impact this can have on [1:12:53] elections i think this is an area where hopefully the entire industry and the government can work [1:12:57] together quickly there's there's many approaches and i'll talk about some of the things we do but before [1:13:03] that i think it's tempting to use the frame of social media um but this is not social media this [1:13:11] is different and so the the response that we need is different you know this is a tool that a user is [1:13:17] using to help generate content more efficiently than before they can change it they can test the [1:13:22] accuracy of it if they don't like it they can get another version um but it still then spreads [1:13:28] through social media or other ways like chat gpt is a you know single player experience where you're just [1:13:34] using this um and so i think as we think about what to do that's that's important to understand [1:13:41] there's a lot that we can and do do there um there's things that the model refuses to generate [1:13:47] we have policies uh we also importantly have monitoring so at scale uh we can detect someone [1:13:54] generating a lot of those tweets even if generating one tweet is okay yeah and of course there's going [1:13:59] to be other platforms and if they're all spouting out fake election information i just i think [1:14:05] what happened um in the past with russian interference and like it's just going to be a tip of the [1:14:10] iceberg when some of those fake ads so that's number one number two um is the impact on intellectual [1:14:18] property and senator blackburn was getting at some of this with song rights um and have serious concerns [1:14:25] about that but news content um so senator kennedy and i have a bill uh that was really quite straightforward [1:14:32] that would simply uh allowed uh the um the news organizations and exemption to be able to negotiate [1:14:40] with basically google and facebook microsoft was supportive of the bill but basically negotiate [1:14:46] with them to get better rates and um be able to not have some leverage and other countries are doing [1:14:53] this australia and the like and so my question is when we already have a study by northwestern predicting that [1:14:59] one-third of the u.s newspapers are that roughly existed two decades are going to go are going to be [1:15:05] gone by 2025 unless you start compensating for everything from book movies books yes but also news [1:15:14] content we're going to lose any realistic content producers and so i'd like your response to that and [1:15:21] of course there is an exemption for copyright in section 230 but i think asking little newspapers to go out [1:15:29] and sue all the time just can't be the answer they're not going to be able to keep up yeah like it is [1:15:36] my hope that tools like what we're creating can help news organizations do better i think having a [1:15:41] vibrant having a vibrant national media is critically important and let's call it round one of the internet [1:15:48] has not been great for that right we're talking here about local that you know report on your high [1:15:52] school for sure scores and a scandal in your city council those kinds of things for sure they're the [1:15:56] ones that are actually getting the worst the little radio stations and broadcast but do you understand [1:16:03] that this could be exponentially worse in terms of local news content if they're not compensated [1:16:09] well because what they need is to be compensated for their content and not have it stolen yeah again [1:16:15] our our model you know the current version of uh gpt-4 ended training in 2021 it's not it's not it's not a [1:16:23] good way to find recent news uh and it's i don't think it's a service that can do a great job of [1:16:28] linking out although maybe with our plugins it's it's possible uh if there are things that we can do [1:16:33] to help local news we would certainly like to again i think it's it's critically important okay um [1:16:38] one last may i add something there yeah but let me just ask you a question you can combine them quick [1:16:42] more transparency on the platforms um senator coons and senator cassidy and i have the platform [1:16:49] accountability transparency act uh to give researchers access to this information of the [1:16:55] algorithms and the like on social media data would that be helpful and then why don't you just say yes [1:17:00] or no and then go at his uh the transparency is absolutely critical here to understand the political [1:17:06] ramifications the bias ramifications and so forth we need transparency about the data we need to know more [1:17:11] about how the models work we need to have scientists have access to them i was just going to amplify your [1:17:16] earlier point about uh local news a lot of news is going to be generated by these systems they're [1:17:21] not reliable newsguard already is a study i'm sorry it's not in my appendix but i will get it to your [1:17:25] office um showing that something like 50 websites are already um generated by bots um we're going to see [1:17:33] much much more of that and it's going to make it even more competitive for the local news organizations [1:17:37] and so the quality of the sort of overall news market is going to decline as we have more generated [1:17:43] content by systems that aren't actually reliable in the content that generated thank you and thank [1:17:47] you on a very timely basis to make the argument why we have to mark up this bill again in june i [1:17:52] appreciate it thank you senator graham thank you mr chairman and senator holley for having this [1:17:59] i'm trying to find out how it is different than social media and learn from the mistakes we've made with [1:18:05] social media uh the idea of not suing social media companies is to allow the internet to flourish because [1:18:11] and if i slander you uh you can sue me uh if you're a billboard company and you put up the slander can [1:18:20] you sue the billboard company we said no basically section 230 is being used by social media companies [1:18:28] to hide to avoid liability uh for activity that other people generate uh when they refuse to comply with [1:18:37] their terms of use a mother calls up the company and says this app is being used to bully my child to [1:18:43] death you promise in the terms of use you would prevent bullying and she calls three times she gets [1:18:51] no response the child kills herself and they can't sue do you all agree we don't want to do that again [1:18:59] yes if i may speak for one second there's a fundamental distinction between reproducing content and generating [1:19:09] content yeah but you you would like liability where people are harmed absolutely yes in fact ibm has [1:19:18] been publicly advocating to condition liability on a reasonable care standard so let me just make sure [1:19:24] i understand the law as it exists today mr allman thank you for coming your company is not claiming [1:19:30] that section 230 applies the tool you have created yeah we're claiming we need to work together to find a [1:19:37] totally new approach i don't think section 230 is even the right framework okay so under the law it [1:19:45] exists today this tool you create if i'm harmed by it can i sue you that is beyond my area of legal [1:19:52] expertise have you ever been sued not for that no have you ever been sued at all uh that your company [1:19:58] yeah open a i get sued huh yeah we've gotten sued before okay and what for um i mean they've mostly [1:20:07] been like pretty frivolous things like i think happens to any company but like the examples my [1:20:13] colleagues have given from artificial intelligence that could literally ruin our lives and can we go [1:20:19] to the company that created that tool and sue them is that your understanding yeah i think there needs to [1:20:25] be clear responsibility by the companies but you're not claiming any kind of legal protection like section [1:20:33] 230 applies to your industry is that correct no i don't think we're i don't i don't think we're saying [1:20:38] anything mr marcus um when it comes to consumers there seems to be like three time tested ways to [1:20:44] protect consumers against any product um statutory schemes which are non-existent here legal systems [1:20:55] which may be appear here but not social media and agencies go back to senator holly the atom bomb [1:21:05] has put a cloud over humanity but nuclear power could be one of the solutions to climate change [1:21:13] so what i'm trying to do is make sure that you just can't go build a nuclear power plant hey bob what [1:21:19] would you like to do today let's go build a nuclear power plant you have a nuclear regulatory commission [1:21:23] that governs how you build a plant and is licensed do you agree mr ultimate that these tools you're [1:21:30] creating should be licensed yeah we've been calling for this we think any that's the simplest way [1:21:36] you get a license and do you agree with me that the simplest way and the most effective way is have an [1:21:42] agency that is more nimble and smarter than congress which should be easy to create [1:21:48] overlooking what you do yes we'd be enthusiastic about that you agree with that mr marcus absolutely [1:21:53] do you agree with that miss montgomery i would have some nuances i think we need to build on what we [1:21:59] have in place already today we don't have an agency regulators wait a minute nope nope nope we don't have [1:22:05] an agency that regulates the technology so should we have one but a lot of the issues i i don't think [1:22:12] so a lot of the okay wait a minute wait a minute so ibm says we don't need an agency uh interesting should [1:22:20] we have a license required for these tools so so what we believe is that we need to that's a simple [1:22:26] question should you get a license to produce one of these tools i think it comes back to some of them [1:22:32] potentially yes so what i said at the onset is that we need to um clearly define risks do you claim [1:22:40] section 230 applies in this area at all we're not a platform company and we've again long advocated for [1:22:47] reasonable care standard in section i just don't understand how you could say that you don't need an [1:22:52] agency to deal with the most transformative technology maybe ever well i i think we have existing is this a [1:23:03] transformative technology that can disrupt life as we know it good and bad i think it's a [1:23:08] transformative technology certainly and the conversations that we're having here today have [1:23:13] been really bringing to light the fact that this is the domains and the issues this one with you has [1:23:18] been very enlightening to me mr altman why are you so willing to have an agency senator we've been clear [1:23:28] about what we think the upsides are and i think you can see from users how much they enjoy how much value [1:23:33] they're getting out of it but we've also been clear about what the downsides are and so that's [1:23:36] why we think we need an agency system it's a major tool to be used by a lot it's a major new technology [1:23:41] if you build it'll be yeah if you make a ladder and the ladder doesn't work you ensue the people [1:23:46] made the ladder but there's some standards out there to make a ladder so that's why we're agreeing [1:23:50] with you yeah that's right i think you're on the right track so here's what my two cents worth for [1:23:55] the committee is that we need to empower an agency that issues and a license and can take it away [1:24:04] wouldn't that be some yes incentive to do it right if you could actually be taken out of business [1:24:10] clearly that should be part of what an agency can do now and you also agree that china's doing [1:24:16] ai research is that right correct this world organization that doesn't exist maybe it will [1:24:23] but if you don't do something about the china part of it you'll never quite get this right do you agree [1:24:29] well that that's why i think it doesn't necessarily have to be a world organization but there has to be [1:24:34] some sort of and there's a lot of options here there has to be some sort of standard some sort [1:24:38] of set of controls that do have global effect you know because you know other people doing this i got [1:24:43] 15 military application how can ai change the warfare and you got one minute i got one minute yeah [1:24:53] all right this is that's a tough question for one minute um this is very far out of my area of [1:24:59] expertise uh but i give you one example a drone can a drone you program you can plug into a drone the [1:25:05] coordinates and it can fly out and it goes over this target and it drops a missile on this car [1:25:11] moving down the road and somebody's watching it could ai create a situation where a drone can select [1:25:17] the target itself i think we shouldn't allow that well can it be done sure thanks thanks senator graham [1:25:26] senator coon thank you senator blumenthal senator holly for convening this hearing for working closely [1:25:32] together to come up with this compelling a panel of witnesses and beginning a series of hearings on [1:25:39] this transformational technology we recognize the immense promise and substantial risks associated [1:25:46] with generative ai technologies we know these models can make us more efficient help us learn [1:25:52] new skills open whole new vistas of creativity but we also know that generative ai [1:25:59] can authoritatively deliver wildly incorrect information it can hallucinate as is often [1:26:06] described it can impersonate loved ones it can encourage self-destructive behaviors and it can shape [1:26:13] public opinion and the outcome of elections congress thus far has demonstrably failed to responsibly [1:26:19] enact meaningful regulation of social media companies with serious harms that have resulted that we don't [1:26:26] fully understand senator klobuchar referenced in her questioning a bipartisan bill that would open up [1:26:32] social media platforms underlying algorithms we have struggled to even do that to understand the [1:26:39] underlying technology and then to move towards responsible regulation we cannot afford to be as late [1:26:46] to responsibly regulating generative ai as we have been to social media because the consequences both [1:26:53] positive and negative will exceed those of social media by orders of magnitude so let me ask a few [1:26:59] questions designed to get at both how we assess the risk what's the role of international regulation and how [1:27:06] does this impact ai mr altman i appreciate your testimony about the ways in which open ai assesses the safety [1:27:14] of your models through a process of iterative deployment the fundamental question embedded in that process [1:27:19] though is how you decide whether or not a model is safe enough to deploy and safe enough to have been built and then [1:27:26] let go into the wild i understand one way to prevent generative ai models from providing harmful content [1:27:34] is to have humans identify that content and then train the algorithm to avoid it there's another approach [1:27:40] that's called constitutional ai that gives the model a set of values or principles to guide its decision making [1:27:47] would it be more effective to give models these kinds of rules instead of trying to require or compel [1:27:53] training the model on all the different potentials for harmful content thank you senator it's it's a [1:27:59] great question uh i'd like to frame it by talking about why we deploy at all like why we put these [1:28:04] systems out into the world there's the obvious answer about there's benefits and people are using it for [1:28:10] all sorts of wonderful things and getting great value and that makes us happy but a big part of why we do it is [1:28:16] that we believe that iterative deployment and giving people in our institutions and you all time to come [1:28:23] to grips with this technology to understand it to find its limitations that benefits the regulations we [1:28:28] need around it what it takes to make it safe that's really important going off to build a super powerful [1:28:34] ai system in secret and then dropping it on the world all at once i think would not go well so a big [1:28:40] part of our strategy is while these systems are still relatively weak and deeply imperfect to find [1:28:46] ways to get people to have experience with them to have contact with reality and to figure out what we [1:28:53] need to do to make it safer and better and that is the only way that i've seen in the history of new [1:28:59] technology and products of this magnitude to get to a very good outcome and so that that interaction with [1:29:05] the world is very important now of course before we put something out uh it needs to meet a bar of [1:29:11] safety and uh and again we spent well over six months with gpt4 after we finished training it going [1:29:18] through all of these different things uh and deciding also what the standards were going to be before we [1:29:23] put something out there trying to find the harms that we knew about uh put it and and and how to address [1:29:28] those one of the things that's been gratifying to us is even some of our biggest critics have looked at gpt4 and [1:29:33] said wow open ai made huge progress on could focus briefly on whether or not a constitutional model [1:29:38] that gives values would be worth it i was just about to get there all right sorry about that um yeah i [1:29:44] think giving the models values up front uh is an extremely important set uh you know rlhf is another [1:29:52] way of doing that same thing but somehow or other you are with synthetic data or human generated data [1:29:57] you're saying here are the values here's what i want you to reflect or here are the wide bounds of [1:30:03] everything that society will allow and then within there you pick as the user you know if you want [1:30:07] value system over here a value system over there um we think that's very important there's multiple [1:30:12] technical approaches but we need to give policy makers and the world as a whole the tools to say [1:30:19] here's the values and implement them thank you ms montgomery you serve on an ai ethics board of a [1:30:25] long established company that has a lot of experience with ai i'm really concerned that generative ai [1:30:31] technologies can undermine the faith of democratic values and the institutions that we have the [1:30:36] chinese are insisting that ai as being developed in china reinforce the core values of the chinese [1:30:43] communist party and the chinese system and i'm concerned about how we promote ai that reinforces [1:30:49] and strengthens open markets open societies and democracy in your testimony you're advocating [1:30:55] for ai regulation tailored to the specific way the technology is being used not the underlying [1:31:00] technology itself and the eu is moving ahead with an ai act which cater categorizes ai products based [1:31:07] on level of risk you all in different ways have said that you view elections and the shaping of [1:31:12] election outcomes and disinformation that can influence elections as one of the highest risk cases [1:31:18] one that's entirely predictable we have attempted so far unsuccessfully to regulate social media after the [1:31:25] demonstrably harmful impacts of social media on our last several elections what advice do you have [1:31:30] for us about what kind of approach we should follow and whether or not the eu direction is the right [1:31:35] one to pursue i mean the conception of the eu ai act is very consistent with this concept of precision [1:31:44] regulation where you're regulating the use of the technology in context so absolutely that approach [1:31:51] makes a ton of sense it's what i advocated for at the onset different rules for different risks so in the [1:31:58] case of elections absolutely any algorithm being used in that context should be required to have disclosure [1:32:06] around the data being used the performance of the model um anything along those lines um is really [1:32:13] important guardrails need to be in place and and on the point just come back to the question of of [1:32:17] whether we need an independent agency i mean i think we don't want to slow down regulation to address real risks [1:32:26] right now right so we have existing regulatory authorities in place who have been clear that [1:32:33] they have the ability to regulate in their respective domains a lot of the issues we're talking about [1:32:38] today span multiple domains elections and the like so if i could um i'll just assert that those existing [1:32:45] regulatory bodies and authorities are under resourced and lack many of the statutory regulatory powers that [1:32:51] they need correct we have failed to deliver on data privacy even though industry has been asking us to [1:32:57] regulate data privacy if i might mr marcus i'm i'm interested also what international bodies are best [1:33:04] positioned to convene multilateral discussions to promote responsible standards we've talked about a [1:33:09] model being cern and nuclear energy i'm concerned about proliferation and non-proliferation we've also [1:33:17] talked i would suggest that the ipcc a un body helped at least provide a scientific baseline of [1:33:24] what's happening in climate change so that even though we may disagree about strategies and globally [1:33:29] we've come to a common understanding of what's happening and what should be the direction of [1:33:34] intervention i'd be interested mr marcus if you could just give us um your thoughts on who's the [1:33:39] right body internationally to convene a conversation and one that could also reflect our values i'm still [1:33:46] feeling my way on that issue i think global politics is not my specialty i'm an ai researcher [1:33:53] but i i have moved towards policy in in recent months really um because of my great concern about [1:33:58] all of these risks i think certainly the un unesco and uh has its guidelines should be involved and at [1:34:05] the table and and maybe things work under them and maybe they don't but they should have a strong [1:34:09] voice and and help to develop this the oecd has also been thinking um greatly about this [1:34:15] number of organizations have internationally i i don't feel like i personally am qualified to [1:34:20] say exactly what the right model is there well thank you i i think we need to pursue this both at [1:34:25] the national level and the international level i'm the chair of the ip subcommittee of the judiciary [1:34:30] committee in june and july we will be having hearings on the impact of ai on patents and copyrights [1:34:36] you can already tell from the questions of others there will be a lot of interest i look forward to [1:34:40] following up with you about that topic i know mr chairman i look forward to helping as much as [1:34:44] possible thank you very much thanks senator coons senator kennedy thank thank you all for being here [1:34:53] permit me to share with you three hypotheses that i would like you to assume for the moment to be true [1:35:03] hypothesis number one many members of congress do not understand artificial intelligence hypothesis number [1:35:16] two that absence of understanding may not prevent congress from plunging in with enthusiasm and trying to [1:35:31] regulate this technology in a way that could hurt this technology hypothesis number three that i would like you [1:35:41] to assume there is likely a berserk wing of the artificial intelligence community that intentionally or [1:35:55] unintentionally could use artificial intelligence to kill all of us and hurt us the entire time that [1:36:05] we are dying assume all of those to be true please tell me in plain english two or three reforms regulations [1:36:18] if any that you would you would implement if you were queen or king for a day ms montgomery i think it comes [1:36:33] back again to transparency and explainability in ai we absolutely need to know and have companies attest what [1:36:41] do you mean by transparency so disclosure of the data that's used to train ai disclosure of the model and how [1:36:51] it performs and making sure that there's continuous governance over these models that we are the leading [1:36:58] edge in governance technology governance governance organizational governance um rules and clarification [1:37:08] that are needed that this which rules i mean this is your chance folks to tell us how to get this right [1:37:14] please use it right i mean i think again the rules should be focused on the use of ai in certain contexts so [1:37:22] if you look at for example so if you look at the eu ai act it has certain um uses of ai that it says are [1:37:31] just simply too dangerous and will be outlawed in the okay so we ought to first pass the law that says you [1:37:37] can use ai for these uses but not others is that is that what you're saying we need to define the highest [1:37:43] risk use is anything else um and then of course uh requiring things like impact assessments and [1:37:50] transparency requiring companies to show their work um protecting data that's used to train ai in the [1:37:59] first place as well professor marcus if you could be specific this is your shot man talk in plain english [1:38:08] and tell me what if any rules we ought to implement and please don't just use concepts i'm looking for [1:38:16] specificity number one a safety review like we use with the fda prior to widespread deployment if you're [1:38:24] going to uh introduce something to 100 million people somebody has to have their eyeballs on it there [1:38:30] you go okay that's a good one number i'm not sure i agree with it but that's a good one what else you [1:38:35] didn't ask for three that you would agree with number two a nimble monitoring agency to follow [1:38:41] what's going on not just pre-review but also post as things are out there in the world with authority [1:38:47] to call things back which we've discussed today and number three would be funding geared towards things [1:38:53] like ai constitution ai that can reason about what it's doing i would not leave things entirely to [1:39:00] current technology which i think is poor at behaving in ethical fashion and behaving in honest [1:39:05] fashion um and so i would have funding to try to basically focus on ai safety research that term [1:39:11] has a lot of complications in my field um there's both safety let's say short term and long term and [1:39:17] i think we need to look at both rather than just funding models to be bigger which is the popular thing [1:39:22] to do we need to find models to be more trustworthy because i want to hear from mr altman mr altman here's [1:39:28] your shot thank you senator uh number one i would form a new agency that licenses any effort above a [1:39:33] certain scale of capabilities and can take that license away and ensure compliance with safety [1:39:38] standards number two i would create a set of safety standards focused on what you said in your third [1:39:43] hypothesis as the dangerous capability evaluations one example that we've used in the past is looking [1:39:48] to see if a model can self-replicate and self-exfiltrate into the wild we can give your office a long other [1:39:53] list of the things that we think are important there but specific tests that a model has to pass before [1:39:58] it can be deployed into the world and then third uh i would require independent audits so not just [1:40:03] from the company or the agency but experts who can say the model is or isn't in compliance with these [1:40:08] stated safety thresholds and these percentages of performance on question x or y can you send me [1:40:13] that information we will do that um would you be qualified to uh to to uh if we promulgated those [1:40:22] rules to administer those rules i love my current job cool are there people out there that would be [1:40:30] qualified we'd be happy to send you recommendations for people out there yes okay you make a lot of [1:40:35] money do you i make no i'm paid enough for health insurance i have no equity in open ai really that's [1:40:40] interesting you need a lawyer i need a what you need a lawyer or an agent i i'm doing this because [1:40:47] i love it thank you mr chairman thanks senator kennedy uh senator hirono thank you mr chairman listening [1:41:00] to to all of you testifying thank you very much for being here clearly ai truly is a game changing tool [1:41:10] and we need to get the regulation of this tool right because my staff for example asked ai it might have [1:41:20] been gpt4 it might have been i don't know one of the other entities to create a song that my favorite band [1:41:29] bts a favorite song a song that they would sing somebody else's song but you know neither of the [1:41:36] artists were involved in creating what sounded like a really genuine song so you can do a lot [1:41:43] we also asked can can there be a speech created talking about the supreme court decision and dobs and [1:41:51] the chaos that it created using my voice my kind of voice and it created a speech that was really good [1:41:57] it almost made me think about you know what do i need my staff for so don't worry that's not where [1:42:03] we're nervous laughter behind you their jobs are safe but there's so much that can be done and one [1:42:09] of the things that you mentioned mr altman altman that intrigued me was you said gpt4 can refuse harmful [1:42:17] requests so you must have put some thought into how your system if i can call it that can refuse harmful [1:42:24] requests what what do you consider a harmful request you can just keep it short yeah uh i'll give a few [1:42:31] examples uh one would be about violent content another would be about content that's encouraging [1:42:36] self-harm another's adult content not that we think adult content is inherently harmful but there's [1:42:41] things that could be associated with that that we cannot reliably enough differentiate so we refuse all [1:42:45] of it so those are some of the more obvious harmful kinds of uh of information but in the election [1:42:51] context for for example i saw a picture of a former president trump being arrested by nypd and that went [1:42:59] viral i don't know is that considered harmful i've seen all kinds of statements attributed to any one of [1:43:06] us that could be put out there that may not be that may not rise to your level of harmful content but [1:43:12] there you have it so two of two of you said that we should have a licensing scheme i can't envision or [1:43:19] imagine right now what kind of our licensing scheme we would be able to create to pretty much regulate [1:43:25] the vastness of of the this game changing tool so are you thinking of an ftc kind of a system an fcc kind [1:43:38] of a system what do the two of you even envision as a potential licensing scheme that would provide the [1:43:45] kind of guard wheels uh that we need to protect our our literally our country from harmful content to [1:43:52] to touch on the first part of what what you said there are things besides uh you know should this [1:43:57] content be generated or not that i think are also important so that image that you mentioned was [1:44:01] generated i think it'd be a great policy to say generated images need to be made clear in all [1:44:06] contexts that they were generated and you know then we still have the image out there but it's we're at [1:44:11] least requiring people to say this was a generated image okay well i mean you don't need an entire [1:44:18] licensing scheme in order to make that where where i think the licensing scheme comes in is uh not with [1:44:26] not for what these models are capable of today because as you pointed out you don't need a new [1:44:31] licensing agency to do that but as we as we head and you know this may take a long time i'm not sure as [1:44:37] we head towards artificial general intelligence and the impact that will have um and the power of that [1:44:42] technology i think we need to treat that as seriously as we treat other very powerful technologies and [1:44:48] that's where i personally think we need such a such a scheme i agree and that is why though by the [1:44:52] time we're talking about agi we're talking about major harms that can occur you through the use of [1:44:59] agi so um professor marcus i mean what kind of a regulatory scheme would you envision and we can't just [1:45:05] come up with something you know that is going to be uh take care of the issues that will arise in the [1:45:12] future especially with agi so what what kind of a scheme would you contemplate well first if i can [1:45:18] can rewind just a moment um i think you really put your finger on the central scientific issue [1:45:23] in terms of the challenges in building artificial intelligence we don't know how to build a system [1:45:28] that understands harm in the full breadth of its meaning so what we do right now is we gather examples and [1:45:33] we say is this like the examples that we have labeled before but that's not broad enough and so [1:45:38] i thought your questioning beautifully outlined um the challenge that ai itself has to face in order to [1:45:44] to really deal with this we want ai itself to understand harm and that may require new technology [1:45:49] so i think that's very important on this second part of your question um the model that i tend to [1:45:55] gravitate towards but i am not an expert here is the fda at least as part of it in terms of you have to [1:46:01] make a safety case and say why the benefits out uh weigh the harms in order to get that license [1:46:07] probably we need elements of multiple agencies i'm not an expert there but i think that the safety case [1:46:12] part of it is incredibly important you have to be able to have external reviewers that are scientifically [1:46:17] qualified look at this and say you have you addressed enough so i'll just give one specific [1:46:21] example auto gpt frightens me that's not something that opening i made but something that [1:46:26] opening i did make call called chat gpt plugins led a few weeks later to some building open source [1:46:33] software called auto gpt and what auto gpt does is it allows systems to access source code access the [1:46:40] internet and so forth and there are a lot of potential let's say cyber security risks there there [1:46:44] should be an external agency that says well we need to be reassured if you're going to release this [1:46:49] product that there aren't going to be cyber security problems or there are ways of addressing it so [1:46:54] professor i am running out of time there there's you know i just want to mention miss mc montgomery [1:47:00] your model is a use model similar to what the eu has come up with but the vast vastness of ai and [1:47:08] the complexities involved i think would require more than looking at the use use of it i think that [1:47:16] based on what i'm hearing today don't you think that that we're probably going to need to do a heck of [1:47:22] a lot more than to focus on what use it is being ai is being used for for example you can ask ai to [1:47:29] come up with a funny joke or something but you can use the same you can ask the same ai tool to generate [1:47:35] something that is like an election fraud kind of a situation so i don't know how you will make a [1:47:40] determination based on where you're going with the use model how to distinguish those kinds of uses of [1:47:47] this tool so i think that if we're going to go toward a licensing kind of a scheme we're going to [1:47:53] need to put a lot of thought into how we're going to come up with an appropriate scheme that is going to [1:47:59] provide the kind of future reference that we need to put in place so i i thank all of you for coming in [1:48:08] and providing further food for thought thank you mr chairman thanks very much senator hirono uh [1:48:16] senator padilla thank you mr chairman i appreciate the flexibility as i've been back and forth to uh [1:48:24] between this committee and uh homeland security committee where there's a hearing going on right [1:48:28] now on the use of ai in government so it's ai day on the hill or at least the senate apparently [1:48:34] uh now for folks watching at home if you never thought about ai until the recent emergence of [1:48:41] generative ai tools the developments in this space may feel like they've just happened all of a sudden [1:48:48] but the fact of the matter is mr chair is that they haven't ai is not new not for government not [1:48:54] for business not for public not for the public in fact the public uses ai all the time and just for [1:49:00] folks to be able to relate want to offer the example of anybody with a smartphone uh many features [1:49:07] on your device leverage ai including suggested replies right when we're text messaging or even to [1:49:14] email autocorrect features including but not limited to spelling in our email and text applications [1:49:22] so i'm frankly excited to explore how we can facilitate positive ai innovation that benefits [1:49:30] society while addressing some of the already known harms and biases that stem from the development [1:49:36] and use of the tools today now with language models becoming increasingly ubiquitous i want to make sure [1:49:44] that there's a focus on ensuring equitable treatment of diverse demographic groups my understanding is that [1:49:51] most research in to evaluating and mitigating fairness harms has been concentrated on the english language [1:50:00] while non-english languages have received comparatively little attention or investment and we've seen [1:50:07] this problem before i'll tell you why i raised this social media companies for example have not [1:50:12] adequately invested in content moderation tools and resources for their non-english uh in in non-english language [1:50:23] and i share this not just out of concern for non-us-based users but so many u.s-based users prefer [1:50:32] a language other than english in their communication so i'm deeply concerned about repeating social media's [1:50:38] failure in ai tools and applications question mr altman and miss montgomery how are open ai ibm ensuring [1:50:47] language and cultural inclusivity that they're in their uh large language models and it's even an area [1:50:57] of focus in the development of your products so bias and equity in technology is a focus of ours and [1:51:06] always has been i think diversity in terms of the development of the tools in terms of their deployment [1:51:11] so having diverse people that are actually training those tools considering the downstream effects [1:51:17] as well we're also very cautious uh very aware of the fact that we can't just be articulating and [1:51:25] calling for these types of things without having the tools and the technology to test for bias and to [1:51:31] apply governance across the life cycle of ai so we were one of the first teams uh and companies to put [1:51:38] toolkits on the market deploy them uh contribute them to open source that will do things like [1:51:44] help to address you know be the technical uh aspects in which we help to address issues like bias [1:51:50] can you speak just for a second specifically to language inclusivity yeah i mean language so we [1:51:56] don't have a consumer platform but we are uh very actively involved with ensuring that the technology [1:52:03] we help to deploy in the large language models that we use in helping our clients to deploy technology [1:52:10] is focused on and available in many languages thank you michelle we think this is really important uh [1:52:17] one example is that we worked with the government of iceland uh which is a language with fewer speakers [1:52:22] than many of the languages that are well represented uh on the internet to ensure that their language [1:52:27] was included in our model uh and we've had many similar conversations and uh i look forward to many [1:52:33] similar partnerships with lower resource languages uh to get them into our models gpt4 is uh unlike [1:52:41] previous models of ours which were good at english and not very good at other languages uh now pretty good [1:52:46] at a large number of languages you can go pretty far down the list ranked by number of speakers and still [1:52:52] get good performance but for these very small languages we're excited about custom partnerships [1:52:57] to include that language uh into our model run and the part of the question you asked about values [1:53:02] and making sure that cultures are included we're equally focused on that uh excited to work with [1:53:07] people who have particular data sets uh and to work to collect a representative set of values from [1:53:12] around the world to draw these wide bounds of what the system can do i also appreciate what you said [1:53:17] about the the benefits of these systems and wanting to make sure we get those to as wide of a group as [1:53:22] possible uh i think this will these systems will have lots of positive impact on a lot of people [1:53:27] but in particular uh underrepresented historically underrepresented groups in technology uh people [1:53:33] who have not had as much access technology around the world this technology seems like it can be a big [1:53:38] lift up okay and uh i know my my uh question was specific to language inclusivity but i'm glad there's [1:53:47] agreement on the broader uh commitment to uh diversity and inclusion and i'll just give a couple more [1:53:53] reasons why i think it's so critical you know the largest actors in this space can afford the massive [1:54:00] amount of data the computing power and they have the financial resources necessary to develop complex [1:54:06] ai systems but in this space we haven't seen from a workforce standpoint the racial and gender diversity [1:54:15] reflective of the united states of america and we risk if we're not thoughtful about it uh contributing [1:54:21] to the development of tools and approaches that only exacerbate uh the bias and inequities that exist [1:54:29] in our society so a lot of follow-up work to do there in my time remaining i do want to ask one more [1:54:35] question uh this committee and the public are right to pay attention to the emergence of generative ai [1:54:40] now this technology has a different opportunity and risk profile than other ai tools and these [1:54:47] applications have felt very tangible for the public due to the nature of the user interface and the [1:54:53] outputs that they produce but i don't think we should lose sight of the broader ai ecosystem as you [1:54:58] consider ai's broader impact on society as well as the design of appropriate safeguards so miss montgomery [1:55:05] in your testimony uh as you noted ai is not you can you highlight some of the different applications [1:55:13] that the public and policy makers should also keep in mind as we consider possible regulations yeah i mean [1:55:21] i think the generative ai systems that are available today are creating new issues that ness that need [1:55:28] to be studied new issues around the potential to generate content that could be extremely misleading [1:55:34] deceptive and the like so those issues absolutely need to be studied but we shouldn't also ignore [1:55:40] the fact that ai is a tool it's been around for a long time it has capabilities beyond just generative [1:55:46] capabilities and again that's why i think going back to this approach where we're regulating ai where [1:55:51] it's touching people and society is a really important way to address it thank you thank you mr chair [1:55:59] thanks senator pia uh senator booker is next but i think he's going to defer to senator ossoff because [1:56:06] senator ossoff is a very big deal i don't know if you know i have a meeting at noon and i'm grateful to [1:56:13] you senator booker for uh yielding your time uh you are as always brilliant and handsome and uh thank you to [1:56:21] the panelists for joining us thank you to the subcommittee leadership for opening this up to all committee [1:56:27] members if we're going to contemplate a regulatory framework uh we're going to have to define what [1:56:33] it is that we're regulating so uh you know mr albin any such law will have to include a section that [1:56:39] defines the scope of regulated activities technologies tools products just take a stab at it yeah thanks [1:56:46] for asking senator ossoff i think i think it's super important i think there are very different levels [1:56:50] here um and i think it's important that any any new approach any new law does not stop the innovation [1:56:57] from happening with smaller companies open source models researchers that are doing work at a smaller [1:57:02] scale uh that's a wonderful part of this ecosystem and of america we don't want to slow that down [1:57:08] there still may need to be some rules there um but i think we could draw a line at systems that need [1:57:15] to be licensed in a very intense way the easiest way to do it i'm not sure if it's the best but the easiest [1:57:21] would be to talk about the amount of compute that goes into such a model so we could you know we could [1:57:26] define a threshold of compute and it'll have to go it'll have to change it could go up or down i could [1:57:31] down as we discover more efficient algorithms that says above this amount of compute you are in this [1:57:37] regime um what i would prefer uh it's hard to do but i think more accurate is to define some capability [1:57:44] thresholds and say a model that can do things x y and z up to you all to decide uh that's now in this [1:57:50] licensing regime but models that are less capable you know we don't want to stop our open source [1:57:54] community we don't want to stop individual researchers we don't stop new startups can [1:57:58] proceed you know with a different framework thank you as concisely as you can please state which [1:58:02] capabilities you'd propose we consider for the purposes of this definition uh i would love rather [1:58:09] than to do that off the cuff to follow up with your office with like a well perhaps opine things opine [1:58:14] understanding that you're just responding uh and you're not making law all right in the spirit of [1:58:19] just opining um i think a model that can uh persuade manipulate influence a person's behavior [1:58:26] or a person's beliefs that would be a good threshold um i think a model that could help [1:58:30] create novel biological agents would be a great threshold things like that i want to talk about [1:58:36] the predictive capabilities of the technology and we're gonna have to think about a lot of very [1:58:41] complicated constitutional questions that arise from it uh with with massive data sets [1:58:47] uh the integrity and accuracy with which such technology can predict future human behavior is [1:58:56] potentially uh pretty significant at the individual level correct i think we don't know the answer to [1:59:02] that for sure but let's say it can at least have some impact there okay so we may be confronted by [1:59:06] situations where for example a law enforcement agency uh deploying such technology seeks some kind of [1:59:13] judicial consent to execute a search or to take some other police action on the basis of a modeled [1:59:19] prediction about some individual's behavior but that's very different from the kind of evidentiary [1:59:27] predicate that normally police would take to a judge in order to get a warrant talk me through how you [1:59:32] think about that issue yeah i think it's very important that we continue to understand that these [1:59:39] are tools that humans use to make human judgments and that we don't take away human judgment i don't think [1:59:44] that people should be prosecuted based off of the output of an ai system for example we have no [1:59:50] uh national privacy law uh the europe has has rolled one out to uh mixed reviews do you think we need [2:00:01] one i think it'd be good and what would be the qualities or purposes of such a law that you think [2:00:07] would make the most sense based on your experience again this is very far out of my area of expertise i [2:00:11] think there's many many people that could that are privacy experts that could weigh on what a law needs [2:00:16] i'd still like i'd still like you to weigh in um i mean i think a minimum is that users should be [2:00:22] able to to sort of opt out from having their data used by companies like ours or the social media [2:00:27] companies it should be easy to delete your data i think those are it should but the thing that i think [2:00:33] is important from my perspective running an ai company is that if you don't want your data used [2:00:38] for training these systems uh you have the right to do that so let's think about how that would be [2:00:42] practically implemented i mean uh as i understand it uh your tool and certainly similar tools one of [2:00:49] the inputs will be um scraping for lack of a better word data off of the open web right as a low cost way [2:00:59] of of gathering information and there's a vast amount of uh information out there about all of us [2:01:07] how would such a restriction on the access or use or analysis of such data be practically implemented [2:01:14] so uh i was speaking about something a little bit different which is the data that someone [2:01:17] generates the questions they ask our system things that they input their training on that data that's [2:01:22] on the public web that's accessible even if we don't train on that the models can certainly link out to [2:01:26] it uh so that was not what i was referring to i think that you know there's ways to have your data or [2:01:32] there should be more ways to have your data taken down from the public web but certainly models with [2:01:36] web browsing capabilities will be able to search the web and link out to it when you think about [2:01:42] implementing a safety or a regulatory regime to constrain such software and to mitigate some risk [2:01:52] is your view that the federal government would make laws such that certain capabilities or [2:02:00] functionalities themselves are forbidden in potential in other words one cannot deploy or execute code [2:02:09] capable of x yes or is it the act itself x only when actually executed well i think both i'm a [2:02:17] believer in defense in depth i think that there should be limits on what a deployed model is capable [2:02:22] of and then what it actually does too how are you thinking about how kids use your product we well you [2:02:29] have to be i mean you have to be 18 or up or have your parents permission at 13 and up to use a product [2:02:34] but we understand that people get around those safeguards all the time and so what we try to do [2:02:38] is just design a safe product and there are decisions that we make that we would allow if [2:02:42] we knew only adults were using it that we just don't allow in the product because we know children will [2:02:46] use it some way or other too in particular given how much these systems are being used in education [2:02:53] we like want to be aware that that's happening i think what and senator blumenthal has done extensive [2:03:00] work investigating this what we've seen repeatedly is that companies whose revenues depend upon volume [2:03:07] of use screen time intensity of use design these systems in order to maximize the engagement of all [2:03:16] users including children with with perverse results in many cases and what i would humbly advise you [2:03:22] is that you get way ahead of this issue the safety for children of your product or i think you're [2:03:30] going to find that senator blumenthal senator holly others on the subcommittee and i um are will look [2:03:37] very harshly on the deployment of technology that harms children we couldn't agree more i think we're [2:03:42] out of time but i'm happy to talk about that if i can respond go ahead well it's up to the chairman [2:03:47] okay um i first of all i think we try to design systems uh that do not maximize for engagement in fact [2:03:55] we're so short on gpus uh the less people use our products the better but we're not an advertising [2:04:00] based model we're not trying to get people to use it more and more um and i think that's that's a [2:04:06] different shape than ad supported social media second uh these systems do have the capability to [2:04:14] to influence in obvious and in very nuanced ways um and i think that's particularly important for the [2:04:20] safety of children but that will that will impact all of us one of the things that we'll do ourselves [2:04:25] uh regulation or not but i think a regulatory approach would be good for also [2:04:29] is requirements about how the values of these systems are set and how these systems respond [2:04:36] to questions that can cause influence um so we'd love to partner with you couldn't agree more on the [2:04:40] importance thank you mr chairman for the record i just want to say that the senator from georgia is [2:04:46] also very handsome and brilliant too but i will uh allow that comment to stand with without objection [2:04:55] without objection okay um mr chairman and ranking member you are now recognized thank you very much [2:05:02] thank you it's nice that we finally got down to the ball guys down here at the end um i just want to [2:05:07] thank you both this has been one of the best hearings i've had this congress and uh just a testimony to you [2:05:11] two and seeing uh the challenges and the opportunities that ai presents so i appreciate you both i want to just [2:05:17] jump in i think very broadly and then i'll get a little more narrow uh sam you said very broadly technology [2:05:23] has been moving like this and we are a lot of people have been talking about regulation [2:05:29] and so i use the example of the automobile what an extraordinary uh piece of technology i mean new [2:05:35] york city did not know what to do with horse manure they were having crises forming commissions and the [2:05:40] automobile comes along ends that problem but at the same time we have tens of thousands of people [2:05:45] dying on highways every day we have emissions crises and the like there are multiple federal agencies [2:05:51] multiple federal agencies that were created uh or are specifically focused on regulating cars um and [2:05:59] and so this idea that this equally transforming technology is coming and for congress to do nothing which is [2:06:07] not what anybody here is is uh calling for little or nothing is is obviously unacceptable uh i really [2:06:14] appreciate uh senator welsh and i who've been going back and forth uh during this hearing and him and bennett [2:06:19] have a bill talking about trying to regulate in this space not doing so for uh social media uh has [2:06:28] been i think very destructive and allowed a lot of things uh to go on that are really causing a lot of [2:06:33] harm and so the question is is what kind of regulation you all have spoken that to a lot of my colleagues [2:06:39] um and i and i want to say miss montgomery and i have to give full disclosure i'm the child of two ibm parents [2:06:45] um uh but i i you know you talked about uh defining the highest risk uses we don't know all of them we [2:06:54] really don't we can't see where this is going uh regulating at the point of risk and you you sort of [2:07:02] called not for an agency and i think when you when somebody else asks you to specify because you don't [2:07:06] want to slow things down we should build on what we have in place but you can envision that we can try to [2:07:12] work on two different ways that ultimately a specific like we have in uh in cars epa nitsa the [2:07:21] federal motor car carrier safety administration all of these things you can imagine something specific [2:07:27] that is uh as mr marcus points out uh a nimble agency that could do monitoring other things you [2:07:33] can imagine the need for something like that correct oh absolutely yeah and and so uh just for the record [2:07:39] then in addition to trying to regulate with what we have now you would encourage congress and my [2:07:45] colleague senator welsh to move forward in trying to figure out the right tailored agency to deal [2:07:50] with what we know and perhaps things that might come up in the future i would encourage congress to [2:07:56] make sure it understands the technology has the skills and resources in place to impose regulatory [2:08:02] requirements on the uses of the technology and to understand emerging risks as well so yes yeah mr marcus [2:08:08] there's no way no way to put this genie in the bottle globally this is it's exploding i appreciate [2:08:14] your thoughts and i shared some with my staff about your ideas of what the international context is but [2:08:20] there's there's no way to stop this moving forward so with that understanding just building on what what [2:08:26] ms montgomery said what kind of encouragement do you have as specifically as possible to forming an agency [2:08:32] to using current rules and regulations can you just put some clarity on what you've already stated [2:08:36] let me just insert there are more genies yet to come from more bottles some genies are already out [2:08:42] but we don't have machines that can really for example self-improve themselves we don't really have [2:08:47] machines that have self-awareness and we might not ever want to go there so there are other genies to be [2:08:52] concerned about on to the main part of your question i i think that we need to have some international [2:08:58] meetings very quickly with people who have expertise in how you grow agencies in the history of growing [2:09:04] agencies we need to do that in the federal level we need to do that in the international level [2:09:08] i'll just emphasize one thing i haven't as much as i would like to which is that i think science has [2:09:13] to be a really important part of it i'll give an example we've talked about misinformation we don't [2:09:17] really have the tools right now to detect and label misinformation with nutrition labels that we would [2:09:22] like to we have to build new technologies for that we don't really have tools yet to detect a wide [2:09:28] uptick in cybercrime probably we probably need new tools there we need science to probably help us [2:09:34] to figure out what we need to build and also what it is that we need to have transparency around understood [2:09:40] understood sam just going to you uh for the little bit of time i have left real quick um first of all [2:09:46] you're a bit of a unicorn when i sat down with you first uh could you explain why non-profit in other [2:09:51] words you're you're you're you're not looking this and you've even capped the vc people just really [2:09:56] quickly i want folks to understand that we started as a non-profit uh really focused on [2:10:03] how this technology was going to be built at the time it was very outside the overton window that [2:10:08] something like agi was even possible that's that shifted a lot um we didn't know at the time how [2:10:12] important scale was going to be but we did know that we wanted to build this um with humanity's best [2:10:19] interest at heart and a belief that this technology could if it goes the way we we want if we can do some [2:10:24] of those things professor marcus mentioned uh really deeply transformed the world and we wanted [2:10:30] to be as much of a force for getting to a positive i'm going to interrupt you i think that's all good [2:10:35] i hope more of that gets on the record the second part of my question as well um i found it fascinating [2:10:41] uh but are you ever going to for a revenue model for return on your investors are you ever going to do [2:10:46] ads uh or something like that i wouldn't say never i don't think like i think there may be people [2:10:53] that we want to offer services to and there's no other model that works but i really like having [2:10:57] a subscription-based model uh we have api developers pay us and we have chat gpt okay can i just jump [2:11:05] real quickly one of my biggest concerns about this space is what i've already seen in the space of uh [2:11:11] web 2 web 3 is this massive corporate concentration it is really terrifying to see how few companies now [2:11:18] control and affect the lives of so many of us and these companies are getting bigger and more powerful [2:11:23] and i see you know open ai backed by microsoft uh anthropic is backed by google uh google has its [2:11:28] own in-house product we know bar so i'm really worried about that and i'm wondering uh if sam you [2:11:35] can give me a quick uh acknowledgement are you worried about the corporate concentration in this [2:11:39] space and what effect it might have uh um uh and the associated risks perhaps with market concentration [2:11:45] in ai and mr market marcus can you answer that as well i think there will be many people that develop [2:11:51] models uh what's happening on the open source community is amazing but there will be a [2:11:55] relatively small number of providers that can make models at the at the true um i think there is [2:12:03] benefits and danger to that because we're talking about all the dangers with ai the fewer of us that [2:12:07] you really have to keep a careful eye on on the absolute like bleeding edge of capabilities there's [2:12:12] benefits there um but then i think there needs to be enough and there will because there's so much value [2:12:17] that consumers have choice that we have different ideas mr marcus real quick there there is a real [2:12:22] risk of a kind of techno technocracy combined with oligarchy where a small number of companies [2:12:27] influence people's beliefs through the nature of these systems again i put something in the wall [2:12:32] in the record about the wall street journal about how these systems can subtly shape our beliefs and [2:12:37] that has enormous influence um on how we live our lives and having a small number of players do that [2:12:43] with data that we don't even know about that scares me sam i'm sorry one more thing i wanted to add uh [2:12:48] one thing that i think is very important is that this what these systems get aligned to whose values [2:12:53] what those bounds are that that is somehow set by society as a whole by governments as a whole and [2:12:59] so creating that data set the aligned that our alignment data set it could be you know an ai [2:13:04] constitution whatever it is that that has got to come very broadly from society thank you very much [2:13:10] mr chairman my time's expired and i guess the best for last thank you senator booker senator well uh [2:13:17] first of all i want to thank uh you senator blumenthal and you senator holly this has been a tremendous [2:13:22] hearing uh senators are noted for their short attention spans but i've sat through this entire [2:13:29] hearing and enjoyed every minute of it you have one of our longer attention spans in the united states [2:13:36] to your great credit well we've had good witnesses and it's an incredibly important issue [2:13:40] and here's just i don't all the questions i have have been asked really but here's a kind of a [2:13:47] takeaway and what i think is the major question that we're going to have to answer as a congress [2:13:55] number one you're here because ai is this extraordinary new technology that everyone says can be [2:14:04] transformative as much as the printing press number two it's really unknown what's going to happen [2:14:11] but there's a big fear you've expressed to all of you about what bad actors can do and will do [2:14:18] if there's no uh rules of the road uh number three as a member who served in the house and now in the [2:14:28] senate i've come to the conclusion that it's impossible for congress to keep up with the speed of [2:14:34] technology and there have been concerns expressed uh in so about social media and now about a ai that [2:14:43] relate to fundamental privacy rights bias rights uh intellectual property uh distant the spread of [2:14:51] disinformation which in many ways for me is the biggest threat because that goes to the core of our [2:14:55] capacity uh for self-governing uh there's the economic transformation which uh can be profound uh there's [2:15:04] safety concerns and and i've come to the conclusion that we absolutely have to have an agency uh what [2:15:12] its scope of engagement is it has to be defined by us uh but i believe that unless we have an agency [2:15:24] that is going to address these questions from social media and in ai uh we really don't have much [2:15:30] of a defense against the bad stuff and the bad stuff will come so uh last year i introduced in the [2:15:38] house side and senator bennett didn't senate side was the end of the year digital commission act and [2:15:43] we're going to be reintroducing that this year and the two things that i want to ask one you've [2:15:50] somewhat answered because i think two of the three of you have said you think we do need an independent [2:15:55] commission you know and congress established an independent commission when railroads were running rampant [2:16:01] over the interest of farmers when wall street had no rules of the road and we had the s sec and i [2:16:07] think we're at that point now but the what the commission does would have to be defined and [2:16:15] circumscribed but also there's always a question about the use of regulatory authority and the [2:16:24] recognition that it can be used for good jd vance actually mentioned that when we were considering his [2:16:29] and senator brown's bill about railroads in that event in east palestine regulation for the public [2:16:37] health but there's also a legitimate concern about regulation getting in the way of things being too [2:16:42] cumbersome and being a negative influence so a two of the three of you have said you think we do need [2:16:51] need an agency uh what are some of the perils of an agency that we would have to be mindful of [2:16:58] in order to make certain that its goals of protecting many of those interests i just mentioned [2:17:04] privacy bias intellectual property disinformation would be the winners and not the losers and i'll [2:17:10] start with you mr altman thank you senator uh one i think america has got to continue to lead [2:17:18] uh this happened in america i'm very proud that it happened in america by the way i i think that's [2:17:23] right and that's why i'd be much more confident if we had our agency as opposed to got involved in [2:17:31] international discussions ultimately you want the rules of the road but i think if we lead and get [2:17:35] rules of the road that work for us that is probably a more effective way to proceed i i i personally [2:17:41] believe there's a way to do both and i think it is important to have the global view on this because [2:17:46] this technology will impact americans and all of us wherever it's developed but uh i think we want [2:17:53] america to lead we want we want so get to the perils issue though because i know well that's one i mean [2:17:58] that is apparel which is you slow down american industry in such a way that china or somebody else [2:18:03] makes faster progress uh a second and i think this can happen with like the regulatory pressure should [2:18:10] be on us it should be on google it should be on the other small set of people in the lead the most [2:18:14] we don't want to slow down smaller startups we don't want to slow down uh open source efforts we [2:18:18] still need them to comply with things they can still you can still cause great harm with a smaller [2:18:22] model but leaving the room and the space for new ideas and new companies uh and independent [2:18:31] researchers to do their work and not putting a regulatory burden that say a company like us could [2:18:36] handle but a smaller one couldn't um i think that's another peril and it's clearly a way that regulation has [2:18:41] gone mr marcus or professor marcus the the other obvious peril is regulatory capture if we make it as [2:18:49] appear as if we are doing something but it's more like green washing and nothing really happens we just [2:18:55] keep out the little players because we put so much burden that only the big players can do it so there are [2:19:01] also those kinds of perils i fully agree with everything that that mr altman said and i would add that to [2:19:05] the list miss montgomery one of the things i would add to the list is the risk of not holding companies [2:19:12] accountable for the harms that they're causing today right so we talk about misinformation in [2:19:18] electoral systems so no agency or no agency we need to hold companies responsible today and accountable [2:19:26] for the ai that they're deploying that disseminates misinformation on things like elections and where [2:19:33] the where the risk you know regulatory agency would do a lot of the things that senator graham was talking [2:19:37] about you know you don't build a nuclear reactor without getting a license you don't build an ai system [2:19:42] without getting a license that gets tested independently i think it's a great analogy we [2:19:47] we need both pre-deployment pre-deployment and post-deployment okay thank you all very much i yield back mr [2:19:53] chairman thanks thanks senator wells let me ask a few more questions you've all been very very patient [2:20:00] and the turnout today which is beyond our subcommittee i think reflects both your value in what you're [2:20:09] contributing as well as the interest in this topic uh there are a number of subjects that we haven't [2:20:15] covered at all but one was uh just alluded to by professor marcus which is the monopolization danger [2:20:29] the dominance of markets that excludes new competition and thereby inhibits or prevents innovation and [2:20:40] invention which we have seen in social media as well as some of the old industries airlines automobiles [2:20:50] and others where consolidation has narrowed competition and so uh i think we need to to focus on [2:21:00] kind of an old area of antitrust which dates more than a century uh still inadequate to deal with the [2:21:07] the challenges we have right now in our economy and certainly we need to be mindful of the way that [2:21:14] rules can enable the big guys to get bigger and exclude innovation and competition and responsible [2:21:21] good guys such as are represented in this industry right now we haven't dealt with national security [2:21:31] there are huge implications for national security i will tell you as a member of the armed services committee [2:21:35] uh classified briefings on this issue have abounded and the threats that are posed by some of our [2:21:45] adversaries china has been mentioned here but the sources of threats to this nation in this space are very [2:21:54] real and urgent uh we're not going to deal with them today but we do need to deal with them and we will [2:22:01] hopefully in this committee and then uh on the issue of a new agency you know i've been doing this stuff [2:22:08] for a while i was attorney general of connecticut for 20 years i was a federal prosecutor the u.s attorney [2:22:15] most of my career has been in enforcement and i will tell you something you can create 10 new agencies [2:22:20] but if you don't give them the resources and i'm talking not just about dollars i'm talking about [2:22:27] scientific expertise you guys will run circles around them and it isn't just the the models or the [2:22:37] generative ai that will run models around run circles around them but it is the scientists in your [2:22:46] companies uh for every success story in government regulation you can think of five failures that's true [2:22:53] of the fda it's true of the iaea it's true of the sec it's true of the whole alphabet list of government [2:23:02] agencies and i hope our experience here will be different but the pandora's box requires more than [2:23:10] just the words or the concepts licensing new agency there's some real hard decision making as ms montgomery [2:23:22] has alluded to about how to frame the rules to fit the risks first do no harm make it effective make it [2:23:34] enforceable make it real i think we need to grapple with the the hard questions here that you know [2:23:43] frankly this initial hearing i think has raised very successfully but not answered and i i thank our [2:23:51] colleagues who have participated and and made these uh very creative suggestions i'm very interested in [2:24:00] uh enforcement i you know literally 15 years ago i think advocated abolishing section 230 what's [2:24:07] old is new again you know now people are talking about abolishing section 230 back then it was [2:24:15] considered completely unrealistic but uh enforcement really does matter i want to ask uh mr altman because [2:24:24] of the the privacy issue and you've suggested uh that you have an interest in protecting the privacy of [2:24:32] the data that may come to you or be available how do you what specific steps you take uh to protect [2:24:40] privacy uh one is that we don't train on any data submitted to our api so if you're a business customer [2:24:48] of ours and submit data we don't train on it at all we do retain it for 30 days solely for the purpose of [2:24:54] trust and safety enforcement um but that's different than training on it if you use chat gpt uh you can opt [2:25:00] out of us training on your data um you can also delete your conversation history or your whole account [2:25:07] um ms montgomery i know you don't deal directly with consumers but do you take steps to protect privacy [2:25:13] as well absolutely and we even filter our large language models for content that includes personal [2:25:21] information that may have been pulled from public data sets as well so we apply additional level of [2:25:25] filtering um professor marcus you made reference to self-awareness self-learning uh already we're [2:25:36] talking about the potential for jail breaks um how soon do you think that new kind of generative [2:25:46] ai will be usable will be practical new ai that is self-aware and so forth where yes i mean i have no [2:25:57] idea on that one i think we don't really understand what self-awareness is and so it's hard [2:26:01] to put a date on it um in terms of self-improvement there's some modest self-improvement in current [2:26:06] systems but one could imagine a lot more and that could happen in two years it could happen in 20 [2:26:11] years um the basic paradigms that haven't been invented yet some of them we might want to discourage [2:26:18] but it's a bit hard to put timelines on them and just going back to enforcement for one second [2:26:22] one thing that is absolutely paramount i think is far greater transparency about what the models are and [2:26:28] what the data are that doesn't necessarily mean everybody in the general public has to know [2:26:32] exactly what's in one of these systems but i think it means that there needs to be some enforcement arm [2:26:37] that can look at these systems can look at the data can perform tests and so forth um let me ask you [2:26:44] uh all of you uh i think there has been a reference to um elections and banning um outputs involving [2:26:58] elections are there other areas where you think what are the what are the other high risk or highest [2:27:03] risk areas where you would either ban or establish especially strict rules miss montgomery the space [2:27:15] around misinformation i think is hugely important one um and coming back to the points of transparency [2:27:22] you know knowing what content was generated by ai is going to be a really critical area that we need to [2:27:29] address any others i think medical misinformation is something to really worry about we have systems [2:27:36] that hallucinate things they're going to hallucinate medical advice some of the advice they'll give is [2:27:40] good some of it's bad we need really tight regulation around that same with psychiatric um advice people [2:27:46] using these things as as kind of ersatz therapists i think we need to be very concerned about that i think [2:27:52] we need to be concerned about internet access for these tools when they can start making requests both of [2:27:58] people and and internet things it's probably okay if they just do search but as they do more intrusive [2:28:04] things on the internet like do we want them to be able to order equipment or order um chemistry and [2:28:09] so forth so as they as we empower these systems more by giving them internet access i think we need to [2:28:15] be concerned about that and then we've hardly talked at all about long-term risks sam alluded to it briefly [2:28:21] i don't think that's where we are right now but as we start to approach machines that have a larger [2:28:26] footprint on the world beyond just having a conversation we need to worry about that and [2:28:30] think about how we're going to regulate that and monitor it and so forth uh in a sense we've been [2:28:36] talking about bad guys or um certain bad actors manipulating ai to do harm manipulating people and [2:28:48] manipulating people but also generative ai can manipulate the manipulators it it can i mean there's there's [2:28:56] many layers of manipulation that are possible and i think we don't yet really understand the consequences [2:29:02] dan dennett just sent me a manuscript last night that will be in the atlantic in a few days on what [2:29:07] he calls counterfeit people um it's a wonderful metaphor these systems are almost like counterfeit [2:29:13] people and we don't really honestly understand what the consequence of that is they're not perfectly human [2:29:20] like yet but they're good enough to fool a lot of the people a lot of the time [2:29:23] and that introduces lots of problems for example cybercrime and how people might try to manipulate [2:29:28] markets and so forth so it's a serious concern in my opening i suggested three principles transparency [2:29:40] accountability and limits on use would you agree that those are a good starting point in my company [2:29:50] 100 and as you also mentioned industry shouldn't wait for congress that's what we're doing here at ibm [2:29:56] there's no reason that absolutely wait for congress yeah professor marcus i think those three would be [2:30:02] a great start i mean there are there are things like the white house bill of rights for example that [2:30:06] show i think a large consensus the unesco guidelines and so forth so a large consensus around what it is [2:30:12] we need and the real question is definitely now how are we going to put some teeth in it try to make [2:30:17] these things actually enforce so for example we don't have transparency yet we all know we want it but [2:30:22] we're not doing enough to enforce it mr altman i certainly agree that those are important points [2:30:28] i would add that and professor marcus touched on this i would add that as we we spend most of the [2:30:34] time today on current risks and i think that's appropriate and i'm very glad we have done it [2:30:39] as these systems do become more capable and i'm not sure how far away that is but maybe not not super [2:30:45] far i think it's important that we also spend time talking about how we're going to confront those [2:30:49] challenges having talked to you privately you know how much i care i agree that you care deeply and [2:30:56] intensely but also that prospect of increased danger or risk resulting from even more complex and capable [2:31:07] ai mechanisms certainly may be closer than a lot of people appreciate let me just add for the record [2:31:15] that i'm sitting next to sam closer than i've ever sat to him except once before in my life [2:31:21] and that his sincerity in talking about those fears is very apparent um physically in a way that just [2:31:27] doesn't communicate on the television screen thank you indicates from here thank you senator holly [2:31:33] thank you again mr chairman for a great hearing thanks to the witnesses so i've been keeping a [2:31:37] little list here of the potential downsides or harms risks of generative ai even in its current form [2:31:44] let's just run through a loss of jobs and this isn't speculative i think your company miss montgomery [2:31:49] has announced that it's it's potentially laying off 7 800 people third of your non-consumer facing [2:31:54] workforce for because of ai so loss of jobs invasion of privacy personal privacy on a scale we've never [2:32:01] before seen manipulation of personal behavior manipulation of personal opinions and potentially [2:32:07] the degradation of free elections in america did i miss anything i mean this is this is quite a list [2:32:13] i noticed that an eclectic group of about a thousand technology and ai leaders everybody from [2:32:20] andrew yang to elon musk recently called for a six-month moratorium on any further ai development [2:32:27] were they right do you join those calls are they right to do that should we should we pause for six [2:32:32] months your characterization is not quite correct um i actually signed that letter about 27 000 people signed [2:32:38] it um it did not call for a ban on all ai research it only called in nor on all ai but only on a very [2:32:47] specific thing which would be systems like gpt5 um every other piece of research that's ever been done [2:32:54] it was actually supportive or neutral about it specifically called for more ai specifically called [2:33:00] for more research on trustworthy and safe ai so you think just so you think that we should take [2:33:06] a moratorium a six-month moratorium or more on anything beyond chat gpt4 i took the letter what [2:33:12] is the famous phrase uh spiritually not literally what was the famous phrase um well i'm asking for [2:33:18] your opinion now though so do my my opinion is that the moratorium that we should focus on is actually [2:33:25] deployment until we have good safety cases um i don't know that we need to pause that particular project [2:33:31] but i do think its emphasis on focusing more on ai safety on trustworthy reliable ai is exactly right [2:33:39] deployment means not making it available to the public yeah so my concern is about things that are [2:33:45] deployed at a scale of let's say 100 million people without any external review i think that we should [2:33:50] think very carefully about doing that what about you mr allman do you agree with that would you would [2:33:55] you pause any further development for six months or longer uh so first of all we after we finished [2:34:00] training gpt4 we waited more than six months to deploy it um we are not currently training what [2:34:05] will be gpt5 we don't have plans to do it in the next six months but i think the frame of the letter [2:34:10] is wrong what matters is audits red teaming safety standards that a model needs to pass before training [2:34:18] if we pause for six months then i'm not what we sure sure what we do then do we pause for another six do [2:34:23] we kind of come up with some rules then the standards that we have developed and that we've used for [2:34:29] gpt4 deployment uh we want to build on those but we think that's the right direction [2:34:33] uh not a calendar clock pause there may be times i expect there will be times when we find something [2:34:40] that we don't understand and we really do need to take a pause but we don't see that yet never mind all [2:34:46] the benefits what would you you don't see what yet you're comfortable with all of the potential [2:34:51] ramifications from the current existing technology i'm sorry we don't see the reasons to not train anyone [2:34:55] for deploying as i mentioned i think there's all sorts of risky behavior and there's limits we put we [2:35:00] have to pull things back sometimes add new ones i mean we don't see something that would stop us from [2:35:04] training the next model uh where we'd be so worried that we'd create something dangerous even in that [2:35:08] process let alone the deployment but that may happen what about you miss montgomery i think we need to [2:35:14] use the time to prioritize ethics and responsible technology as opposed to posing development well [2:35:21] wouldn't a pause in development help the development of protocols for safety standards and ethics i'm not [2:35:27] sure how practical it is to pause but we absolutely should be prioritizing safety protocols okay the [2:35:33] the point about practicality leads me to this i'm interested in this talk about an agency and you know [2:35:39] maybe that would work although having seen how agencies work in this government they usually get [2:35:45] captured by the interests that they're supposed to regulate they usually get controlled by the people [2:35:51] who they're supposed to be watching i mean that's just been our history for a hundred years maybe this [2:35:55] agency would be different i have a little different idea why don't we just let people sue you why don't [2:36:00] we just make you liable in court we can do that we know how to do that we can pass a statute we can [2:36:05] create a federal right of action that will allow private individuals who are harmed by this technology [2:36:10] to get into court and to bring evidence into court and it can be anybody i mean you want to talk about [2:36:16] crowdsourcing we'll just open the courthouse doors we'll define a broad right of action private right [2:36:22] of action private citizens be class actions we'll just open it up we'll allow people to go into [2:36:27] court we'll allow them present evidence they say that they were harmed by they were given medical [2:36:32] misinformation they were given election misinformation whatever why not do that mr altman i mean please [2:36:40] forgive my ignorance can't can't people sue us well you're not protection by protected by section 230 [2:36:46] but there's not currently a i don't think a federal right of action private right of action that says [2:36:52] that if you are harmed by generative ai technology we will guarantee you the ability to get into court [2:36:57] well i think there's like a lot of other laws where if you know technology harms you uh there's [2:37:03] standards that we could be sued under unless i'm really misunderstanding how things work [2:37:07] if the question is are more are clearer laws about the specifics of this technology and consumer [2:37:14] protections a good thing i would say definitely yes the laws that we have today were designed long [2:37:20] before we had artificial intelligence and i do not think they give us enough coverage uh the plan [2:37:25] that you propose i think is a hypothetical would certainly make a lot of lawyers wealthy but i think [2:37:29] it would be too slow to affect a lot of the things that we care about and there are gaps in the law for [2:37:34] example we don't really wait you think it'd be slower than congress yes i do in some ways really [2:37:41] well litigation can take a decade or more oh but the threat of litigation is a powerful tool i mean how [2:37:47] would ibm like to be sued for a hundred billion dollars in no way asking to take litigation off [2:37:51] the table among the tools but i think for example if i can uh continue um we there are areas like [2:37:58] copyright where we don't really have laws we don't really have a way of thinking about wholesale [2:38:03] misinformation as opposed to individual pieces of it where say a foreign actor might make billions of [2:38:08] pieces of misinformation or a local actor we have some laws around market manipulation we could apply [2:38:13] but we get in a lot of situations where we don't really know which laws apply there would be loopholes [2:38:18] the system is really not um thought through in fact we don't even know that 230 does or does not [2:38:24] apply here as far as i know i think that that's something a lot of people speculated about this [2:38:28] afternoon but it's not solid we could fix that well the question is how oh easy you you just you just [2:38:37] it it would be easy for us to say that section 230 doesn't apply to generative ai i'm just going to [2:38:41] give you the last word i'll use this term just on the point of you suggested miss montgomery a duty of [2:38:46] care which i think fits the idea of a private right of action no that's exactly right and also ai is not [2:38:53] a shield right so so if a company discriminates in granting credit for example or in the hiring process [2:39:02] the by virtue of the fact that they relied too significantly on an ai tool they're responsible [2:39:08] for that today regardless of whether they used a tool or a human to make that decision i'm going [2:39:14] to turn to senator booker for some final questions but i just uh want to make a quick point here on the [2:39:22] on the issue of the moratorium i think we need to be careful the world won't wait the rest of [2:39:29] the global scientific community isn't going to pause we have adversaries that are moving ahead [2:39:35] and sticking our head in the sand is not the answer safeguards and protections yes but a flat [2:39:46] stop sign sticking our head in the sand i would be very very worried without militating for any sort [2:39:55] of pause i would just again emphasize there is a difference between research which surely we need [2:40:00] to do to keep pace with our foreign rivals and deployment at really massive scale you know you could [2:40:06] deploy things at the scale of a million people or 10 million people but not a hundred million people [2:40:10] or billion people and if there are risks you might find them out sooner and be able um to close the [2:40:16] barn doors before the horses leave rather than after senator booker yeah i i just there will be no pause [2:40:22] i mean there's no enforcement body to force a paw it's just not not going to happen it's nice to call [2:40:27] for it uh for any just reasons or whatsoever but i'm forgive me for sounding skeptical nobody's pausing [2:40:33] this thing is i would agree i would agree i mean i don't think it's a realistic thing in the world [2:40:38] the reason i personally signed the letter was to call attention to how serious the problems were [2:40:43] and to emphasize spending more of our efforts on trustworthy and safe ai rather than just making [2:40:48] a bigger version of something we already know to be unreliable yeah so i'm i'm a futurist i love [2:40:55] exciting about the future and i guess there's a famous question uh if you couldn't control for your race [2:41:00] your gender where you would land on the planet earth or what time in humanity would you want to [2:41:04] be born everyone would say right now it's the it's still the best time to be alive because of technology [2:41:10] innovation and everything and i'm excited about what the future holds but the destructiveness that [2:41:15] i've also seen as a person that's seen the transformative technologies of uh of of a lot of the [2:41:22] technologies of the last 25 years is is what really concerns me and one of the things especially with [2:41:28] um companies that are designed to want to keep my attention on screens and i'm not just talking [2:41:34] about new media uh 24-hour cable news is a great example of people that want to keep your eyes on [2:41:41] screens uh i have a lot of uh concerns about the corporate intention and and sam this is again why i [2:41:47] find your story so fascinating to me and your values that i believe in uh from our conversations so [2:41:54] compelling to me but but absent that i really want to just explore what happens when these companies [2:42:02] that are already controlling so much of our lives a lot has been written about the fang companies [2:42:07] what happens when they are the ones that are dominating this technology as they did before [2:42:13] so professor marcus does that have any concern the the role that corporate power corporate [2:42:19] concentration has in this realm that a few companies might might control this whole area i radically [2:42:25] changed the shape of my own life in the last few months and it was because of what happened with [2:42:30] microsoft releasing sydney and it didn't go the way i thought it would in one way it did which is i [2:42:36] anticipated the hallucinations i wrote an essay which i have in the appendix what to expect when you're [2:42:41] expecting gpt4 and i said that it would still be a good tool for misinformation that it would still have [2:42:46] trouble with physical reasoning psychological reasoning that would hallucinate and then along [2:42:50] came sydney and the initial press reports were quite favorable and then there was the famous article [2:42:55] by kevin roos in which it recommended he get a divorce and i had seen tay and i had seen galactica [2:43:02] from meta and those had been pulled after they had problems and sydney clearly had problems what i would [2:43:08] have done had i run microsoft which clearly i do not would have been to temporarily withdraw it from the [2:43:13] market and they didn't and that was a wake-up call to me and a reminder that even if you have a company [2:43:19] like open ai that is a non-profit and sam's values i think have come clear today other people can buy [2:43:25] those companies and do what they like with them and you know maybe we have a stable set of actors now [2:43:31] but the amount of power that these systems have to shape our views and our lives is really really [2:43:37] significant and that doesn't even get into the risks that someone might repurpose them deliberately for [2:43:42] all kinds of bad purposes and so in the middle of february i stopped writing much about technical [2:43:48] issues in ai which is most of what i've written about for the last decade and said i need to work [2:43:52] on policy this is frightening and sam i want to give you an opportunity it's my sort of last [2:43:59] question or so um don't you have concerns about i mean you you i i graduated from stanford the i know [2:44:07] so many of the players in the valley uh from vc folks angel folks to a lot of founders of companies [2:44:13] that we all know do do you have some concern about about a few players with with extraordinary [2:44:20] resources and power power to influence washington i mean i i see us i love i'm a big believer in the [2:44:26] free market but the reason why i walk into a bodega and a twinkie is cheaper than an apple or a happy meal [2:44:32] costs less than a bucket of salad is because of the way the government tips the scales to pick winners [2:44:38] and losers so the free market is is is not what it should be when you have large corporate power that [2:44:43] can even influence the game here do you have some concerns about that in this in this next era of [2:44:49] technological innovation yeah i mean again that's that's so much of why we started open ai we have [2:44:53] huge concerns about that uh i think it's important to democratize the inputs to these systems the values [2:45:00] that we're going to align to um and i think it's also important to give people wide use of these [2:45:05] tools when we started the api strategy uh which is a big part of how we make our systems available for [2:45:10] anyone to use there was a huge amount of skepticism over that and it does come with challenges that's [2:45:14] for sure but we think putting this in the hands of a lot of people and not in the hands of a few companies [2:45:21] is really quite important and we are seeing the resultant innovation boom from that [2:45:26] um but but it is absolutely true that the number of companies that can train the true frontier models [2:45:33] is going to be small just because of the resources required and so i think there needs to be incredible [2:45:37] scrutiny on us and our competitors uh i think there is a rich and exciting industry happening of [2:45:43] incredibly good research and new startups that are not just using our models but creating their own and i [2:45:48] think it's important to make sure that whatever regulatory stuff happens whatever new agencies may or may [2:45:52] not happen we we preserve that fire because that's critical i'm a big believer in the democratizing [2:45:58] potential of technology but i've seen the promise of that fail time and time again uh where people [2:46:06] said oh this is going to have a big democratizing force my team works on a lot of issues about the [2:46:11] reinforcing of of of bias through algorithms the failure to advertise certain opportunities and certain [2:46:18] zip codes um but you seem to be saying and i heard this with web three that this is going to be die [2:46:23] five uh decentralized finance all these things are going to happen but this is not this seems to me [2:46:29] not even to offer that promise because the people who are designing these it takes so much power energy [2:46:35] resources are you saying that this that my dreams of technology further democratizing opportunity and more [2:46:42] um are are possible within a technology that is ultimately i think going to be very centralized to [2:46:49] a few players who already control so much so this point that i made about use of use of the model and [2:46:58] building on top of it as it this is really a new platform right it is definitely important to talk [2:47:02] about who's going to create the models i want to do that i also think it's really important to decide [2:47:07] to whose values we're going to align these models but in terms of using the models um the people that [2:47:13] build on top of the open ai api do incredible things and it's you know people frequently comment [2:47:19] like i can't believe you get this much technology for this little money uh and so what people are [2:47:23] the companies people are building putting ai everywhere using our api which does let us put [2:47:28] safeguards in place uh i think that's quite exciting and i think that is how it is being democrat not not [2:47:34] how it's going to be but how it is being democratized right now um there is a whole new cambrian explosion [2:47:39] of new businesses new products new services happening by lots of different companies on top [2:47:44] of these models so i'll say chairman as i close that i have most industries resist even reasonable [2:47:50] regulation from seatbelt laws uh to we've been talking a lot recently about rail safety um that [2:47:56] the only way we're going to see the democratization of values i think and while there are noble companies [2:48:01] out there is if we create rules of the road that enforce certain safety measures like we've seen with [2:48:07] other technology thank you thanks senator booker uh and i i couldn't agree more that um in terms of [2:48:16] consumer protection which i've been doing for a while uh participation by the industry is tremendously [2:48:24] important and not just rhetorically but in real terms because we have a lot of industries that come [2:48:30] before us and say oh we're all in favor of rules but not those rules those rules we don't like [2:48:37] and it's every rule in fact that they don't like and i sense that there is a willingness to participate [2:48:44] here that uh is genuine and authentic uh i thought about uh asking chat gpt to do a new version of [2:48:54] don't stop thinking about tomorrow because that's what we need to be doing here and as senator hawley has [2:49:01] pointed out congress doesn't always move at the pace of technology and that may be a reason why we need [2:49:10] a new agency but we also need to recognize the world rest of the world is going to be moving [2:49:15] as well and uh you've been enormously helpful in focusing us and illuminating some of these questions [2:49:22] and performed a great service by being here today so thank you to every one of uh our witnesses and uh i'm [2:49:32] going to close the hearing uh leave the record open for one week in case anyone wants to submit [2:49:38] anything i encourage any of you who have either manuscripts that are going to be published or [2:49:44] observations uh from your companies uh to to submit them to us and uh we look forward to our next hearing [2:49:52] uh this one is closed