Building the rules for AI in America | Runway Summit SF 2026

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Building the rules for AI in America | Runway Summit SF 2026 · Channel: Runway · 24:58 · Watch on YouTube

Topics: AI governance and regulation, Markets and economy, Local and open-source models, AI safety

Summary

A panel of AI policy experts discusses the growing urgency of regulating AI in the United States, including public trust, immediate harms, competition, and frontier-model risks. They debate the benefits and risks of open-weight models, and how safety measures can address harm without concentrating capability in a few large companies. The speakers also examine antitrust, industry self-regulation, and California’s evolving approach to AI oversight. They conclude that consistency matters, but the substance of rules—and whether they support safety, competition, and open innovation—is more important than avoiding state-level variation alone.

Key points

The panel describes a growing public trust deficit and says AI safety issues require attention now, alongside the United States’ strengths in capital, chips, and talent.
Speakers distinguish immediate concerns such as deepfakes, fraud, and cyberattacks from more speculative catastrophic risks, arguing that both deserve attention.
They warn that access to frontier AI could become concentrated among a few companies, and stress that risk regulation should not deter broad diffusion of the technology.
Open-weight models can be inspected, adapted, and deployed in private environments, but their independent use and modification also create oversight challenges.
The panel discusses antitrust scrutiny of cooperation among AI companies, while noting that voluntary safety commitments may be pursued individually and could be enforceable under consumer-protection law if companies fail to follow stated practices.
California’s SB 53 requires frontier model developers to publish safety frameworks, report certain critical safety incidents, and protect whistleblowers.
California is considering stronger requirements, including third-party evaluations, evaluator standards, and ways to regain control of models exhibiting dangerous behavior.
The speakers say regulatory fragmentation can burden smaller developers, but argue that rules’ requirements and their compatibility with open innovation matter more than the number of frameworks.
Panelists express support for a strong federal standard, while cautioning that a national framework could also fail to work for smaller labs or open-weight models.

Chapters

Start	Chapter	Summary
00:00	Introducing the panel and the policy debate	The moderator introduces the panelists and asks which AI policy issues deserve the most attention.
01:33	Trust, immediate risks, and access to AI	The panel discusses the public trust deficit, present-day harms, emerging risks, and the possibility that frontier AI access will be concentrated.
05:11	Competition and existing consumer-protection rules	Maureen Ohlhausen describes the role of merger review and the FTC’s existing authority to address consumer harms.
08:19	Open-weight models: opportunity and risk	Ben Brooks explains why open models matter to organizations and discusses their benefits, risks, and possible safeguards.
12:28	Frontier safety and antitrust	The panel considers whether safety requirements could become a regulatory moat and how voluntary commitments intersect with antitrust and FTC enforcement.
17:42	California’s approach to AI oversight	Vivek Viswanathan outlines California’s existing requirements and potential steps involving evaluations, standards, and regaining control of dangerous models.
21:20	State and federal rules, and regulatory consistency	The panel weighs compliance burdens against the importance of strong, workable standards that preserve competition and open innovation.

Quotes

we now have a problem that is not in the future and needs to be dealt with right now. — 02:04

It is deep fakes. It is fraud. It is cyber attacks on critical infrastructure. — 04:08

You need to be able to inspect those models before deployment. — 09:52

Can we raise the bar? Can we do more on evaluation? Can we do more on mitigation? — 22:21

the keyword there, of course, to Ben’s point is strong — 24:27

Concepts and entities

Runway, Vivek Viswanathan, Ben Brooks, Black Forest Labs, Maureen K. Ohlhausen, Federal Trade Commission, open-weight models, SB 53, Flux Reaction, NVIDIA, Gavin Newsom, Andrew Ferguson

Full transcript (original language)

00:00 Please welcome to the stage senior counselor to the governor Vivek Viswanathan, Black Forest Labs head of public policy Ben Brooks, Wilson Sonsini partner and former FTC acting chair and commissioner Maureen K. Ohlhausen, and Runway head of public policy Evan Engstrom. >> All right. Hello, I’m Evan Engstrom. I’m the head of public policy here at Runway. Over the past few weeks, it’s been a wild time to work on AI policy issues as

00:31 questions around AI safety and foreign competition have brought the regulatory world into the mainstream. If like many people this is all new to you, you’re unsure of what’s going on, it’s a little scary, don’t worry. We have a fantastic group here to help break it all down for you. AI policy experts representing all aspects of the regulatory world. Starting to my left, Vivek Viswanathan, the senior counselor to Governor Gavin Newsom on AI and technology issues, Ben Brooks, head of

01:02 public policy at Black Forest Labs, and Maureen Ohlhausen, partner at Wilson Sonsini and former acting commissioner chair and commissioner at the Federal Trade Commission. Let’s start broad. There’s been a lot going on. Hard to keep track of all the developments in the AI policy space. If you’re just getting grounded in what’s happening, you’re catching up, what should people be paying attention to? In your mind, what are the most pressing issues? How do you separate the

01:33 signal from the noise? Vivek, we’ll start with you. >> Sure. So first, thanks Evan for having us and to the Runway team. I will say as you mentioned Evan in the last several weeks, the concern around these issue has skyrocketed. The good news is that the US looking broadly still in a very strong position when it comes to capital, training on the most advanced chips, and the best talent in the world for maintaining the models that really are at the frontier. The bad news, as everybody in this room knows, is we had

02:04 now have a trust deficit that has built up over the last several months and years that is growing. These increasing safety incidents, some of the stories are pretty harrowing, and recognizing that we now have a problem that is not in the future and needs to be dealt with right now. And you’re seeing physical manifestations of this trust deficit, looking at issues like data centers that have popped up like a lightning rod Infinity’s across the country. But at a broader level, I think for a number of years the discussion was about whether to regulate. And I think now we’re

02:35 seeing that being sunsetted, and the question is not what do you regulate, but how do you regulate? And you know, for now, we’re coming up on 4 years since Chat GPT 3.5 was released. That was in 2022. In 2023, on Capitol Hill, they had a series of forums with leading technology leaders to consider possible legislation. We had a big presidential election in 2024. In 2025, we had the rise of these coding agents becoming increasingly powerful. In the last year, we’ve seen agentic AI,

03:06 both the good and the bad, but what we haven’t seen is a comprehensive set of legislation to govern this framework, and I think people are wise enough to the need for that right now. >> Makes sense, Ben. >> I mean, I’d agree. I think if you look since Chat GPT to now, there’s been this tsunami of legislation at the state and federal level. Just this session alone, there’s I think 800 AI bills in Congress, about 1,500, 2,000 at a state level. And I think there’s still a lack of consensus about what risks matter,

03:37 where in the tech stack can we intervene, and what does regulatory intervention actually look like? And I think we’ve started to see some focus over this summer around cyber, loss of control, some of these sort of forward-looking catastrophic risks being much more prominent than they were historically. I think one challenge in my mind is that I’m kind of twofold. One, we’re taking some of the oxygen away from acute concrete here and now risks that still deserve a lot of attention. You ask the average American family, you ask the

04:08 average policy maker, what do they care about in AI? It’s not loss of control through recursive self-improvement. It is deep fakes. It is fraud. It is cyber attacks on critical infrastructure. Um and so we need to be paying attention to these risks alongside the more forward-looking speculative um risks that have dominated headlines. I think there’s another piece of this, which is we are at risk of maybe losing sight of some of the economic risks pertaining to frontier AI and frontier capability. We have a handful of Bay Area companies saying that AI is going to be critical

04:40 infrastructure, frontier capability and access to frontier capability is going to be the difference between sinking or swimming in the future economy. Um but I fear that we are to some extent sleepwalking into an environment where you can only rent that capability uh through the paywall of two or three companies. And I think it’s vitally important that as we address these sort of catastrophic risks, um we do so in a way that promotes rather than deters the diffusion of this technology and these models and this research more broadly outside of uh

05:11 these big uh big tech labs. >> I mean, that’s a perfect uh transition to your area of expertise in competition. What are you looking at? What is most prominent uh in your mind? >> Uh so so in my mind, coming from an FTC background, and that’s where where I practice in both antitrust and consumer protection, uh starting with kind of the competition side of things, is we’re seeing as companies um are trying to innovate and compete uh and invent the

05:43 future and and for US companies to to maintain their leadership, that they need the inputs to continue to develop AI. And those inputs, they might be chips, they might be power, they might be, you know, other capabilities. And making sure that our merger policy in particular is understanding the need for these um for these resources to be pulled together in an

06:13 efficient way. And And we’ve seen, I think, a lot of a lot of mergers happening that are pro-competitive and bringing these complimentary pieces together. So, we’re seeing a lot of activity in that space. And so far, I think the antitrust agencies are calibrating correctly. Right? They’re not trying to say, “Well, let’s kind of slow down everything through a merger tool, through a merger review, simply because we want to slow the pace of innovation,” which isn’t really what antitrust is

06:45 supposed to be doing. But then, on the the second the second half of the equation at the Federal Trade Commission, we talk about regulation. But the FTC Act has been in place a long time, and it prevents deception and also unfair acts or practices. And those are things that cause substantial injury to consumers that consumers can’t reasonably avoid and that aren’t outweighed by benefits to competition. So, the FTC has for a long time been using these authorities

07:17 to try to ensure that new technologies, whatever they may be, and now they’re very focused on AI, aren’t causing consume these kinds of consumer harms. So, when we’re talking about new new regulations, I think it’s important not to lose sight of the the regulation the more general regulation that that’s been in place. We We talk more about it, but there’s been some FTC announcements just in the past really few hours about focusing on some of these issues. So, I think that’s

07:48 something I know I’m paying a lot of attention to and you know, with with clients making sure that they’re aware of kind of the government expectations here and that they’re ready to to address those and you know, make sure that these kind of substantial harms aren’t happening to consumers. >> Excellent. I’d like to turn back to that in a moment. The recent FTC news always breaking news in the AI policy space it seems these days. But I want to first

08:19 dive into an issue that has gotten a lot of attention over the past couple of months open weight models. So, in the LLM context the focus has really been on the risk of Chinese open weight models competing with perhaps building off of American Frontier AI models, but that’s really just one facet of the debate. Ben, you work at a lab that is developing open weights in video and image and robotics. We just announced open weight robotics model.

08:51 So, there’s a whole angle to this open weight debate that I think is getting lost. Can you speak a little bit about what you think policy makers are getting right, what they should be thinking about, what are the risks, what are the appropriate mitigations? >> Yeah, so I think you know, as a bit of context open models open weights are a an important part of our DNA. Our founders co-developed three of the six most liked open models of all time on hugging face. We released as you say this action prediction model flux reaction last week that’s topping the videos leaderboards.

09:22 So, it’s a really exciting time in the visual space because in the visual space the US and Europe really do have a chance to lead the world in good Frontier open weight models. I think what policy makers are generally forgetting is kind of twofold. One piece is around why open models matter. Like what’s the opportunity here? I think if you’re an organization of any kind, whether it’s a big bank or a small garage band startup, AI is going to be a critical part of your tech stack. You need to be able to inspect those models before deployment. You need to be able to modify them for specific

09:52 purposes. Um you need to be able to modify them and run them in a secure private environment where your data isn’t constantly being exposed to someone else’s API. Um and so that’s why it’s really important for organizations to have access to good, capable frontier or near frontier open models. And I think a lot of policy makers don’t appreciate why this will become important over the coming years. I think for America as a whole, you know, you think of models as a kind of a vehicle for projecting soft power abroad. China understands this. China is currently being more American than America in some ways. If you ship models, you shape the global

10:25 tech stack in ways that will last a very long time in indelible ways. Uh and equally, you create a lot of dependencies on your homegrown labs. And that’s why in the language model space, the best open weight models are coming out of China, and why in the visual space, other than Black Forest Labs and a and handful of others, uh most of those visual multimodal models are coming out of China as well. Alibaba, ByteDance, and Tencent. I think a lot of policy makers don’t understand or don’t appreciate maybe um the risk side of the ledger, either. I think there is, you know, to be very

10:56 clear, there are no silver bullets in the model layer. You can’t fix a lot of these risks in the model alone. The properties that make open weight models useful, which is that you can run them independently and you can modify them independently, um also pose unique challenges from a risk management and an oversight perspective. And so the question then becomes, how do we think about the risk of the model, the capabilities of the model in the context of the wider tech stack and the wider ecosystem? What opportunities do we have to reduce the risk, not just in the model, but in other parts of that tech

11:27 stack? And how can we work together with other players in this ecosystem to bring that risk down to an acceptable level. We’ve done a lot of work in pre-training and post-training our open weight models to make sure that it is as hard as possible to use those models for nefarious purposes like generating child sexual abuse material, non-consensual intimate imagery. Um and we’ve done in third-party independent red teaming that shows that our models have 10 times fewer vulnerabilities than the open weight models coming out of China. So, you can take a few things away from that as a policymaker. One is that there is huge

11:58 variation across the industry in terms of the best practices applied by different developers. I think another piece is that there is a lot you can do in the model itself to bring these risks down and to reduce these vulnerabilities. But, the third piece is that there is still risk. It isn’t COVID zero. You can’t extinguish that risk in the model itself. And so, we need to think along with authorities, civil society, the wider developer community about how what we can do to kind of bring that risk down, keep it down, and to to raise the bar over time. >> Yeah, on the risk side of course and

12:28 like the physical AI is in particular, we need confidence in the safety and security of the models. You know, with it as they say, you know, with enough eyes all bugs are shallow. The value of of open weights is you can lower that uh that risk tolerance to get people more comfortable with deployment. So, it’s it’s interesting to think about open weights as really a pro-safety feature in a lot of ways as well. Um on the safety side, Marine, the other big issue we’ve been seeing dominate headlines relates to frontier model safety, pacing the frontier. Uh the current holder of your former role

12:59 at FTC chair, Andrew Ferguson, said recently that some of the larger industry players are trying to use this safety debate to build a regulatory moat uh by asking for onerous uh obligations, uh pre-release reviews that maybe only the bigger labs can handle. Um what is your overall view on the state of competition in the AI space? How do you see the role of competition regulators like your former colleagues at the FTC in shaping AI policy? >> Yes, so um I was actually at that event

13:30 when he made those remarks and I had um interviewed the head of the antitrust division uh on a similar topic uh and that was about 2 weeks ago. And in that 2 weeks, we’ve seen some very interesting developments. I think as this discussion has moved forward. So, from an antitrust viewpoint, uh for entities competitors to get together and agree not to uh expand output or develop uh you know, a new new capabilities,

14:02 that raises, you know, antitrust concerns uh because uh they should maintain independent centers of decision-making and and re- remain competitive. But on the other hand, I think it became clear there these concerns were, you know, had genuine motivations behind them, right? Not just, you know, moat digging, but uh you know, how do how does an industry who is facing these kinds of concerns react as an industry?

14:32 And there’s actually a long history of self-regulation. Uh so, the in some ways, these are not new issues. I’ve seen them occur across many, many industries. Uh but the the um I would say the speed at which these things are coming up in AI is is particularly uh notable. So, just yesterday, we saw an announcement uh at the White House uh of a number of the big frontier model companies um and um Nvidia, you know, some some other people

15:03 in the ecosystem uh signing an agreement for them to individually take steps to protect safety. Um so, they’re maintaining that individual uh approach. They’re not agreeing that they’re all going to um you know, I’ll only do it if you do it, kind of uh, safety uh um, precautions. Uh, and we saw that uh, Chair Ferguson was at the event. So, I think um, what we’ve seen is this kind of fast

15:35 evolution to say, well, what’s a way that we can address these goals and stay clear of an antitrust concern? Because otherwise, for entities to do it collectively, they would need some sort of antitrust, you know, immunity for that, which the antitrust agencies had said they, you know, they had concerns about. Um, so, uh, so that will be interesting. We’ll see, we’ll see how that, how that develops. Uh, but it was a very, I think, consistent with traditional antitrust principles approach and a

16:06 self-regulatory approach. And there’s some, there’ve been some questions raised already, well, if they make this promise, uh, but they’re not making a a promise, it’s not a law, uh, it’s just a promise, what tools can be used uh, if they don’t adhere to that promise? So, that is actually a very traditional part of self-regulation, where if an entity says, I’m going to adhere to these best practices or third-party developed um, standards or standards that I’ve publicly said I will, um,

16:38 that I will adopt, and they don’t uh, do it, then the FTC, using its consumer protection authority, could bring an enforcement action. And we’ve seen that in past regular, uh, uh, self-regulatory spaces. Uh, then interestingly, just this morning in the New York Post, there was a story suggesting that the FTC is also starting um, to to look into some of the the um, practices and safety practices of some of the AI companies. Um,

17:10 now, the FTC, I mentioned its enforcement powers, it also has the ability to engage in industry-wide studies. So, it’s potentially that that’s kind of what they’ll pursue there. But I But that That’s what I think you’re seeing happen just like in a very accelerated manner in this consumer protection and antitrust intersection that’s becoming more intertwined every day. >> Yeah, I think it’s the area where the federal regulatory apparatus is moving most quickly, certainly.

17:42 Vivek, on the state level, things are moving even faster. You are the senior counselor for Governor Newsom on AI issues. You have the challenging position of actually having to craft rules to regulate the industry. And California has moved a lot faster than the federal government. There’s been inaction as federal policymakers debate whether to preempt state laws. Meanwhile, Governor Newsom has signed many AI bills in just the past few months governing things like AI chatbots, third-party auditors, synthetic media disclosures. It’s very active. How do you see the role of states in

18:13 regulating the AI sector? What is California specifically trying to solve for and and what would you say you think the state should prioritize in the next year? >> Yeah, absolutely. So, first of all, our fundamental law in California, SB 53, basically requires frontier model developers to publish a safety framework to make sure that any critical safety incidents that touch on catastrophic risk are reported and then as well protections for whistleblowers who, if they see something, can say something

18:44 without fear of penalty. And so, the governor signed that into law last year and that provides some transparency into what is happening. The next step that I think we are all recognizing now is twofold. Number one, seeing that we have a much wider range of loss of control and safety incidents than I think some may have anticipated at the outset or at least they wanted to see how the technology would develop. And in particular, I think one thing that we’ve all seen in recent weeks is that maybe before there was a focus on deployed

19:15 models or testing right before deployment, but a lot of the incidents that we’ve all heard and read about are actually now happening in the train training phase, in fact, even when they are supposedly in a sandbox. And so, the range of scenarios that you have to consider is much wider than it was before. And then the second issue is that under the current law, you have to publish a safety framework. There are some pressures to adhere to it, but in many ways, certainly not the federal government and even at the California level, the heavy stick has not really been

19:45 brought down in terms of saying, you have to adhere to these types of standards. You have to have third-party evaluators come in. They have to meet a certain level. I think what the White House put out yesterday, as Maureen mentioned, it was these companies agreeing to a voluntary set of commitments. There may be some pressures associated with that that could lead to a better outcome than we have currently, but what the governor is looking at, and he put out an executive order on this about 10 days ago, is basically saying, number one, should we actually be requiring these third-party evaluations

20:17 under law? Should they in fact be embedded in the companies the way in other industries, banks, you have greater visibility with embedded auditors? Should these third-party evaluators have to meet some type of standards so that you make sure that they are sufficiently rigorous? Who is involved in setting those standards? And then, and this got a lot of news, should we advance the development of some type of kill switch or other way to regain control if these models engage in very dangerous behavior on an ongoing basis? And so, none of these questions have easy answers, but

20:50 there is growing concern about them both in the technology community and among the broader public. And so, the governor has signed a number of laws, as you mentioned, on specific topics related to kids safety, related to deep fakes, related to AI in the economy, and the transformative impact that we’re about to see, but I think a lot of both the future requirements as well as the capabilities that government needs to help govern and guide the development of this technology in a safe way. That is for the coming months and so it’s something that the

21:20 state is very focused on and we are paying close attention to what other states are doing as well as what the federal government is doing. >> To that last point, when we talk about state versus federal regulation, there’s often this claim of well, a patchwork of of regulatory approaches across 50 states presents a difficult compliance burden. Ben, as somebody who is working at a lab that has to deal with that potential patchwork, is it a real concern? If if it is, what would you think a better approach would

21:51 be that you’d recommend to to policy makers? >> Yeah, I mean I think I mean it is a concern. Fragmentation disharmony between different frameworks is a concern, but I I think the tech industry in general just fixates a little bit too much on preemption and the fragmentation of law across the country. I would rather as developer 50 different frameworks so long as they are consistent with open innovation and visual intelligence rather than one framework from one state or the federal government that isn’t. And so I think we need to ask ourselves more about the substance of the

22:21 underlying rules. Do we have whether we’re talking about kill switches, export controls, evaluation regimes, self-regulatory auditing? Are the standards being set in a way that requires you as a developer to exercise control over every copy of that model anywhere in the world, in which case this is a very bad idea in terms of open innovation and competition at the frontier? Or are they consistent with open innovation and with competition? Can we raise the bar? Can we do more on evaluation? Can we do more on mitigation? Can we impose new duties of

22:52 care on developers? But we can we do so in a way that’s that’s compatible with the widespread diffusion of the technology? Those are the questions we need to ask ourselves. And so I think um while a lot of the sort of public discourse around state regulation has been dominated by this concern that there’ll be 50 different frameworks, they’ll all pull in a different direction, and developers will be a sort of underwater or the sky will fall down. Uh I think the real question is what exactly do those frameworks require of developers? What is the standard of compliance? And is that standard

23:23 consistent again with widespread diffusion, widespread adoption? >> Okay. Can I Can I just jump jump in on this because this is a debate that’s happened you know, across across industries. And I think one of the um from a competition viewpoint, one of the under-appreciated things is that if you do have so many different regulatory regimes to comply with, that is uh difficult for smaller players. And it actually ends up being bigger players who have the resources to hire

23:56 law firms and professionals and and overseers to try to comply with all those different frameworks. So, having consistency is is very important, but having uh if it is inconsistent, it can create this kind of drag on smaller players in the market. >> And I’ll just say one thing on that, which is you know, speaking for the governor’s office, um he himself has also said a strong federal standard would make a lot of sense. You know, we prioritize startups in California, don’t want to have a crazy compliance burden

24:27 on them. But the keyword there, of course, to Ben’s point is strong, you know, and so right now um Congress hasn’t passed any type of comprehensive uh AI safety or security law. And so I think that’s something that needs to be resolved. >> And I will say most of the most exotic and eclectic ideas have come out of Washington. So, there is a risk with a national framework that you actually end up with a framework that doesn’t work for small labs, doesn’t work for open weights, doesn’t work for visual intelligence. Uh and to some extent, California, New York, a few other states have been

24:58 experimenting with some frameworks and with some ideas that might look very interventionist at first glance, but there is a pathway to making that work, and that might actually be the preferable approach. >> I think we’re going to see a lot more action over the next year at the state and federal level. Thank you so much, uh, for your insights uh, and helping break down a complex topic for our audience. Thank you. >> [applause]