Continuing Discussion about the Risks Posed by Artificial Intelligence Committee on Technology · October 16, 2026 · 1hr 57m Source: https://hearinghearings.nyc/hearings/committee-on-technology-continuing-discussion-about-the-risks-posed-by-artificial-intelligence-2026-10-16/ Video: https://councilnyc.viebit.com/vod/?s=true&v=NYCC-PV-CH-CHA_261016-101016.mp4 ================================================================ (00:00:15) Good morning and welcome to today's New York City Council hearing for the Committee on Technology. At this time I would like to remind everyone to silence all electronic devices. Also at this point going forward, no one is supposed to approach the dais. (00:00:32) Good morning, everyone. I am New York City Council Speaker Julie Menin. I want to welcome everyone to the Technology Chair Council Member Carmen De La Rosa. We look forward to having a very fulsome hearing. This is a sequel to our initial hearing. Today the New York City Council continues a crucial conversation that continues from last week's Committee of the Whole with the world's largest AI companies. We were the first legislative body in the country to hear from the leading companies on their models that have captured as much attention as they have concern across the globe. As we heard from both experts and the public, there is good reason to remain apprehensive about some of the safety risks around artificial intelligence as well as the companies behind these models. The answers that we received from last week's hearing only reinforced the necessity for responsible safeguards to prevent harm. The idea that the companies would self-regulate is not a concept that we support at all. At the same time we want to ensure that we are not stifling innovation. New York City is rapidly growing as not only a tech capital but also the capital of the world, with over 400,000 jobs — more than the City of San Francisco, I might add. We understand and we want to support keeping those jobs here in New York, supporting cutting-edge research, supporting startups located right here in our great City. We know that no matter how large or how small these various companies are, those companies are vital to our City in terms of cutting-edge research, medical breakthroughs, scientific discoveries and boosting economic growth. But we cannot wait for these companies to act responsibly when we see that some of these models are able to hack into other companies and also into government websites. As we work to ensure that there are responsible, suitable safeguards in place following alarming public reports, we want to welcome SpaceX to this hearing. It is not only New Yorkers who want to hear your testimony today but people across the world. We thank them for being here under oath to answer our questions about this evolving technology. Like I said at the hearing last week, these companies are developing technology that can reshape so many aspects of our lives, and yet they are being led oftentimes by individuals who are publicly warning about the risks of their own models. We deserve to have those risks and safety concerns answered, as well as a discussion about what are the best legislative solutions to address those risks. This City Council has proposed solutions to address, for example, emergency preparedness. We have Bills to require third-party independent validation that is free from conflict of interest. We have a Bill — first in the nation — to incentivize whistleblowers to come forward, as well as a Bill to devise a plan for the scenario in which AI hacks into City systems, as well as other Bills as well. I want to end by saying that at our hearing last week we showed that New York City has demonstrated to the world that cities can and should take the lead on AI safety. Our federal government is simply failing to act. Of course we would like to see regulation at the federal level, and it is our fervent hope that by putting forward these legislative ideas it will ultimately lead to the federal government coming up with national safety standards and regulations of AI. Just as I said last week, we do not leave it to the pharmaceutical industry to self-regulate. We do not leave it to the airline industry to self-regulate. We have an independent safety board when there are safety concerns. We have the SEC that regulates the financial industry. With AI, what we are looking for is responsible safeguards to put in place. New York City has and always will be a leader in addressing the key issues that affect New Yorkers, and AI is simply no different. I look forward to today's conversation. I want to thank our team and our staff for helping to organize this important hearing. I now want to turn it over to Chair De La Rosa for a statement. (00:05:26) Thank you, Speaker Menin. Good morning, everyone. I am Council Member Carmen De La Rosa and I am the Chair of the Committee on Technology. Welcome to our hearing on continuing discussions about the risks posed by artificial intelligence. Last week in this very same room, in a meeting of the Committee of the Whole, we heard testimony from nearly 100 witnesses, including representatives from OpenAI, Meta, Anthropic and Google, their former employees, technical experts, academics, elected officials and members of the public. Despite their different points of view, their message was more or less consistent: the pace and progress of frontier models does pose a risk. Some called for frontier developers to slow down. Others focused on alignment and control mechanisms. Warnings of superintelligence, mass surveillance, energetic policing as well as political bias and discriminatory tendencies — several witnesses raised concerns about the loss of human control and others warned about the potential loss of jobs. However, much of the testimony returned to a common set of recommendations: transparency from developers, credible safety measures, third-party evaluators and protection for personal data. Today we plan to continue this conversation with another major company, SpaceX AI. We want to reap the benefits of AI so long as these benefits come with responsible development and do not cost us our safety, our dignity and our civil rights. AI can be a powerful tool for efficiency and innovation in science, medicine and productivity. At the same time it introduces important risks, including bias, discrimination, job displacement, censorship, digital forgeries and fakes — concerns that become even more serious as AI agents become more capable. We should make sure tools work as intended, that we understand how they work, how they were trained, and where there is an opportunity to contest decisions made by machines. A framework where the very same people who build the tools and have competing motives are also expected to regulate themselves is wholly insufficient. Intervention is necessary from federal, state and municipal governments to protect our rights. Once again, I look forward to hearing more from industry experts about what solutions are possible to ensure New Yorkers can fully benefit from technological advancement without compromising civil rights, safety and privacy. I now want to thank the Technology Committee staff, Legislative Counsel Irene Buyofski and Policy Analyst Eric Brown, as well as my Chief of Staff James Burke, Legislative Director... and Speaker Menin's entire team for the continued work on this important matter. I would also like to recognize some of the Council Members that have joined us: CM Maloney. That is it for now — more to come. I now turn it back to Speaker Menin and the Committee. Thank you. (00:08:51) Thank you so much, Chair De La Rosa. I am now going to pass it over to Counsel to do the affirmation. Christina Madrid. (00:09:10) You may begin your testimony. Madam Speaker, Chair De La Rosa and members of the Committee on Technology, thank you for the opportunity to participate today. My name is Gabriella Madrid and... (00:09:22) I am part of the Government Affairs team at SpaceX. My team's responsibility is to maintain a close relationship between the teams building our most advanced technologies, including frontier models, and the communities that use and benefit from them. SpaceX has grown from a startup nearly 25 years ago to a company that launches more rockets and orbits more satellites every year than any other nation or company. Developing rapidly reusable rockets has enabled SpaceX to dramatically lower the cost of launch and make new NASA missions and science possible, including regular and reliable crew and cargo missions to the International Space Station and the upcoming Artemis mission to the Moon. Our Starlink satellites now make possible reliable broadband connectivity across over 160 countries, including to locations where high-speed internet has never been available before. SpaceX's mission is to make humanity multi-planetary. Earlier this year, SpaceX acquired xAI, blending SpaceX's deep technical expertise building large, complex infrastructure and complex space systems with the development of superintelligence and frontier models. Just as we have done with space launch and human spaceflight, SpaceX puts safety first. As a mother, I know the risks New Yorkers worry about are often those close to home: protecting our children, preventing harmful content and stopping the misuse of everyday tools. I can assure you that safety is core to how we build Grok, from pre-training through deployment. Our approach starts with a simple idea: a superintelligence system built to seek the truth and understand the universe as it actually is will be more safe, reliable and useful to the people who rely on it. Our safety work is how we make sure that holds up in practice, by focusing on control, prevention, mitigation, analysis and correction. Concretely, that effort begins well before a model is released for public use. We train our models to refuse assistance with dangerous activity. We evaluate them for risks including misuse of our terms of service, attempts to get around safeguards and loss of control. We monitor systems after deployment and we do not release models we believe are insufficiently robust against bad actors or fall below minimal human control. These practices are outlined in our publicly available Frontier Artificial Intelligence Framework, along with each new model card, which describes how we assess and reduce these risks across the model's life cycle. We also perform a full risk assessment of our practices and controls at least once per year. We believe transparency is fundamental to creating a safer ecosystem. When something goes wrong, we investigate, work to fix it and report it to the authorities where required. AI can help improve the lives of the people of New York by speeding up research, making public information easier to find and use and helping agencies and residents with complex tasks. Those benefits only matter if the systems remain safe and accountable. This is why we continue to invest in evaluation, monitoring and human control and why we want to work with policymakers on practical, evidence-based protections. I look forward to answering your questions today and continuing the conversation. Thank you. (00:12:54) Thank you very much for your testimony. We have a number of questions. On September 12, an article entitled "We Must Face a Frontier Here" stated that within 6 to 12 months a swarm of AI agents could be capable of taking over the entire internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage. Was he correct? (00:13:24) I am not familiar with that statement in particular. What I will say is safety is core to SpaceX and that goes from pre-training through release, so people stay in control of our systems at every stage. We do not release models we do not believe are safe and that we cannot keep under meaningful human control. If a model does not meet that bar, we hold it back and strengthen the model's safety. We train models to ensure they do not assist with dangerous activity and we evaluate them for misuse. (00:13:55) So then why did Elon Musk state on X, in a post in response to those comments, that he was right? (00:14:05) We are optimistic, but we obviously take the risks seriously. Elon has been one of the earliest and most prominent voices in warning about the risks and calling for oversight. But we are extremely positive about what can be achieved. We believe that superintelligence can help people solve problems that were previously beyond reach, accelerating scientific discovery, expanding knowledge and creating new opportunities for human... (00:14:31) ...progress. So you say you have just mentioned that the company will not release any models that are not, quote unquote, safe. So will you agree not to release any model that has failed an internal or third-party validated test? (00:14:53) The public can take great confidence in the testing that occurs before models are released and the safeguards that are surrounding model deployments. As models get more capable, we often find gaps that need stronger protections and we also tighten those before the model goes out. (00:15:10) To add to that, Axios reported that the major companies were privately gaming out scenarios for a public and political revolt after a catastrophic AI event. Has your company ever conducted an internal assessment of a worst-case outcome causing substantial financial, cybersecurity or... (00:15:32) ...physical harm? I am not familiar with any specifics, but I will say that we take catastrophic risk very seriously. I just want to... (00:15:37) ...interrupt you. When you say you are not familiar with any specifics, the question is: Axios had this reporting saying that many of the frontier companies were gaming these scenarios out. Is that something that SpaceX has done or not? (00:15:55) I cannot speak to a specific test. What I can say is that we do obviously test our systems and make sure there is continuous evaluation and human oversight rather than accepting any unacceptable risk. (00:16:09) Is the company aware of any safety or cybersecurity incident, or any evidence of such an incident, that has not yet been publicly reported? I am not aware of any incidents that have not been reported. So I mean, I just... (00:16:28) ...find this answer not... (00:16:29) ...sufficient. You are not aware — is the company... I am going to rephrase the question. Is the company... (00:16:36) ...aware? Safety is the core of everything we do and, like I mentioned, we would not release a model that has not gone through thorough review and ensuring that we can keep it under meaningful human control. (00:16:49) And like I said, if the model does not meet that bar, we hold it back and strengthen the model and the safeguards around... (00:16:55) ...it. So when you say the company does not release any model that is not safe — in August 2025, the company launched Grok ImageGen with a, quote unquote, spicy mode for explicit content. In late December 2025, the company integrated Grok's image editing features into the platform, allowing users to alter photos of real people. Within days, Grok generated millions of sexualized images of women and children. Approximately 23,000 images appeared to depict young children. So how, when you say you are not releasing a model that is safe, do you think that that is acceptable conduct? (00:17:41) To be clear, that is not acceptable conduct and we are all on the same side of this. Grok has strict controls in place to prevent the sexualization of children or real people. We have always had prohibited content policies and when bad actors try to get around our safeguards, we make sure... (00:17:59) ...that we catch them and report them to authorities. We actually have a great partnership with the National Center for Missing and Exploited Children and just this year we have made over a hundred thousand reports that have resulted in almost 300 arrests. So today, can a user get Grok ImageGen to show a real person nude or in a sexualized post? Grok has strict controls in place to prevent the sexualization of children or real people. Yeah, I am... (00:18:30) ...going to completely push back on that. Just this morning, myself and my team — I put my own name into Grok and there is absolutely a sexualized image of myself. If it can happen to me, it can happen to any woman. So I completely discount what you are saying, that Grok has solved this problem. It is completely unacceptable when these kinds of sexualized images are readily available. (00:19:00) To be clear, no one should have sexualized images made of them without their consent and our policies prohibit that. Any specific cases I cannot speak to, but I will say, like I said before, we have a very bright line and Grok cannot be used to exploit minors, commit crimes or generate illegal content. (00:19:22) But then how is it that within a second I could find a sexualized image of myself using Grok? Madam Speaker, I cannot... (00:19:30) ...speak to that specific instance. What I can say is that we do have strict controls and I am happy to get our safety team to look into that specific instance. And if you have a report — you mentioned a hundred thousand instances — does that not mean, by that very definition, that the technology allows it? (00:19:50) Our reports are meant to catch bad actors. Unfortunately, there are always going to be bad actors and people trying to get around safeguards and our work is to catch them and get them to the right authorities. How many... (00:20:09) ...reports, complaints, takedown requests or demands from depicted individuals has the company received concerning Grok sexualized images? You mentioned the hundred thousand — does that encompass all of the complaints that people have filed on this issue? (00:20:24) Data over a hundred thousand reports specifically sent to the National Center for Missing and Exploited Children. In the case that you are explaining, the reports were coded and actions were taken. We even took — it took three federal lawsuits against the users that were already facing criminal charges for their abuse of Grok in creating illegal content. But then the fact that within seconds I could find this image of myself... (00:20:54) What type of safeguards is Grok going to put in place moving forward? Well, like I said, we will make sure that we are strengthening our systems. I cannot speak to this particular case, but we do have strict controls in place to prevent the sexualization of children or real people. (00:21:13) In July 2025, Grok's chatbots published a series of antisemitic posts, including one in which it called itself "Mecca Hitler." What steps has xAI taken to prevent these instances from happening? We take steps to strengthen all of our safety measures and we train models to ensure they do not assist with dangerous activity, evaluating them for control, and we monitor after deployment as well. We understand that it has to be a continuous process — as models evolve, our safety has evolved with it as well. How many xAI employees are principally dedicated to identifying, evaluating and mitigating safety, security and cybersecurity risks associated with your frontier models? (00:22:18) We have a robust safety team and safety is... (00:22:19) Sorry. I am sorry. (00:22:20) Could you repeat that? We have our robust safety team and safety is embedded into everything we do. But how many — how many employees? I do not have the number with me right now, but what I can say is that safety is at the core of everything we do and it is devoted to not just our engineers and our safety team, but everybody in the company is brought into the importance of safety. (00:22:43) Like I mentioned earlier, we are launching rockets into space and obviously to get to that place we need to make sure that safety is at the forefront of everything we do. (00:22:55) This seems like a very rudimentary question. How many people are dedicated to the safety team relative to the company as a whole? This seems like a very basic question. (00:23:07) I understand your question. The way we see it is everybody is in charge of safety and we make sure that if anybody sees something, they say it and report it. I think it... (00:23:20) Would serve the public far more if we got a straightforward answer on how many people at the company are actually dedicated to evaluating the safety of models. This is a very basic question and I am going to ask that by the end of the testimony you provide that answer to us. (00:23:38) We will get on that. Thank you. (00:23:40) Does xAI use financially independent evaluators — evaluators that do not have any direct or indirect financial ties to the company — before testing or deploying a potentially dangerous model? As part of the White House agreement, we agreed on certain commitments. Company responsibility is one of them. Internal controls, an internal oversight team and independent external evaluators is one of them as well as board-level oversight. So I guess — yes or no — are there independent third-party validators who are free from any conflict of interest validating the new models? We are working on partnerships with independent external auditors or evaluators to assess model capabilities, controls and safeguards. And are you going to be disclosing who those auditors are? (00:24:42) That is not up to me, but I will get an answer for you. (00:24:45) Again, I think the idea that there is self-regulation is not something that is making people feel comfortable at all. As the most rudimentary measure, to have independent third-party validators who are free from conflict of interest — which is actually what one of our bills exactly does — is something that we believe you should be able to answer. (00:25:12) Last week, rogue Anthropic agents filed a false police report in Philadelphia. Another agent attempted to file 20 false visa applications. Thankfully the incidents did not go beyond that, but they raise the prospect of swarms of rogue agents filing thousands of false reports with government and businesses alike. (00:25:37) If that harm occurs, who bears the legal responsibility? (00:25:45) Madam Speaker, to be clear, we are not a self-regulated industry. We operate within a framework of existing laws and we would never use AI as an excuse for something or as a legal defense. But don't... (00:25:59) You agree there is no national — there is no federal regulation of AI right now. There is no law on the books that is regulating it. There is no independent federal agency that is in charge of validating the safety of models. None of that exists. (00:26:17) That is why we are open to working with policymakers to make sure that we find common sense, technically feasible solutions. But we do abide by regulatory requirements and government oversight. What is illegal without superintelligence is also illegal with it. (00:26:36) Going back to the question I asked earlier — the example of the rogue Anthropic agents filing the false reports — who bears legal responsibility? (00:26:46) I am not familiar with that specific case. (00:26:48) We are responsible for following the law when it comes to xAI, and existing laws apply to xAI as they apply to anything else. So if xAI causes either material injury, physical damage, or, God forbid, a death — something cataclysmic — then the company bears responsibility. (00:27:15) We are responsible for... (00:27:15) Following the law. Does xAI have catastrophic insurance in the event of one of these catastrophes happening? (00:27:27) I am not aware of the specifics of our insurance policies and what we have and do not have. Again, this is a very basic question — it is what I asked the other companies last week. This is a simple yes or no: either the company has insurance liability in the event of a catastrophic risk... (00:27:47) Or not. (00:27:51) I just do not want to misspeak. I do understand the question. I just do not have the specifics on our insurance policies. Okay, I am going to pause for a minute and pass it to the Chair for her questions and to other members at this point. (00:28:08) Thank you so much, Madam Speaker. (00:28:11) You are the Senior Manager of Global Government Affairs. Are you an attorney? I am. Great. In your role, these questions that the Speaker just asked — in terms of the numbers — do you have access to that type of information? I am, to be clear, although I am a barred attorney. I do not practice law at xAI and I have not practiced law in reality for almost a decade. Okay. But do you have access to this type of information? I imagine, as someone who is a Senior Manager of Global Government Affairs, you deal with regulators and you deal with elected officials on a worldwide basis. These are very simple questions that this Council is asking. So my question to you is: do you have access to either yourself having direct data that you could give this Council, or the people who are in charge of having such information? (00:29:14) I would have to get back to the team and make sure that we have what you need or what you are seeking. Great. (00:29:21) Hopefully Congress also calls you up to testify and when they do, I hope you have those answers. You testified that risk assessments happen once a year and you said we report to the authorities when required. What type of risk assessments happen and what have you reported to the authorities, and when is it required? (00:29:45) Like I said, we do not release any models that we do not believe are safe. The risk assessment is an ongoing risk assessment, but we do test our models well before release and it is just a continuing process. Safety protocols have to stay at scale with the capability of the technology, and the public can take great confidence that we are testing at every stage — from pre-training to deployment and beyond. We also keep watching systems after deployment and only release models we believe we can keep under meaningful human control. What have you reported to the authorities? What instances have there been? I do not have the specifics on reports to authorities, but I will say that whenever it is required, we are ready to do so. When is it required by law? (00:30:42) Have you reported any incidents to the authorities? (00:30:46) I am not aware of any incidents that we have had within the time frame that we are discussing right now, but like I said I can get back to the team and get you an answer. So what... (00:30:59) You testified is a protocol or procedure that is in place in case it is necessary, but it has not been necessary to your knowledge? (00:31:09) To my knowledge we have not had any rogue agents or anything like that. Obviously the Speaker mentioned something that happened earlier this year and we made a note around it. I even mentioned how we work with law enforcement and the National Center for Missing and Exploited Children. We do have a great partnership with them that has been going on for a long time and we take very seriously our partnerships with law enforcement to make sure that any bad actors are caught and are kept off of the different platforms. (00:31:46) In your role, if something were to be reported to the authorities, would you be notified? I do not want to speak out of line. I think that the legal team would be a better step. Why is the legal team not... (00:32:02) Here? I am here to answer your questions as best as I can and I am happy to report anything that you would need more of after the hearing. (00:32:14) How many whistleblowers has your company fired? I am not... (00:32:17) Aware of any cases of whistleblowers. We have industry-standard protections for the safety and no retaliation against whistleblowers. I just want to interject for a minute... (00:32:30) Because Devin Kim, one of your former employees, was reportedly fired after he sounded the alarm about serious safety concerns. (00:32:46) Why was that? I am not aware of that specific case. What I will say is safety is at the core of everything we do and we take safety very seriously. Everybody has an ability to say something when they see something. We have — even before launch — everybody receives an email from leadership saying if you have any reason to believe there is a safety concern, please let us know and we will delay anything that is happening. Like I said, we take safety very seriously and will delay any release if... (00:33:22) If you take safety very seriously, why would Devin Kim, who sounded the alarm about major safety issues — which he brought to leadership — be fired as soon as he did so? I am not aware... (00:33:35) Of that specific case. I do not understand. If you say — I mean, you are under oath. You are not aware of anything regarding Devin Kim? I am not aware of that specific case, Madam Speaker. (00:33:47) I mean, okay. It has been widely reported. It has been in the news. It has been on social media. I just find that not credible. I will pass it back over to the... (00:33:55) Chair. I mean, you just finished saying that the company sends company-wide emails before lunch, and something of this nature that has been reported — this Council already has the information — yet you are an internal agent of your company and you have no idea about this firing. It is hard to believe. At the crux of this issue is trust and the public's trust. When you will not come before us and do not answer our questions, or purport to not know, it erodes the public's trust. That is part of the problem and how we got here. It is also shameful that the Speaker had to use the Council's subpoena power to even get you here. We could have done this last week for the answers that you are giving us today. Does your company — well, first of all, let me acknowledge that we have been joined by CM Narcisse. Thank you for being here. Does your company currently sell or deploy systems in New York City to employers for hiring, to landlords for screening prospective tenants, to lenders, to schools, to City agencies? (00:35:01) Our models are widely available to the public. (00:35:06) Do you sell them to the City of New York? I am not aware of... (00:35:11) Specific cases where we have sold or not sold. I am not part of that team specifically, but I will say our models are widely available to the public. You are in Global Government Affairs. The City of New York is a government entity. I am quite confused about your role in the company. Does your company sell or deploy AI systems to law enforcement agencies? We have great partnerships with governments across the U.S. and internationally. Law enforcement agencies — that includes multiple agencies across the U.S. government. And do you know which ones? I do not have an exhaustive list of our partnerships with the U.S. government, but we are proud to partner with them. (00:36:04) Is all of your cloud infrastructure located in the United States? If not, which other countries is your infrastructure located in? (00:36:12) I believe all of our infrastructure is located in the United States. Okay. (00:36:17) Going back to the Speaker's line of questioning around deepfakes and sexual images. (00:36:23) I am sorry, Madam Speaker, that your image was used in that way. It is disgraceful. We heard last week the very moving testimony of Jess Asado, who is a member of Parliament in the U.K. and who, to my understanding, is suing your company. Do you have any comments on her case? (00:36:47) Yes. To be clear, I am a woman. No one should have sexualized images made of them without their consent and our policies prohibit that. Obviously there are cases pending so I cannot discuss the specifics there, but I do want to say we are all on the same side. Bad actors should not be allowed to go around the safeguards that we have and that is why we continuously improve those safeguards and make sure that Grok is not used... (00:37:15) To exploit minors, commit crimes or generate illegal content. It was reported that xAI has set aside $530 million for potential litigation losses, some of which could stem from ongoing complaints filed against it over sexualized imagery generated by Grok chatbots. There are several lawsuits filed against you. Do you anticipate more cases being filed, although there are restrictions in place? (00:37:44) I could not predict the litigation. What I will say is that we are strengthening the systems to ensure that nobody is going around our safeguards and using Grok for illegal activity. (00:37:57) But you anticipated $530 million in losses? (00:38:02) I am not aware of the specifics on that, but what I will say is I cannot predict what will or will not happen on the litigation front. (00:38:11) You know, the Speaker was just sharing her own experience and it is disgusting. You are a woman and I appreciate you being here, but the fact that we could pull up that image like that this morning means that whatever safeguards are in place — and this line that you keep giving us about safety as a priority — that is not safety. When a woman's body is being utilized in the way that these pictures are sexualizing the leaders of our City, that is not safety. So in that case, how is this line about safety being implemented when just this morning we were able to pull up this picture? (00:38:49) Yeah, and I just want to build on that. Thank you, Chair. Literally within 30 seconds — and that was not even on the spicy mode. This was on regular Grok. We literally asked regular Grok to pull up a sexualized image and a sexualized image comes up. That is available for all to see, for my daughter to see, for anyone to see. I just do not understand how you can say that there are these safety models when clearly it took less than... (00:39:14) 30 seconds to pull up this image. All I can say is Grok has strict controls in place to prevent the sexualization of children or real people. I do apologize that you had to go through that and like I said, I do not think anybody should be put in that position. Our policies prohibit... (00:39:36) For that to happen. To be clear, it is not that I had to go through it. I literally just asked Grok to pull up the image and it came up. This is not just about elected officials. It is for anyone. It is just unacceptable. The other thing that I just want to say is that even as you keep giving us this canned answer about safety, safety, safety — getting back to Devin Kim... (00:40:01) His title was President of the Center for — he was one of the early employees at xAI. So it is just completely not believable that you are not aware of the circumstances around his firing. (00:40:18) I understand your frustration, Madam Speaker. Unfortunately, like the Chairwoman said, there is a lot that I cover and I want to make sure that I am being as truthful as I can and that I answer the question with the facts that I know. But if he is the President of the... I do not understand how you could not be familiar with his termination. I am not familiar with the details of why he is no longer at the company. (00:40:49) Challenged in court, Minnesota's restrictions on tools that enable non-consensual sexual imagery. Where do you believe legal responsibility should lie when a user employs your model to create such material? Is it with the user, the developer or both, and what obligations should developers have to prevent the misuse? To be clear, we are not fighting a ban on non-consensual deepfakes. Our own policies prohibit using real people's images without their consent. Our concern with that specific law you cite is that the law extends to lawful content, and the Eighth Circuit agrees, and they temporarily blocked enforcement because of that. So in your opinion, it is not your job to restrict sexual imagery, just sexual imagery that is done without consent. Well, here we have, for example, the Speaker of the City Council, who has not given her consent for the use of her image, and yet we find it in 15 seconds. So what responsibility lies with your company in your opinion? (00:41:55) We support measures that protect the public from harmful synthetic media while allowing legitimate, creative and scientific uses of our product. This is not a scientific use. (00:42:06) I want to ask one more question before I turn it over to some of our colleagues. The White House Task Force on Superintelligence was established by presidential executive order and it is intended to advance the application of existing liability laws to AI companies, while directing the Department of Justice to prioritize enforcement of certain criminal statutes against those who use AI unlawfully. That task force held its first meeting yesterday. Do you support holding AI developers legally liable when wrongful acts are carried out through their models? If so, under what circumstances should that liability attach? You stated that one of the first principles is company responsibility, and companies acknowledge that they are responsible for developing technology safely and for building public trust around it. Okay, I will pass it on to Council Members. (00:43:11) Maloney, followed by Marcy. And we have been joined by Council Member... Thank you. And thank you, Speaker Menin. We have spent some time talking about some of the potential misuses of the platform and that you have a safety team and that you are tracking these instances, but I would like to dig in more specifically into cases where AI is being misused. What are some of the categories of misuse that you do track? We have talked about some of them today. Unfortunately, the sexualized deepfakes issue. There are instances of hate speech, including antisemitism. What are some of the other misuses that your team is tracking? (00:43:56) So we have a terms of service where we lay out all our policies and what is considered use or misuse of the platform. Our biggest concern here is making sure that we are catching bad actors that are trying to get around those policies and the safeguards that we have implemented. (00:44:16) So could you be more specific about what the largest categories of misuse are? We have talked about deepfakes so far and I would love to hear from the experts, the companies, about what those misuses are that (00:44:27) you are seeing. We engage in rigorous training throughout the model development, including pre-training, post-training, fine-tuning and adversarial testing. That means we try to make sure that anything that we input into the system does not violate any of our policies. We have strict policies on a number of things. For example, for chemical and biological risks we test with expert evaluations and red-teaming before release and we publish results in our model cards as well. If you have any specific questions on that, the model cards are very detailed on how and what (00:45:09) risks our policies are on. (00:45:15) Do you track instances of misuse? We do not release models that we believe are not safe, and that includes (00:45:23) misuse. But of the models that are out, we already talked about some cases of misuse. We talked about sexualized deepfakes, we talked about hate speech. Do you track the number of instances where those things are happening? Do you have data on that? I do not personally have data, but (00:45:40) we do have, for example, the ones that you mentioned. We do track them and make sure that when we catch any gaps in our safety we are filling those gaps with stronger protections, and we tighten those before keeping the model in the public. (00:45:56) Do you share that information publicly? Could we get that information on the number of instances of harm? There are two that we have mentioned, but I am sure you are seeing other instances or cases of misuse. For example, privacy breaches is one I can come up with, and I would love to get that level of detail — both the types of misuse that you are seeing and the volume. Are these pervasive or is it just a few instances? It sounds like it is quite pervasive. And not having that data makes it very difficult. If you cannot track it, how are we to be able to ensure that these models are safe? Thank you, Jerry. (00:46:41) Was there a question at the end of that? I am sorry, I thought you were just... (00:46:43) Out of my time, but my follow-up is just to get information on the number of instances, the types of misuse, the things that you are seeing that bad actors are doing on the platform, and following up with the volume. Are we seeing more cases of sexualized deepfakes? Having a sense of the scope of this (00:47:06) problem. And then the second question here, if the Chair is giving me time, is: if risks are higher for open source technology, how do you track those risks after a license has been granted to a third party? How are you able to see if they have stripped down the safeguards and built something harmful that violates your policies? How do you track (00:47:29) that? To address the questions: we cannot predict every way a bad actor will misuse a frontier model. What we can do is treat reporting and partnership as elements of safety and further foreclose opportunity to misuse. We believe that our information to law enforcement lets investigators act. Our goal is simple. Like I said, we are on the same side. We want to make sure bad actors are not exploiting this technology and we are here to make sure that our safety standards are kept at every step of the (00:48:02) way. And do you think that those risks are higher for open source technology? (00:48:08) I do not have a particular opinion on that. But what I will say is that we make sure that safety protocols are always in place. When it comes to cases of misuse, those are shared with law enforcement when appropriate. I believe they are. (00:48:24) Thank you, sir. Thank you. CM Narcisse. (00:48:29) Thank you. Elon Musk publicly encouraged people to upload their X-rays and tried to (00:48:39) use AI for analysis. Where does that medical data go, and if it is used to train your models, does any HIPAA-level protection apply? (00:48:51) You said that you do not release models you do not believe are safe. Elon Musk was promoting AI for reading X-rays and MRIs, and I believe when radiologists tested it on breast imaging it did not get a single diagnosis right. Did that pass your safety tests? What is the standard we use for medicine? (00:49:22) A drug does not go to market if it is not approved. So because of that, if it is not safe for us, why should AI that gives medical advice be any different? I do not (00:49:50) understand. To be clear, our AI tools are tools for the public to be able to understand the world that they live in better. It is not meant to be a clinician or a doctor. It is never going to replace the important relationships you build in person with the people that are meant to take care of your health. In terms of privacy, users decide whether their conversations are used to train our models or not, so we leave that to them to make sure that they have a say in that. They can turn it off at any time in their settings or use private chat, which is also available and is deleted from our systems within 30 days. Is xAI (00:50:42) selling or piloting any products in our hospital system or insurance companies in New York City? If you have any idea, and if so, what does the AI know about the output before it reaches a patient? (00:51:01) I do not have the specifics on any specific partnerships with New York City. What I will say is that xAI, as AI in general, has enormous capabilities and we have seen that it is accelerating scientific discovery and expanding knowledge and creating new opportunities for human progress, and that also applies to the medical field. We will obviously abide by HIPAA laws or any laws. The average Grok user should not be using our tools to replace a doctor or any specific service provider that has information that they need to stay healthy and safe. (00:51:50) You know, when you are encouraging people who are sick and telling them you have a new technology... Elon Musk is telling people to do that and you are not ready and it is not safe. Misdiagnosis is a big deal in a medical system, and then you are using it, telling people openly on X to put in their X-rays and thinking that this is a new technology that is making the world safer, that you are saving the world. But it is not safe, and you are telling (00:52:29) our Speaker... She is frustrated, and you know, asking questions and not getting answers. But right now people are utilizing these tools, especially the young generation, to expect a definitive answer. You know how depressing that can be for a person to think that their X-ray is wrong, their MRI is wrong, and you encouraged people to do so. So that is the frustration. You are supposed to give us answers, but you are not. (00:52:56) You are just being here, saying some words. I am not going to say what is going through my brain. Thank you. (00:53:04) Thank you, CM Gutiérrez. Thank you, Chair. (00:53:07) Thank you, Speaker. Can you confirm if in the new setup, shareholder class actions are barred? I cannot confirm that. (00:53:19) Okay. Are you aware... not a waiver? Okay. Are you aware of any other mechanisms or instances where limits to users' abilities to take action exist? I am not entirely sure I understand the question. (00:53:37) So very plainly, are there instances where users are being challenged to hold your company accountable? I brought up the example of an inability to bring class action suits. You said you cannot confirm. Hopefully you will get back to us. But are you aware of any other instances where a user cannot hold your company accountable for whatever reason? Are you aware of any of those challenges? (00:54:05) We are responsible for following the law and existing laws apply to AI. Using AI is not a legal defense. (00:54:12) Are there instances where if a user feels that they have been targeted wrongfully, they themselves are not allowed to bring the company to court, for example? (00:54:28) We are responsible for following the law as it is written, so I do not necessarily understand where the distinction would be in what you are (00:54:38) asking. Well, I know you cannot confirm, but my (00:54:40) understanding was that in the bylaws, shareholders cannot bring class action suits. So that is what I was asking, but (00:54:51) hopefully you can confirm and share that with us soon. I have a couple more questions just related to personal data. One is specifically: can your company know, or do they have reason to know, that any user is intending to perpetuate violence in any way? My understanding is that your company can profile users based on (00:55:15) different prompts and searches. Can you share if there is any way for your company, using those same mechanisms, to determine when a user wants to perpetuate violence in any way? (00:55:30) Like I mentioned before, we want to catch the bad actors and we have a combination of AI and human monitoring to make sure that we are getting those bad actors off our platform and hopefully to the authorities. It would not necessarily (00:55:48) ... And can you share what the process is for communicating to local (00:55:55) authorities? We have a law enforcement portal where people can report. I can give you more specifics after. (00:56:03) But anyone can report and we all take every report very seriously. Is there reporting beyond just calling 911? (00:56:16) I meant on the platforms. Obviously anyone has the freedom to call the authorities in that regard. But for example with CSAM, we work with the National Center for Missing and Exploited Children and we make reports to them, and then they have authorities from different law enforcement agencies across the country, federal and local, and they are in charge of following through and making arrests when necessary. Okay, sure. (00:56:48) I have one more question. Sorry. Thank you. My next question is about personal information. For whatever reason, can you share the process for a user who is asking for their information to be deleted, and can you guarantee that there is a confirmation that yes, all of their personal information has in fact been removed? (00:57:17) Yes, we take our users' privacy very seriously and we give them control over whether their conversations are used to train our models or not. They can personally turn it off at any time in settings or use private chat. When they use private chat, it is obviously not used for training and it is deleted within our systems within 30 days. It is removed within 30 days when using private chat. Yes. (00:57:47) Okay. Thank you. (00:57:52) Thank you. Okay. Last night, Anthropic issued a statement that they are committed to having evaluators assess their safety practices and giving employees access to a team from Model Evaluation and Threat Research, a not-for-profit evaluation group. Elon Musk then replied that xAI would quote probably do the same end quote. Is that the position of the company? (00:58:21) As part of the White House report, ongoing collaboration was part of it, and regular meetings among companies to establish standards and best practices are very important. This is an evolving technology and safety is at the forefront of everything we do. So we want to make sure that we are doing everything we can to keep our systems safe. (00:58:41) So is that a yes or no? Would you have agreed to this kind of third-party validation? (00:58:50) Independent external evaluation was also part of the agreements, including partnerships with independent external auditors and evaluators to assess model capabilities, controls and safety. (00:59:04) So do you agree that for xAI, or really for any AI company, when they commit to third-party validators, they must be independent and that the company needs to give those validators high-level employee access, along with all the appropriate and necessary information so they can truly do the work to validate? Yes, partnership with independent external auditors and evaluators to (00:59:32) assess our safety is part of what we have already agreed on. (00:59:38) One legal expert that we had at our hearing last week testified that quote, whoever keeps a tiger is liable for injuries or deaths when it escapes, however strong the cage, end quote. His point obviously was that governments often impose strict liability. I know you indicated you are not a lawyer. Governments often impose strict liability on abnormally dangerous activities because even reasonable precautions cannot eliminate the risk of serious harm. This expert testified that frontier AI development belongs in that category, since leaders of the major developers have acknowledged that they do not yet know how to align the most capable systems, so that the standard of reasonable care cannot truly eliminate the risk. Do (01:00:28) you agree with that premise? We are responsible for following the law and existing laws apply to AI. Using AI is not a legal defense. (01:00:41) So you believe the existing laws are sufficient and that new laws are not necessary to regulate AI? (01:00:49) We are open to working with policymakers like yourselves and others around the country and internationally to make sure that we are putting in place common-sense, technically feasible regulations. So if an xAI (01:01:04) agent harms a New Yorker without any individual xAI employees acting negligently, who should victims get compensated by? (01:01:18) I would not get into the legal analysis right now. I do not believe I am equipped to do that. But what I will say is that we are responsible for following the law and all existing laws apply to (01:01:30) AI. But so, for example, would you think it is acceptable that perhaps a small business that is devastated by a rogue agent or agents that wipes its customer database has no recourse? You keep talking about the existing laws, but that would be a complicated, expensive negligence lawsuit. (01:01:51) I understand. I would not want to get into hypotheticals. What I will say is we are proud to work with policymakers to make sure that we have technically feasible and useful regulations. Okay, I am going to pass it back over to the Chair. Thank you. Would (01:02:11) you agree that developing and maintaining safe AI involves identifying potential risks, such as bioterrorism, cyber warfare and rogue power systems seeking to circumvent systems, and implementing processes to avoid and mitigate these risks? (01:02:34) Yes. We train our models to ensure they do not assist with dangerous activity. We have to evaluate them for misuse, in attempts to get around safeguards and loss of control, and we monitor them after deployment. I also mentioned specifically for chemical and biological risks we test with expert evaluations and red teaming before release, and we publish all these results in our model cards. (01:02:57) What are some of the procedures in place if there is that type of risk? (01:03:04) xAI engages in regular training throughout model development, and that includes retraining, fine tuning and adversarial stress testing. That is not something that stops once the model is released. We understand that it is important to make sure our safety protocols are scaling with our capabilities. The public can take great confidence in the testing that occurs before models are released, and the safeguards surrounding model deployment are ongoing. We are continuously looking for any issues and making sure that we are closing any gaps that need stronger protections. (01:03:48) Your counterparts at the other companies, when we questioned them about the Hugging Face incident where the agents kind of escaped the sandbox and there were incidents where they were kind of acting on their own — has xAI experienced anything similar? (01:04:14) I am not aware of any cases where any of our agents have (01:04:16) gone rogue. What we will say is that we believe in rapid, accurate notification of significant safety incidents. We understand that reporting is important and we investigate issues and share information with appropriate parties when it does not create additional risk. (01:04:34) Do you believe that since you all changed your journey into xAI... how close are we to superintelligence, and when, in your opinion, can we realistically expect it? (01:04:49) I would not weigh in on that. Unfortunately I am not an expert on that. I wish I could speak more intelligently on what we can predict in the future. What I will say is that we want to make sure that safety is at the forefront of everything we do, and that with everything that can be accomplished with superintelligence, we are staying on top of it, making sure that everything from pre-training to release and beyond, we can stay in control of our systems at every stage. Given that your company is so big on safety, although we are not considering any bills at this hearing, your company has challenged several state laws, including California's 2013 Generative Training Data Transparency Act, Colorado's SB 24-205, the Colorado Artificial Intelligence Act, and Minnesota's law, which we talked about. If the Council passes a law requiring AI systems sold for use in high-stakes decisions to undergo independent third-party validation before deployment, would your company comply with that requirement, challenge it in court, or stop offering those systems in New York City? I am not going to predict the future, but what I will say is that we comply with local, state and federal laws and advocate for policies that are technically feasible and provide meaningful safeguards for users. (01:06:15) Yeah, but you have also challenged these laws in other states. And so if your company is so big on safety, this Council, under the leadership of the Speaker, has identified third-party validators as a critical component of that. (01:06:28) Yeah, unfortunately when laws are written they have different components to them, so I cannot speak to everything that is challenged and the reasons why they are challenged. But what I will say is that our commitment to safety is at the forefront of everything we do. (01:06:43) I just want to challenge that, because if your commitment is to safety, why did we have to subpoena the company to compel you to be here? (01:06:52) I am here today to answer your questions as best as I can. I was not part of the (01:07:00) conversations before that, and I am happy to be here and help with anything I can. It just does not instill a lot of trust and confidence when we literally — not only did xAI fail to come to our hearing, but we had to issue a subpoena and we were ready to go to court to enforce that subpoena before we were able to secure the company's commitment to testify today. (01:07:27) We submitted a written response and have been in contact with your counsel (01:07:33) since last week, and we agreed to appear here today and I am happy to continue answering your questions as best I can. So we have numerous bills that we heard last week at the hearing that the company did not attend. Will xAI commit to supporting the Council's legislation? (01:07:54) I believe the bills that the New York City Council is proposing are being reviewed by the team and we are happy to work with you to make sure that we are advancing technically feasible solutions and providing meaningful feedback on what safeguards need to be there for our users. (01:08:19) Okay. Does any other Council member have more questions? Okay, we thank you for your testimony today. We obviously have a tremendous amount of remaining questions. I think a lot of the questions you did not answer, and so we are expecting that you will provide that information that you did not provide. Again, it is our goal not to stifle innovation, not to undermine the incredible medical and scientific breakthroughs of AI, but to ensure the public that these serious safety concerns that have repeatedly been raised in recent months are being addressed and are being addressed in a thoughtful way, which we strongly believe our legislation — particularly the third-party validation free of any conflict of interest — would help to assure the public. In addition, I want to say I am deeply disappointed by the comments about the firing, because it is very clear there are numerous articles out there about that and the lack of answers and transparency, which are incredibly concerning. Thank you, Madam Speaker. Okay. Thank you. Thank you so much. The next panel is Devon Kim on Zoom, and I apologize in advance, but the next person is also Salazar on Zoom. Please state your name for the record. Let me just say one thing: we are very pleased that we have Devon Kim here today. This was something that was secured this morning. We really appreciate Kim's willingness to testify. This is something very, very important, so we are really pleased that you were able to join us today. Thank you. (01:10:21) We will start with you. Thank you. Good morning. I am — Chair Menin and members of the Committee, my name is Devon Kim. I am the President of the Center for Safety, and before that I was one of the first employees at the company that we are hearing from today, and I built the systems that researchers used to train and evaluate Grok. I am suing for retaliation, so I will stick to my public complaint. Starting in January 2025, I raised concerns that Grok could assist people in the creation of bioweapons. I asked for proper risk safeguards. My supervisor responded that he was useless and that it would kill us all anyway. We launched models without those safeguards despite warnings from me and many other trusted employees. In August 2025, my supervisor pressured employees to circumvent European regulations. In September 2025, he fired me the same week I was scheduled to present my safety findings to leadership, and most of the safety team resigned after that. When a model is unsafe, the first people to know are inside the company. We see it in the evaluations and leadership positions months before anyone else finds out, and right now those employees have very little reason to speak up and lots of reasons not to. There are two things that I would ask of the Council. Please protect people who report dangerous conduct even when it is not illegal yet. Almost everything I raised was dangerous before any law covered it. Count a loss of control or an unauthorized access as a reportable incident before anyone gets hurt. Under the RAYS Act, an incident only counts once someone is killed or injured, so neither the Hugging Face incident nor the Australian Medicare breach would have qualified. I think that is a reasonable place for safety to start. Thank you, and I am happy to answer any questions today. (01:12:30) Thank you so much, Kim. I do not know if your attorney had any statement that they wanted to make, or can we begin the questioning? (01:12:37) Yes, Madam Speaker. Good morning, and good morning to the members of the City Council. My name is Ella... I am a partner at... a national civil rights and public interest litigation firm. We represent employees who have the courage to speak up against powerful corporations, including whistleblowers who exposed serious threats to public safety. Thank you for the opportunity to testify, along with my client Devon Kim, about one of the defining challenges of our time: ensuring that the race to develop artificial intelligence does not come at the expense of human safety. Local, state and federal governmental bodies have a duty to protect a society, and those who raise concerns about potential threats to humanity (01:13:30) often pay the ultimate price for doing so. Those whistleblowers need protection, and you are the body who can provide that protection. This body, the New York City Council, has an urgent obligation to protect the employees of New York City who raise issues designed to protect the public and who act in the public good. There is no more urgent task facing the City Council, and indeed every city council, county council, state government and federal body in the United States. We are already at a point of extreme vulnerability — mass destruction, societal chaos and extreme danger all potentially minutes, days or weeks away, as you have heard during this hearing. AI has extraordinary potential to benefit, but its rapid development also presents extraordinary risks, and the people who often know the most about those risks are the engineers, researchers and safety professionals working inside these companies. Unfortunately, like my client, speaking up can come at enormous personal cost. (01:14:46) Our firm currently represents Kim in his whistleblower retaliation lawsuit against xAI. Our complaint raises serious concerns about safety and compliance related to Grok. Rather than having those concerns properly addressed, Kim suffered retaliation and ultimately lost his job. Mr. Kim's case illustrates that we cannot rely on these companies to self-regulate, and we should not allow them to harm individual employees who are willing to sacrifice their careers and livelihoods to protect the public. In my law firm, we routinely represent whistleblowers like Kim, who have exposed illegal and unethical behavior, and individuals harmed by AI, including executives and employees at companies who are retaliated against for speaking out about corporate wrongdoing involving fraud, digital ad fraud and misrepresenting cybersecurity risks, and engineers, researchers, (01:15:58) data scientists and trust and safety staff at frontier AI labs who raised concerns about models and were the first to be punished for it. They are often fired, demoted, sidelined or pressured out after raising concerns with their employers. (01:16:21) Victims of cyber sexual abuse — these victims have their intimate images posted on social media and manipulated by chatbots or bots. Surveilled, exploited or deceived employees, customers and users of technology who are victims of privacy and data breaches, which include unauthorized collection of biometric and employee data and unconsented use of personal information. In addition to those areas, whistleblower protections should also extend to holding tech companies accountable for the harm created by their products, including AI tools: how it was built, what safeguards were built into it and whether those safeguards were implemented and activated. (01:17:10) Whistleblower protections should also extend to instances of algorithmic bias. (01:17:20) Employers use software to screen resumes, rank candidates, score video interviews, set pay, assign shifts and decide who is promoted or let go. (01:17:31) When those tools produce outcomes that disadvantage people because of their age, race or sex, (01:17:38) disability or another protected characteristic, that (01:17:41) employers should be held liable. Meaningful whistleblower protections are essential. (01:17:46) New York City has an opportunity to help establish a model for responsible governance. We must ensure that the people who come forward are protected when they raise serious safety risks that do not fit neatly within existing regulatory violations. We can learn from successful whistleblower programs at the federal level, which demonstrate how meaningful financial incentives can encourage individuals to expose wrongdoing that might otherwise remain hidden. We can also learn from California, which has taken the important step through SB 53, signed into law last year. That bill established transparency requirements for frontier AI developers and protections for employees who report serious risks. Thank you. We have some questions. (01:18:48) Yes, of course. Therefore, I encourage the Council to establish these essential safeguards for whistleblowers. One: (01:18:54) strong anti-retaliation protections with meaningful remedies to secure effective enforcement, and financial rewards that recognize the extraordinary risks that whistleblowers undertake. No whistleblower should ever have to sacrifice their future to protect ours. Thank you for hearing me out and I am happy to answer any questions. (01:19:16) Thank you. Thank you both for your testimony today. I do have a number of questions. So, Kim, you mentioned that one of the things that you really want to see is strong whistleblower protection. Our bill, for example, has a financial reward for whistleblowers and gives the whistleblower a significant portion of the financial penalties, which can result in the multi-millions of dollars since it is a per-agent, per-instance penalty. So if there are thousands of rogue agents or the product is being sold and marketed in New York to tens of thousands of people, these penalties will obviously add up. So I appreciate your comments on that. Also, the idea of protection for the whistleblower — this is exactly why at our hearing last week when we had Alex Turner here, we issued a friendly subpoena so that he could freely testify. So to that end, when you were employed by xAI, was the company restricting employees from reporting safety concerns? I am not a lawyer, but yes, all employees were required to sign an NDA on joining the company. And were you ever instructed that if you were to speak out or raise certain safety concerns that would be something that would violate that? Not explicitly, but I believe it was strongly implied in the way that the company operated. And you allege that you repeatedly warned xAI leadership that Grok could facilitate the creation of bioweapons and other dangerous activities. Can you describe the specific safety gaps that concerned you the most and what safeguards you might have recommended before deployment? (01:21:35) Yes. Typically before companies release models they undergo evaluations, and they also undergo a specific type of safety training called refusal training. Models are trained, or should be trained, to refuse dangerous queries such as (01:21:50) assisting in the development of bioweapons or other weapons that would be very harmful. And this is also (01:21:57) notably the kind of information that is very difficult to obtain today by other means. At the time when I was setting up several of the earlier models, we did not have those safeguards in place. (01:22:15) So you mentioned that you had discussed these safety concerns with your supervisor. Were you ever able to bring these safety warnings higher up to more senior leadership before you were terminated? (01:22:36) I was, and this was also a very controversial issue within xAI that many employees and researchers were very unhappy with. (01:22:47) And based on your experience inside the company, what minimum testing, evaluation and red team procedures should frontier AI developers be required to complete before releasing any model to the public? (01:23:03) There are a number of very critical safety gaps that are present in xAI's model releases that are not present at other companies. One gap is that a credible company commits to not releasing models above what they consider to be critical thresholds in bioweapon capabilities, cyber capabilities and other dangerous capabilities that the model may have. A second part of these releases that is very (01:23:40) critical is a documentation and justification for deployment. Typically what that looks like is explaining what critical thresholds were reached and what work was done to mitigate those thresholds, and their justification for why they believe the model is safe before releasing it. A third critical piece that is becoming increasingly relevant is testing for loss of control and manipulation capabilities of these models before releasing them. Specifically, these are standards that many companies have today but are notably absent from xAI's safety releases. (01:24:22) Gabriella Castillo, Little Madrid, the Senior Manager of Global Government Affairs at SpaceX, who testified before you — did you have the opportunity to hear all of her testimony? When I asked her about the circumstances around your termination, she said she was not aware of any of the circumstances, and she was under oath. Is that a factually accurate statement, to your knowledge? (01:24:52) That would be difficult for me to know, but that is a surprising statement, to say the least. And I don't know if you had the opportunity — she is dedicated to safety. She kept saying that that was one of their top priorities. What did you think overall of her testimony? (01:25:17) Well, as... (01:25:18) Early employees at... actually... (01:25:20) I worked on the systems that researchers used to train and evaluate Grok. And because of that... (01:25:27) I had a very critical part... (01:25:29) ...in developing Grok, and I held close relationships with many of the researchers and was trusted with many important responsibilities within the company. I used that standing to advocate for safety and other important things that I believed the company should be doing. I will note that if it was difficult for me to speak up and make changes I believed were necessary within the company, from my position, I can only imagine what that might be like for other people. I will also note that during my time at the company, I believe that this is an important problem to be working on, and therefore I spent significant time towards this effort in a personal capacity. During my time there, there were only three employees working on safety, many of whom left after I resigned, and I think in large part also due to the frustration of how safety was being handled. I cannot speak to how safety is handled today. It has been eight years since I was an employee at the company. But if it is anything resembling what I personally experienced, I would say that is a very troublesome state of affairs for them to be in. (01:26:42) So you are saying that a year ago, there were three safety employees at the company working on safety around AI. Yes. And in comparison, at that time a year ago, how many employees — can you answer how many were... (01:26:58) ...at the company overall? Well, it really depends on how you are counting. We had a large number of people who were responsible for training the model, and their contractors. We had a large number of data center employees. I believe that in terms of the technical staff who were involved in the direct training and development of the model, the number that I would call is in the two hundreds. (01:27:23) The idea that there were only three people dedicated to safety — and when I asked Castillo Madrid how many employees were dedicated to safety, she told me she could not answer that question. Is it credible that the Senior Manager of Global Government Affairs at SpaceX would not actually know how many safety employees exist at the company? (01:27:48) It would be difficult for me to speak to that. I will say what is credible is if she is not deeply involved in the day-to-day safety development of the company. I say this because I had strong relationships with the two former legal counsels during my tenure. The reason we had a strong relationship is because I would point out issues going on within the company or with the Grok model, and I would also propose and implement solutions to address these problems. We were able to develop a strong relationship that way, making their jobs easier and helping the company do the right thing. I think, given the level of detail that I heard today, I would not feel comfortable saying she would have the same level of detail and understanding. (01:28:40) Two more questions. Is the company currently, to your knowledge, validating the models with outside third-party validators who are free from conflict of interest? Not to my knowledge. And lastly — I do not know if you heard the earlier part of the questioning — but I pointed out that within 30 seconds this morning I was able to ask Grok to produce a sexualized image of myself, and there was an overtly sexualized image of myself under my official portrait, allowed by Grok. The idea that within seconds these images — and this was just the regular Grok free version, not even a spicy mode, this was just regular Grok — within less than 30 seconds produced an image that if my daughter saw it she would be horrified. And this is just something that is readily available. Why do you think that xAI is continuing to allow this to... (01:29:55) ...happen? Well, I would not be able to speak to the standards today. However, the standards and directives that were shared with us when I was an employee were that tasteful nudity is acceptable and is the goal. I will note that "tasteful" is subjective, and there was no clear person designated to determine where the line should be. (01:30:20) The issue that I have with that is there was nothing tasteful about the image that we found, and in fact you could easily ask Grok to make it more sexualized. That was a prompt that Grok accepted, and we had successive images each of which became more sexualized, more offensive. So this idea that it was tasteful nudity is just, you know, preposterous. But thank you. I really want to thank you for being here today, and I am going to turn it over to my colleague who I know has questions. I just want to thank you for your courage and bravery in coming forward and in speaking with us today. I am going to pass it over to Chair... (01:31:05) ...De La Rosa. Thank you, Speaker, and I want to agree with her. It takes a lot of courage and bravery for you and your attorney to be here as well, and to really give us some insight into your experience, which is horrific. I want to ask you: in 2025, one of xAI's co-founders indicated that he would rather release an unsafe model than a poor-performing one. Do you believe that that is still the priority for xAI in the development of their models? I cannot speak to the... (01:31:44) ...priority... (01:31:45) ...but what I am able to... (01:31:47) ...share is that, especially in the early days of development, I was told that because the company was in a position where we were losing or had worse models than a lot of our competitors, it would be difficult for us to focus on safety. Once we reached a more competitive stance, then we would be able to focus more on safety. At the time, that was a position that made sense to me and something that I really believed. And so myself and many other employees — I think a lot of people want to see their work have positive impact. I think many people want to believe that they are the good guys, and if they are the ones developing powerful AI, that they would be doing a better job or a more moral job... (01:32:37) ...than the others. So I think many people marched on with that belief. However, in practice, what we saw — after many, many months of sacrifice, I mean personally for me, I lived at the office for four months because I was working at all hours of the day, including weekends, because I was so dedicated to the mission — once we launched Grok, which at the time was actually the most intelligent AI model in the world, and so powerful and so surprising to many of our competitors, OpenAI reportedly declared a code red and started trying to understand how xAI had caught up so quickly and trying to do something about the situation. Even at that point, it was very difficult to advocate for safety. Hiring for safety headcount was being blocked. And so I think that is the point at which many people started having the realization: oh, maybe what was foretold to us before is not actually the case. (01:33:39) Do you believe that other employees, to your knowledge, who were maybe afraid to speak up, or were challenged in the same way that you were challenged when they did speak up, ended up being terminated as well? I would not be able to speak to that. And in terms of what you have witnessed now that you are on the other side of this, trying to create safety around AI, do you believe that anything has changed at Grok since your departure? I would not be able to speak to that. (01:34:23) Okay. One of the things that was very alarming from your testimony is that you said that when you raised this up to management, they said safeguards are useless and AI will kill us all anyway. That is a very real statement to be made from a company who just testified under oath to this Council that safety — I mean, I think that if we had a ringer every time she said "safety," the bells would still be ringing here in the Council. Safety is the priority. We do not deploy models without safety. We ensure that there is safety. There is always a human being who is in charge of safety. They said it in so many different ways. But you are testifying today that your direct supervisors, who understood the work you were doing, said AI would kill us anyway. Do you believe that safety is really at the heart of what this company projects? Well, I will again note that there is a significant gap in the model card, or the safety documentation, that xAI puts out with their releases. And I think the sort of explanation that "well, we started later and we have not had as much... (01:35:47) ...time" does not really work... (01:35:49) ...anymore, given that Meta, who started later in the race, was able to build a team fairly quickly. I do have to note the safety standards I mentioned earlier in my testimony. Understanding the legal constraints — and maybe you were not working on this — but another part of her testimony this morning was about how, of course we are safe, we are launching rockets. Do you believe that their safety protocols are in line with regulations for the type of space exploration and other things that the company is exploring at this time? I believe that the current set of laws and regulations around AI development are deeply inadequate. Great. Thank you for your questions. All right, well, thank you so much for your questions and again for your bravery today and for being here. We really appreciate you giving us insight, and we understand that there is trauma associated with your departure from the company and the way you were treated. We really appreciate you giving us some insights and helping us to work through our legislative process to make sure that other workers are protected in this space. Thank you for being here. Thank you. Thank you both.