Webinar Video: Congress and the Future of AI: Model Testing, Preemption, and Open Weight Policy
Updated: Oct 15
The Bottom Line
Artificial intelligence presents real risks, but poorly designed regulation can create risks of its own. Congress should replace opaque and improvised policymaking with transparent rules, invest in independent technical expertise, and provide national clarity without displacing the states’ traditional authority. Above all, lawmakers must avoid turning legitimate safety concerns into an open-ended permission system that protects incumbent firms from competition.
On Aug. 28, the Committee for Justice convened a panel to examine three questions at the center of the congressional AI debate: how advanced models should be tested, whether federal law should preempt state AI regulations, and why continued access to open-weight models matters for American innovation.
CFJ Executive Director Ashley Baker moderated the discussion with Kevin Frazier, director of the AI Innovation and Law Program at the University of Texas School of Law; Neil Siefring, senior fellow at Alliance for the Future; and Paul Steidler, senior fellow at the Lexington Institute.
The panel did not produce a single regulatory blueprint—and that was one of its strengths. Its members agreed that AI policy must become more transparent, technically informed, and nationally coherent. They differed, however, over where assurance testing should end and mandatory government review should begin.
Congress Has Moved From “Whether” to “What”
As Frazier explained, Congress has spent several years asking whether it should legislate on AI. It is now confronting the harder question of what it should legislate.
He divided current proposals into three broad categories. The first concerns frontier-model governance: who may release the most advanced models, under what conditions, and with what assurances. The second addresses AI’s economic effects, including workforce disruption, small-business adoption, and access to computing resources. The third deals with physical and digital infrastructure, from research computing to data centers and electricity demand.
Despite the volume of proposals, Frazier was skeptical that Congress would enact major AI legislation during the current session. That delay does not mean policymaking has stopped. It means consequential decisions may instead be made through executive action, voluntary agreements, state legislation, and private negotiations.
That is precisely why institutional design matters. A policy can address an immediate problem and still be the wrong long-term governing structure.
Testing Without an FDA for AI
All three panelists recognized a legitimate role for AI testing. Their disagreement concerned the form that testing should take and the legal consequences that should follow.
Steidler focused on models that could create imminent catastrophic risks, particularly through cyberattacks against financial institutions, hospitals, utilities, and other critical infrastructure. He supported a mandatory, time-limited review of the most sophisticated frontier systems before release.
At the same time, Steidler rejected an FDA-style approach to AI. Drug approval is famously prolonged, expensive, and restrictive. Importing that structure into AI could delay beneficial products, limit individual choice, and turn technical evaluation into a general government license to innovate.
Frazier also supported a testing framework. He emphasized that advanced models can display capabilities that were not apparent before deployment and that reliable evaluation is technically difficult, expensive, and resource-intensive. Testing is therefore necessary both to discover hazards and to give prospective users better information about systems they may place in sensitive environments.
But Frazier argued that any federal system must be authorized by Congress and administered transparently. Secret or ad hoc arrangements between executive officials and selected technology companies are not a substitute for public, prospective, and predictable law.
Siefring added another essential caution: testing should focus on identifiable threats and actual uses, not treat the mere existence of a model as proof of danger. Congress must also examine the effects of compliance requirements throughout the AI ecosystem. Large laboratories can hire specialized lawyers and absorb substantial fixed costs. Startups and smaller developers often cannot.
As Siefring put it, “Regulation can become a moat.”
That insight should shape every AI proposal. A rule may be facially neutral yet still protect incumbent firms by increasing the cost of entry. A safety regime that only the largest companies can navigate may reduce the very experimentation and competitive pressure that help expose weaknesses and produce better systems.
Expertise Without a New Bureaucracy
Congress cannot write detailed technical requirements that will remain current as AI develops. By the time lawmakers define a particular model, benchmark, or testing procedure in statute, the technology may already have changed.
Frazier therefore urged Congress to set the governing framework while relying on institutions capable of updating technical practices. He pointed to the Rules Enabling Act as one possible analogy: Congress establishes authority and procedures but does not itself write every operational detail.
He also called for greater investment in independent expertise. As leading researchers leave universities and enter a small number of frontier laboratories, knowledge about the technology becomes increasingly concentrated inside the companies government is supposed to evaluate. Frazier described this as the rise of a “Silicon Tower.”
The panelists saw a possible role for the Center for AI Standards and Innovation, which works within the National Institute of Standards and Technology on AI measurement, testing, voluntary standards, and national-security risks. Steidler favored strengthening the center and drawing on outside evaluators. Siefring agreed that government needs better access to expertise but cautioned against reflexively creating another bureaucracy. Congress should first use existing institutions, university partnerships, fellowships, and private expertise more effectively.
The proper division of labor is critical. Congress should establish public objectives, legal limits, oversight, and accountability. Technical experts can then develop and revise measurement practices without requiring Congress to amend a statute whenever the technology changes.
Preemption Without Erasing the States
AI systems operate across state lines. A model is not developed in California, deployed in Arizona, and then stopped at the New Mexico border. Fifty inconsistent development regimes would impose particularly heavy costs on smaller companies and could make the rules of the most restrictive states national by default.
Siefring therefore argued for meaningful federal preemption. But he emphasized that preemption does not mean laissez-faire, nor does it mean eliminating the states’ traditional police powers. States would continue to enforce generally applicable laws governing fraud, discrimination, contracts, consumer protection, property, and personal injury.
Frazier proposed a related distinction between AI development and AI deployment. The federal government would establish nationally uniform rules for development-level questions, such as training requirements, pre-deployment evaluations, and model-release standards. States would retain greater authority over how AI is deployed in particular settings, such as hospitals, schools, professional services, or state agencies.
That division is a promising starting point, although it will require careful drafting. “Development” and “deployment” are not self-defining categories, and nominally local requirements can sometimes affect how a product must be designed nationwide.
The goal should be a federal framework clear enough to provide real national certainty, yet narrow enough to preserve state authority over local harms and context-specific uses. Preemption should not become immunity, just as state regulation should not become a backdoor national licensing system.
Why Open Weights Matter
Open-weight models allow researchers and developers to download, study, modify, and build on a model’s underlying parameters. Siefring argued that maintaining access to those models is essential to American technological leadership.
The international competition over AI is not merely a contest over which country produces the most powerful model. It is also a contest over which country creates the most attractive ecosystem for researchers, entrepreneurs, and developers.
Restricting open weights will not end experimentation. It may simply determine where that experimentation occurs. If researchers cannot build on advanced models in the United States, they will look for jurisdictions and platforms where they can.
Open models can present legitimate security questions, but those concerns should be addressed through evidence-based and risk-specific measures. A categorical approach would sacrifice competition, independent research, and decentralized experimentation without guaranteeing that dangerous capabilities disappear.
Data Centers Are Not AI Regulation
The panel also addressed the political backlash against data centers. Steidler stressed that data-center siting is fundamentally a local land-use question, even though public anxiety about AI increasingly shapes those disputes.
The distinction is relevant because concerns about a proposed facility’s water use, electricity demand, noise, or effects on neighboring property should be evaluated on their own facts. General unease about AI should not become a substitute for that analysis.
Likewise, Congress should not use infrastructure policy as an indirect means of restricting AI development that lawmakers have chosen not to regulate directly. Limiting the electricity, computing resources, or facilities available to developers can function as regulation “through the loading dock”—with less transparency and fewer procedural protections than a direct law.
Rules That Can Learn
The webinar revealed a broad area of agreement beneath the panelists’ differences.
Congress should not tolerate an indefinite system of secret, voluntary, and improvised executive policymaking. It should strengthen access to independent expertise, focus federal attention on demonstrable risks, provide greater national certainty, and preserve the states’ ability to address traditional legal harms. It should also resist moratoria and hyper-detailed statutory requirements that could become obsolete almost immediately.
The hardest question remains whether mandatory pre-release review can be confined to truly extraordinary risks without evolving into a general permission system. If Congress chooses that path, the scope, criteria, deadlines, transparency requirements, and avenues for review must be unmistakably clear.
AI policy does not require a choice between ignoring risk and freezing innovation. It requires institutions capable of learning as quickly as the technology changes. The objective should be rules that help discover problems, allocate responsibility, and permit correction—not regulatory moats that presume government can certify uncertainty away.
Learn more about the webinar and its panelists here.
Related CFJ work: “Certified Uncertain: Why AI Safety Needs Assurance Tests, Not Permission Slips” and “A Federal AI Moratorium Is the America-First Policy the Constitution Requires”.
Transcript
Ashley Baker: Good afternoon, everyone, and welcome to the Committee for Justice’s virtual panel discussion, “Congress and the Future of AI: Model Testing, Preemption, and Open Weight Policy.” I am your host, Ashley Baker, and I am the Executive Director of the Committee for Justice. I will now introduce our panelists.
Professor Kevin Frazier leads the AI Innovation and Law program at the University of Texas School of Law and serves as Senior Editor at Lawfare. His research focuses on how to design and update institutions to preserve the rule of law in the age of AI. Frazier has testified before Congress on AI regulation and regularly advises state and federal policymakers on how to govern emerging technologies. His scholarship and commentary has been published in leading law reviews, such as the Tennessee Law Review. You can listen to his two cents on all things AI policy via the Scaling Laws podcast or Appleseed AI Substack.
Neil Siefring is Vice President of Hilltop Advocacy, a government affairs and strategic consulting firm, and a senior fellow at Alliance for the Future, where he works on artificial intelligence policy, innovation, and the role of Congress in shaping a national framework for AI. Neil spent 15 years working on and around Capitol Hill, including in senior positions for members of the U.S. House of Representatives. His experience includes developing and advancing legislation, advising members of Congress on policy and legislative strategy, working with congressional leadership and committee staff, and building coalitions to advance policy priorities. He also served as an Appropriations Legislative Director to the Chairman of the House Appropriations Subcommittee. Before joining Hilltop Advocacy, Neil served as Director of Government Relations at FreedomWorks. He holds a BA from American University’s School of International Service, a Master’s in Philosophy from the University of Cambridge, and a JD from William & Mary Law School.
Paul Steidler is a senior fellow of the Lexington Institute, a public policy think tank based in Arlington, Virginia. He researches, studies, and discusses federal government policy on technology and logistics issues. Steidler identifies ways the federal government can work more efficiently and documents the full costs, direct and indirect, of federal government policies. His work includes regularly preparing op-eds, developing in-depth position papers, interviewing policymakers, providing comments to agencies’ dockets, and helping to organize events. Steidler’s work and views have appeared in news outlets such as the New York Times, the Wall Street Journal, Financial Times, Associated Press, the Washington Post, The Hill, National Public Radio, Bloomberg, and the New York Post. He graduated from Villanova University with a BA in Political Science and Honors. Steidler resides in Reston, Virginia.
So, before we get started, a quick note about audience questions. We welcome questions from the audience, time permitting. If you have questions for the panelists, you may go ahead and submit them through the Q&A function of Zoom webinar at any point during the webinar, and we’ll get to them whenever they’re relevant and whenever I can work them in.
So, let’s start here first with Kevin, who is going to set the stage a little bit for us on where this debate currently stands. Kevin, Congress has spent several years debating whether to legislate on artificial intelligence, and now it’s debating what exactly it should legislate. So my question is, where does that debate actually stand right now, and what is moving in each chamber? And from a realistic perspective, what is realistically likely to pass?
Kevin Frazier: Yeah, well, Ashley, thank you so much for the invitation to be a part of this great panel. It’s flattering to be among such esteemed guests, and I’m happy to dive into where the heck things stand. Unfortunately, I know we only have 60 minutes. I feel like I could probably go for six hours talking about where things stand, because everyone and their mom at this point is trying to introduce some degree of AI legislation. When folks go back home to their district, they like to talk about what they’re doing on AI, and being able to point to something has led to a proliferation of AI bills in three main buckets that I want to hit on.
So, the first main bucket to discuss is frontier AI governance. This bucket focuses on who should release the most advanced models, when, to whom, and based off of what assurances. And right now, when it comes to thinking through which labs, for example, should have the authority to release certain models at certain points of time, we are being governed by an ad hoc, voluntary, secret framework imposed by the White House. I will say briefly that this is not a good approach, in my opinion, with respect to things like the rule of law. As you noted, Ashley, I’m a professor at the University of Texas, and when I teach my students about the rule of law, I mention things like the law being public, the law being prospective, and the law being predictable. And the White House framework right now, determining when and how AI models are going to be released, doesn’t satisfy any of those conditions.
As a result, we see that Congress is contemplating its own approach to frontier AI governance. The most well-known bill right now is the Frontier Act, which was introduced by Representatives Trahan and Obernolte. As it stands, it’s getting a lot of attention and a lot of play on X, but there is a big gulf between something that’s debated on X versus something that’s actually going to make its way through the legislature. My hunch, if I were a betting man, would be that we are not going to see any action on the Frontier Act by the end of this Congress. It is a very lengthy and very substantive bill that would introduce a pretty novel form of regulation of some of these frontier models, and so I suspect that’s going to require some additional deliberation on the Hill before we see any movement.
Now, the second bucket that I want to focus on is bills addressing the economic effects of AI, so trying to help anticipate and respond to perceived and real economic disruption posed by AI. This, too, is an area in which we’ve seen bipartisan bills be introduced. There’s the PREPARE Act, for example, that’s been led by Senator Warner. That’s really trying to say, how can we improve data collection, such that we have better warning signs about which professions are being impacted when and how by AI? Relatedly, we’ve seen legislation like the AI for Main Street Act that’s trying to make sure that we don’t create a world of AI haves and AI have-nots. So how can we help, for example, small businesses learn how to use the latest and greatest AI and make sure that they’re leaning into this moment as well?
Then the third bucket that I’ll refer to is AI infrastructure. This is everything from how we address the power needs of AI to the compute needs of training the frontier models. So, there’s the CREATE AI Act, which would codify and increase funding for the National AI Research Resource, commonly referred to as NAIRR. I think this is something that perhaps could receive bipartisan support sooner rather than later, because we’re seeing that the need for access to compute, not only for the leading frontier model developers but across the economy, is going to become more important. We’ve also seen consideration of the Remote Access Security Act, which would help determine who has access to the compute that we know is especially important to training the latest and greatest AI. I suspect that, too, may have a greater chance of moving, given that it has some national security implications, and given that we’re increasingly aware of the fact that China perhaps isn’t as far behind the U.S. as some people imagined previously.
And then the final thing I’ll touch on within that AI infrastructure bucket is data centers, everyone’s hot topic these days. One thing that I think Congress may move on, and I hope they move on, is the idea of codifying the Ratepayer Protection Pledge that many of the leading developers have already agreed to at the executive level in a sort of informal manner. We could codify expectations that developers don’t pass on the costs of increases in electricity from data center development to consumers. So that’s my attempt at a quick, comprehensive approach to what the heck is going on on the Hill. But again, the top-line summary would be, if I had any money, I would not bet on substantive legislation passing by the end of this year.
Ashley Baker: You’re right, there’s a lot going on, and that’s a very salient point regarding national security. I’m going to ask Paul to follow up on that a little bit. So, Paul, what do you think are the risks of releasing models that could be exploited for cyberattacks? Are these systemic to the U.S. economy, and how high would you say these risks actually are?
Paul Steidler: Yes, thank you for the question, Ashley. Neil, excellent information. I would just encourage everyone to take that to heart. There are real risks with AI, but I think it’s important to explain that. I’m with an old-school conservative think tank, which is why I have a suit and tie on here today. And we very much believe that the economy functions best, and Americans are best served, when there is a minimal amount of regulation. Having said that, some regulation is necessary, and on the AI front, there are a lot of calls for moratoriums, and taking a time-out on it, and things along those lines. But the most serious risk that needs to be looked at is that of models that come out that provide imminent catastrophic risks.
And this is not a theoretical exercise. We went through this back in May and June, when Anthropic wanted to release its Mythos model, and this led to just some extraordinary activity in the White House. You had Vice President Vance, who was convening meetings with top people of executive agencies and the tech companies to talk about the risks that were inherent here. Something that is not as well known is that the Federal Reserve Chair, Jerome Powell, and Treasury Secretary Bessent met with the CEOs of major financial institutions to talk about what these risks were and how they were going to be addressed. The net result is that the Mythos model did not come out, and it did not come out in part because Anthropic warned about it and gave some warnings in that regard.
But the dangers are whatever you can conceive of, and these are not theoretical. It’s your bank account being wiped out. It’s your water company having its system hacked, so that unsafe amounts of chlorine and other pollutants go into the water supply. It’s about small businesses and hospitals, which do not have the guard against how the Mythos model could have been misused by nefarious actors, being able to protect their patients, being able to protect their business records. We really dodged a bullet there as a country, and a bullet that came quite close to hurting the U.S. economy.
I understand Kevin’s point about the administration making this policy on its own, and all of that is correct, but the corollary to that is, it worked. It stopped catastrophe from taking place. Is it the best way to do things going forward? I certainly don’t think so. There are a lot of questions that arise about keeping that policy in place. But we have to understand, there are some risks that are put out about AI leading to 20% unemployment, or nuclear missiles going off, or other stuff like that, which is, quite frankly, nonsense. But there are other real risks that come into play. As Kevin talked about the bevy of legislation here, we would submit legislation that’s going to govern models, that’s going to make sure that unsafe models do not come out, and that that is a required thing. In other words, the American people are not relying on the good intentions of Anthropic or another company; we will have a system in place to stop that.
I will just add to Kevin’s point: this executive order that has come out to govern these models, we don’t even know what’s in there. It’s a confidential order. That’s troubling. I don’t think there have been any congressional hearings about this, or letters that have been sent. That’s pretty staggering. We just cannot have AI policy made by elites in government and elites in tech companies. Sooner or later, they’re going to miss something, and the odds of catastrophe are going to be higher than if we have more systemic, widespread, known and honored practices. So, I guess much like Kevin, I could filibuster on this, but I’ll just leave it there for now.
Ashley Baker: Thank you, Paul. So, speaking of high-risk settings and supposedly high-risk models, I have a question for Neil on this, since there seems to be a lot of conversation around these high-risk settings. Where do you think Congress can actually play a constructive role without creating a regulatory system that slows innovation?
Neil Siefring: Well, thank you very much for the question, Ashley. Thank you for hosting, and thank you, Kevin and Paul. I’ve learned so much from both of you on X and elsewhere, so it’s a thrill to be part of this panel. I think one of the things that Congress has to do is make sure that we have federal testing that is not so rigid that it cannot keep up with emerging technology. We see how very quickly things move. In the AI space, we’re really at moment one of AI technology in its current iteration, so things are going to be moving really, really quickly. We have to make sure, also, that we’re looking at actual uses, not perceived uses, and not just the existence of models.
This gets back to Kevin’s outstanding legal analysis that he was providing earlier with his framework. It’s important to bear in mind what we don’t know, but it’s also important to say the existence of a model is not a de facto threat from the model. We need to actually have a regime that’s going to be able to keep up with the changes that keep happening. The key question Congress should be asking is, is any regulation that Congress passes going to make AI safer? Or is it going to impose burdens and a bureaucracy that’s not necessary for the tech and doesn’t meet the threats? So you have to have a clear knowledge of the model, a clear articulation and identification of the threats, and an ability to look across the stack to say what ripple effects federal regulations would have, not just on the larger companies, but on mid-sized and smaller-sized companies as well.
One of the challenges we have right now: Kevin, you come from law school, I went to law school. What’s the framework that we use to analyze this? The framework’s being developed right now, and that’s part of the challenge. We don’t have a three-pronged test to look at, for example. And that’s where Congress has an important role, to clearly tell the courts, this is our expectation.
Another quick wrinkle in this issue, too, is that many of these AI issues, data center issues, open-weight model issues, do not fall along typical party-line patterns. So it’s not easy to say Republicans think this, Democrats think that. The overlap portion of the Venn diagram is very wide here, so it’s a disorienting time on Capitol Hill, because you have odd political alliances taking place on these issues that have not taken place before. Sometimes the tech even outpaces the formulation of those alliances, and then people have to recalibrate.
In terms of what Congress may do, I thought Kevin’s analysis was outstanding. I would add one thing to that. We’re heading into a lame duck, and Kevin was alluding to this, that changes the calculation, because I don’t think you’re going to see massive AI legislation unless you have a full committee markup at House Energy and Commerce and Senate Commerce. There has to be a full process, and I don’t know that that’s going to happen over the next few months.
Ashley Baker: So, getting a little bit more into the implications and the nuts and bolts of these proposals, I have a follow-up question first for Neil. One concern about any of these proposals is that any large company, whether it be Anthropic, OpenAI, any large company, can absorb these compliance costs in a way that these smaller startups, smaller companies, cannot really necessarily deal with. So how should Congress think about this sort of risk?
Neil Siefring: It’s a great question. Regulation can become a moat. And it’s a moat that cannot be crossed by companies we can’t even think of right now, because they’re not even fully developed in the minds of the potential developers. So we have to make sure that compliance is foreseeable, it’s reasonable, and it’s actionable across the stack. What does that mean in real life? A smaller-sized AI company has to have really clear rules of the road, and should be able to comply with a minimum of effort in order to make sure that their product is safe and can come to market.
A lot of this fight, though, and I’d be interested in how Kevin and Paul and you actually think about this, is not necessarily big versus small. There’s a very complex AI ecosystem that exists now, and will be deepening over time, where smaller companies are relying on bigger companies, and bigger companies are learning from smaller companies. So we don’t need to pit large against small, small against medium. We need to say, how can we make sure that compliance is easy to understand, transparent, and simple to execute? Those have to be the basic rules. That also ties into open model weights, which I’m sure we’ll get to later as well.
Ashley Baker: And that’s a very good concern, that things are moving really quickly. I have a follow-up question for you, but Kevin and Paul, feel free to weigh in as well. AI obviously is changing quickly. We have new models every couple of weeks. How can Congress really avoid writing testing standards today that become obsolete as the technology evolves?
Neil Siefring: Kevin, sorry, go ahead.
Kevin Frazier: Oh, I think here we have to recognize that Congress has found ways to regulate fast-moving technologies previously, but that requires delegating to appropriately expert bodies, bodies that have the ability to stay abreast of the latest technological change. Here, I would like to see a lot more investment and consideration into how, for example, we can empower and increase the Center for AI Standards and Innovation, also known as CAISI. CAISI is a repository right now of experts in AI who are staying on the vanguard of the latest AI models. We’ve seen that a similar body, for example, the UK AISI, Jim noted that some of these real risks, real cyber risks in particular, have been identified and investigated by the UK AISI with greater detail and greater information than we’ve seen our own equivalent body, again, CAISI, be able to produce, because the UK, remember, little UK, is funding their version of our CAISI to a much greater extent.
And I know that some folks, when they hear the idea that we need to bolster and increase the number of experts in the federal government, may be thinking, oh, by golly, here we go, we’re just going to grow the federal government even more. And I hear that, and it’s a concern I have as well, but right now, we have to analyze these questions from a standpoint of “compared to what?” And the “compared to what” right now is the fact that AI expertise is being increasingly housed within a handful of AI labs. The folks who actually understand this technology, the people who are able to probe what sort of regulation is necessary to protect Americans and to also leverage this technology, are all being housed in San Francisco.
And I’m not being dramatic when I say that we are actually seeing the formation of what I like to refer to as the Silicon Tower. All of the folks who were studying AI at universities, including here at the University of Texas, are getting lured to San Francisco to work for these labs, and we don’t have a repository of independent expertise that Congress can lean on and call on to help create and adjust regulation as necessary as we see this technology evolve.
And so this is where I think that novel approaches, like the Frontier Act, which calls for the creation of independent verification organizations overseen by something like CAISI or by the Department of Commerce, could be a real positive model. I’ve shared elsewhere, and I’ll put this in the chat, that I have some constitutional issues about private delegation, which is something that Congress will need to address. But the fact of the matter is we cannot expect the 535 members of Congress to be diving into the latest details of AI. We need to be thinking about creative approaches to bring in expertise and make sure that these laws and regulations can be adjusted frequently.
Ashley Baker: Paul, I know you’ve had some specific thoughts about this.
Paul Steidler: Yes, thank you, Ashley. So, as far as legislation is concerned, we’ve talked about the profound hurdles for getting anything enacted in this session of Congress. I think it’s important that all capital and focus be on legislation that’s going to prevent catastrophic events, that’s going to prevent the really bad stuff from going down. And one of the key things there, as Kevin mentioned, is funding for CAISI, the Center for AI Standards and Innovation, which is part of the Commerce Department.
Just some numbers on this. This is a tiny organization. It has $20 million in funding this year. It should have at least $100 million in funding. By the way, you don’t need government bureaucrats to be the ones who are going to evaluate AI models. You can have industry peers. You can have other people with consulting expertise do that. CAISI has also had some turnover in its leadership recently, so it’s an organization that’s quite turbulent and problematic.
On the good news front, Michael Kratsios, the White House Advisor on Science and Technology Policy, testified to the House Science Committee last month that he wants to see its budget increased. He thinks the organization is important. So certainly bolstering CAISI, and looking for ways in which models can be systematically looked at for catastrophic risk, as opposed to managed for some type of societal, beneficial purpose, which gets into non-catastrophic issues. I would advise policymakers to proceed in that manner.
Ashley Baker: Neil, do you have any concluding thoughts on that?
Neil Siefring: I would agree with both Kevin and Paul. We’ve got a lot of great resources in the federal government right now that we can fund at more appropriate levels, and that we can utilize to try and get answers to some of these questions. Before we build more of a bureaucratic infrastructure, let’s look at the tools that are available to us right now, see how we can leverage those, and make those more effective.
Kevin brought up something interesting about the Silicon Tower. What are things that we can leverage in the U.S. government right now to have fellowships? What models can we use to say, hey, listen, when you’re doing your work in Silicon Valley, which is extremely valuable, can we have a fellowship for you for a year? How can we make that an industry-wide expectation? How can we make that something attractive, where we’ve got that cross-pollination taking place?
So, again, it’s making sure we’re not pitting one against the other, which no one here is doing, but it’s easy for the larger conversation to go that way. Let’s look at where we’ve got commonalities, and let’s see how we can leverage what we’ve already got to make this happen. There’s a big temptation to have a brand-new bureaucracy. Let’s look at what exists and make sure that we can leverage that appropriately.
Ashley Baker: And Kevin, you went into your views on what should be done with CAISI. A little bit more specifically, though, is there a testing regime that you would actually support, or do you think that’s just a contradiction in terms?
Kevin Frazier: I certainly think we need a testing regime, because as Paul noted, and as Neil has highlighted, we’re seeing that these models have ever greater degrees of capabilities, and it’s also worth noting that they have emergent capabilities. So, to the extent that we’re able to identify what we think a model is capable of right before deployment, it may be the case that they introduce new concerns and new risks after they’ve been deployed. So it’s not akin to crash-testing a car, right? We know that we can run a car against the wall however many times, and we can discover its capabilities pretty easily.
Testing is a very expensive and a very complex exercise, and so when folks hear about this latest Hugging Face OpenAI incident, or the reports from Anthropic and from Meta that they’ve also seen some of their testing environments fail, yes, that should be alarming, but it also should be a reflection of the fact that testing is very hard and resource-intensive. And so, if you are a fan of the potential of AI, as I am, to help with things like improving education, improving access to justice, and improving science, then we need reliable mechanisms to make sure people know what they’re dealing with when they adopt these tools and integrate them into sensitive contexts.
So, I do think creating a testing framework is necessary. I think it must be something that’s authorized by Congress, it cannot be this clandestine, secret approach that’s being implemented by the White House right now, and it’s something that does require extensive resources. And I’ll just flag for a second that developing this kind of approach, one that can be trusted, that can be transparent, and that can allow for greater adoption of these tools, is paramount.
We are getting absolutely smoked by China when it comes to adoption of these tools. You look right now, only 3.3% of American households actually have a paid subscription to AI. 3.3%. That is such a staggeringly low number. If we don’t have folks who are using these tools on a regular basis, who are learning their strengths and weaknesses, then we are going to be caught flat-footed in a very interconnected global economy.
Ashley Baker: That’s a really interesting statistic. So I have a follow-up question. If Congress does enact legislation on model testing or any other related issues, and I’m going to first direct this at Paul, because I know this is probably a concern of his, won’t there be a Pandora’s box opened for all types of onerous or restrictive measures on AI and AI development?
Paul Steidler: Well, the short answer is no, because the box has already been opened. States are already putting forward 1,700 bills on AI. There’s been model legislation that’s been enacted by New York, California, and Illinois. There’s a bevy of proposals at the federal level for everything from an AI moratorium from Bernie Sanders and AOC, to proposals to have a kill switch on AI systems, and things along those lines.
So, the reality is, and this builds off of something I heard Kevin say, we already do testing and evaluation on a lot of stuff. You can’t just put a nuclear reactor out there and start operating it someplace. You can’t just put a jet engine into a plane without federal review of this. So, I think one of the things that was unfortunate when the Mythos discussion came up, you had Kevin Hassett of the Council of Economic Advisers talk about how he wanted an FDA approach to AI regulation. And FDA regulation is very long, it’s very drawn out. It also impedes the individual choice that a lot of people make, or feel that they should be able to make. So there are a lot of problems with the FDA and drug evaluation, but we’re talking about things with AI similar to nuclear reactors and jet engines that have a societal risk. It’s one thing for me to take a risk on something; it’s another thing for somebody to take a risk that’s going to wipe out my bank account, or that’s going to pollute my water at home.
I think at the end of the day, though, these issues are common sense. Americans have concerns about AI. I think seeing some type of bipartisan way in which you take steps to have safeguards in place, that really, everybody should be able to agree upon. It shouldn’t be that hard to agree that you’re going to have a 30-day window for mandatory, not voluntary as we have now, a mandatory review of the most sophisticated frontier systems before they go forward. I think that might even give the American people a lot more favorable view of AI, and maybe get that very low statistic of 3.3% subscription up significantly. Because this is a technology that is fundamentally friendly and fundamentally empowering, and can do a great deal to improve people’s lives.
Ashley Baker: So, I want to take a step back to a bit of a broader level for a minute and talk a bit about the institutional, or the constitutional, separation-of-powers concerns. Kevin, you’ve studied how to design institutions that preserve the rule of law as technology changes. You’ve testified before Congress on this. So my question is, is Congress really the right institution to be writing technical testing standards, or should it be setting the terms and leaving the technical work to someone else?
Kevin Frazier: Yeah, it’s a fascinating question, and it gets to the Constitution. What I always love to do, and my students hate me for it, is make them go read the Articles of Confederation, and I make them see the juxtaposition of what Congress was capable of under the Articles and what Congress is capable of under our Constitution. And it’s often forgotten that the language that went to the Committee of Detail back at the Constitutional Convention, and apologies for nerding out, but if I’m going to nerd out, I think here is as good a place as any, the original language that went to the Committee of Detail emphasized that Congress needed to have the authority to legislate in all matters that require competency and that require harmony across the Union.
And when we think about frontier AI governance, when it comes to models that we know already have a billion-dollar influence on our economy, if not a trillion-dollar influence in the near future, that have hundreds of millions of users and billions of users worldwide, this is certainly an instance in which we need to have uniform regulation, such that we are able to have a clear approach to our AI strategy.
Now, as to whether it should be Congress that is sitting over the precise details of what a model should be held accountable to, Congress is not well-suited to do that. The fact of the matter is, this technology is moving so quickly. As I remind my students, today we’re talking about large language models. Well, in San Francisco, they’re talking about things known as world models. They’re talking about neolabs that are small and agile and are testing new kinds of AI. And so the idea that Congress will be able to keep pace with all of the new advances that are being made in AI is just not aligned with the institution’s ability to act.
This is where we need to be creative, and we can look to prior instances in which Congress has realized it doesn’t have the expertise, and it doesn’t have the independence necessary, to develop good policy. So, by way of example, we’ve seen the Rules Enabling Act. Again, sorry to nerd out on everyone. The Rules Enabling Act is how we get our rules of procedure for our court system. Congress doesn’t preside over the nitty-gritty details of who should be served within how many days, via what process, so on and so forth. Instead, they’ve delegated that task to a handful of experts who are subject to their own procedural hurdles, but can move a heck of a lot faster when it comes to developing expert-informed legislation and policy. And so I would love to see something like an REA equivalent for AI, and don’t worry, as every academic has said, I have a paper coming out on that, and I’ll send that out down the road.
But I would also emphasize that states can and should play a role in a limited fashion. Paul mentioned earlier that we have seen legislation, for example, in California, in Connecticut, in New York. And a lot of that legislation has focused on transparency. We saw this with the Hugging Face OpenAI incident, in which OpenAI decided of its own volition, and this is something to be applauded, to reach out to two organizations to perform an audit of that incident. Well, that’s great, but it was voluntary. OpenAI chose who was going to monitor their own behavior, and decided when and how they were going to disclose certain aspects of that incident.
If states are going to say, hey, we need to force these companies to share certain information such that we can make better policy, that’s something that can add to a more robust, data-driven AI policy ecosystem. And so the bifurcation here that I would really stress is that states can and should play a role in helping us better understand this technology. When it comes to shaping how AI frontier systems are going to be deployed, or, excuse me, how they’re going to be developed, that needs to be a matter that’s left to Congress so that we have a uniform standard as to when and how these models are going to be released.
Ashley Baker: So this does unearth a lot of broader questions and a broader debate over federalism and the role of states versus Congress. Neil, you’re a bit of an expert on the role of Congress and federalism. There are a lot of critics of federal AI preemption, actually a surprising number of critics, and they argue that Washington would be preventing states from protecting their own citizens. What would be your response to this?
Neil Siefring: Well, I would say, first of all, that the debate is a really healthy one, because a lot of the debates around our government, how it was created, how it exists, deal with that inherent tension between the federal government and the states. So that’s a very productive conversation to have. It’s not something that should be feared. It’s something that should be defined. When we talk about preemption, it sounds like a very scary phrase. Preemption sounds like it means you’re not going to get something you want. But preemption has already been used by the federal government in a number of different areas: airline prices, routes, and services; ERISA; medical devices; nuclear safety; trucking; food and drug labeling; railroads. So a precedent exists for this.
An AI model’s not developed in California and then stops in Arizona. It’s inherently interstate in its function. So a role for federal preemption, I think, constitutionally, is very clear. President Trump had his national framework on AI, which was outstanding, and we should be working to try and implement that, and there have been steps to try and implement that, to make sure that we’re preempting a patchwork of 50 state laws and 50 different regimes on AI in the United States. We need that for several reasons. One, a small company cannot compete. They do not have a phalanx of lawyers, and they will look at the bureaucracy they’re facing, and they will say, there’s no point, we can’t push this through. So it’s going to stifle innovation in the United States.
Preemption does not mean states play no role. I have to be very, very clear about that. States have an important function in areas like fraud, discrimination, contracts, consumer protection, and traditional police laws. They have every right to, and should, engage in those areas. But we cannot have the United States held hostage on its AI policy by the two poles of California and New York, because those are the states that have the highest legal compliance costs. Companies and law firms up and down the AI stack will say that’s the regime that we have to comply with. How do we know that? We’ve seen that happen in a ton of different areas of the law.
So preemption does not mean laissez-faire. That is a strawman argument that we hear repeated over and over and over again. Preemption means a national AI framework that is built around transparency, knowability, and actionability, that helps companies from the small ones to the large ones compete, and that respects traditional state powers in the AI sphere. So, again, we don’t need to pit one against the other. These two can work in harmony. And it shows you the wisdom of the founders. The same challenges they were looking at at the Constitutional Convention are the very same challenges we’re looking at right now with a technology they could never have imagined. It’s okay and it’s healthy; we shouldn’t fear it, but we have to have the conversation.
Ashley Baker: That’s a good point about the need during the Convention to regulate commerce. But, Neil, a little bit more of a specific question: what would a good federal preemption provision look like? And how do you prevent it from either becoming so weak that it doesn’t really create national certainty on AI development for the companies, or so broad that it unnecessarily displaces state authority?
Neil Siefring: Make it meaningful enough that it provides actual rules of the road and tells companies what can and cannot happen, while giving them the space and the freedom to innovate. Not making it so difficult and so onerous to comply with, but making it something that’s reachable, knowable, and understandable by these small and mid-sized companies. Taking some of those general things, looking at a framework as a statement of principles around what AI should look like and how it should be pursued, that gives Congress the space to fill in those specifics within certain parameters and rules of the road.
Ashley Baker: I think that makes sense. I’m going to turn to Kevin to talk a little bit more about the interstate commerce issue. AI is obviously interstate commerce, and that obviously puts it squarely within Congress’s authority. But having the authority to do so is not the same as designing it well, so to speak. So what does Congress most often get wrong when it preempts, and how would you avoid that here?
Kevin Frazier: Yeah, so on the language that you can use on preemption, there are whole law reviews written on the proper scope of different terms and what they mean in terms of the preemptive effect of that language. So here, as we’ve seen from Congress and from various committees, they’re studying closely how to properly scope the delineation between state and federal authority with respect to this preemption question. As Neil noted, there’s a role for states to play, but we need to make sure we’re specific about that role.
And one of the best delineations I’ve seen so far, one that I’ve written on, one that Matt Perault has written on at a16z, for example, has been this delineation between deployment and development. So, deployment of AI refers to if you are, for example, a hospital that is going to use an AI system in a clinical setting, or if you are a higher education institution, or so on and so forth, and you are going to deploy AI to an end user. That’s a question that states can dictate, and that states can inform based off of their own values and priorities.
So, for example, if California wants to say that before AI is used in a hospital setting, there has to be some degree of patient consent for that, fine. You do you, California, that’s excellent. If in Florida they just want to err on the side of AI being deployed without substantive inquiry from the patient, for example, sure, Florida, you do you. That has no impact on the other states.
The issue is when we move into the territory of development of AI systems. What training data can be used, for example, to train these models? What standards are imposed on the testing and evaluation of these models before they’re deployed? If we have 50 different approaches for that stage of the AI process, well, that’s just infeasible for most labs, really for any lab, to be able to comply with. And so that’s where we need to see the proper delineation in terms of the preemptive language that actually gets set out, where there’s some bifurcation recognizing that states have a very real interest in defending and protecting the values of their communities, while not inhibiting our ability to continue to push the frontier of AI forward.
Obviously, getting that line right is really tricky, and that’s why we’ve been debating it for so long. I’ll tell you something that would make drawing that line easier: more research. Not to feed a fed horse, as my friends at PETA would say, but investing in things like CAISI, the Center for AI Standards and Innovation, investing in expertise, sending money to universities so that we can recruit and retain our professors and not see all of them move to the Silicon Tower. Those are steps we can take such that we can have better policy and better, properly scoped preemptive language.
Ashley Baker: Okay, so I have one more question for Neil, then I’d like to turn back to Paul here. So, Neil, for people who aren’t steeped in AI policy, why should they care whether powerful open-weight AI models remain available?
Neil Siefring: In order to maintain American dominance in AI. That’s the straight-up Onvartist answer, because the open models are the underlying numerical parameters, or the weights, that are available for people to download. They can build on them, they can modify them, they can play with them, they can use them. If you shut that down, researchers are not going to stop on AI. They’re going to continue, and they’re going to look elsewhere, where they can find open-weight models. And where will they look? They will look to China.
China understands the importance of this, because they know that the race for AI is, in part, the race to create the ecosystem that developers and researchers want to work in, will be most productive in, and will be most comfortable in. So the United States has to make sure that we have these open-weight model systems, so that our researchers are building and innovating here in the United States, and making sure that we’re leading in AI.
Ashley Baker: So, Paul, I want to get back to the federalism question a bit, or the national versus state question. Why not just allow the Trump administration to make current policy in consultation with the tech community? And what can Congress add to this, knowledge-wise, policy-wise? Why is this beneficial?
Paul Steidler: Right. Just because something has worked well, or worked relatively well up to a certain point, does not mean it’s the best way to go forward here. One of the questions that arises is, who in the administration is making the policy, or is going to be making the policy? Is it going to be J.D. Vance? Is it going to be Scott Bessent? Is it going to be David Sacks? There are a lot of different people who have different views about AI and AI policy.
So it would be incumbent, I think, on the administration, and I think in its best interests, to say what it wants to see as far as AI model evaluation goes. Because right now, for the administration, if there is an AI incident that comes up, if there’s another Mythos company out there, or similar technology being developed by somebody we don’t know about who does not want to comply with these voluntary administration guidelines, and there’s a mess that takes place, that’s squarely on the administration. They have all the liability for that going forward.
The other thing I would add, and I realize I’m skipping around here, is we don’t even know what’s in the August 4th executive order. Presumably, it’s accepted by the tech community. We haven’t had a crisis with the Mythos stuff, so I guess in that sense it’s good, but we don’t know what we don’t know here. And again, I think the best policies that we have are developed in sunlight, or at least partially in sunlight, with open discussion.
I’m very much reminded in the current debate of the 2008 financial crisis and the TARP situation that took place, where you basically had a crisis that was looming, and a bunch of people in government and the corporate sector came to Congress with a gun at Congress’s head and said, you have to pass this, or we’re going to be in another depression. If we get to that point on AI, where we have to pass something in order to guard against catastrophic stuff, the mistrust in our institutions, if you think it’s bad now, is going to escalate like something fierce. I don’t think we’ve recovered as a country yet from the 2008 debacle in terms of lost trust, in terms of a sense that average people are not having their views or concerns heard about commonsense steps that should be put in place here.
So, again, we don’t know what we don’t know in terms of what the administration has done here. The country will be best served, and Kevin articulated this at the beginning of the discussion better than I can, when laws are made in a way in which they’re known, they’re debated, they’re appreciated. And if we don’t get there, we’re going to eventually careen into another crisis. It might be in five months, it might be in five years, it might be in ten years. And that also has a big problem as far as tech development goes. We don’t know who is holding off doing some AI stuff because of the testing stuff that’s out there, or what decisions and what investments may not take place.
So, I think the administration has a strong hand to play in terms of forging legislation, perhaps even bipartisan legislation, and it would serve the country well to do it. I also think it would serve the administration’s political interests, and that it would also, importantly, promote the advent and use of AI here. It would take some of the stigma away that’s been increasingly coming upon us.
Ashley Baker: So I have one more question for Paul, and then I have a closing question for all three of you. Paul, I know you’ve written a lot about data centers. To what extent is the rising opposition to data centers affecting these debates that you just referenced, and Congress’s whole approach to AI regulation?
Paul Steidler: Yeah, well, it’s not good. In one sense, the two issues are completely separate. The decision to build a data center is a local land use project that should be made by local and county government. You could have circumstances where the federal government gets involved in legislating it, but fundamentally, it’s a community decision. I would also point out, we built 4,500 data centers in this country before controversy ensued on this. They were generally well accepted, generally great members of the community. They signed long-term agreements to be good corporate citizens and pay taxes. That’s changed because you have so many being developed now, they’re bigger, the concerns are more, but there are still going to be many, many places where they’re welcomed.
I think also one of the things that’s compounding the opposition to data centers is a growing concern about AI that is out there. People are fearful of it. They have heard a lot of negative stuff. The Mythos stuff, I think, is something that augurs for general concerns, but we’ve also heard things like, we’re bound for another situation like 2001: A Space Odyssey, the machines are going to take over, everything’s going to end, it’s going to be apocalyptic.
So the issues are generally unrelated, but I think it would help promote data centers to the extent that we can make clear that AI is safe, that steps have been taken to make it even safer, and to accent the vast amount of benefits that come into play for medicine, agriculture, education, tutoring assistance. We are, as a country, just losing sight of these profound, profound opportunities out there over some things that can be managed and can be addressed and should be kept in perspective.
Ashley Baker: Okay, so I have one last question for each of you, and you have less than a minute each. We’ll go Neil, and then Kevin, and then Paul. What is one thing that Congress should do on AI, or let’s say be thinking about on AI, before the end of this session? And what is one thing, maybe, that it shouldn’t do? Neil, I’ll go with you first.
Neil Siefring: In one minute. If they’re able to do it, and I don’t think they are, but if there was some situation in which they could work out what real federal preemption of state and local laws looks like, while at the same time making it extremely clear that states have a role to play in AI in their natural, normal police functions, that would be one of the best things that could happen to try and make sure that that bottleneck doesn’t continue to happen at the state level.
What Congress should not do is create heavily detailed regulatory requirements on models for small, mid, to large-sized companies. And they should fund, I’ll go back to the first part of it, the experts we already have in the federal government who can give us the advice that we need, and can give Congress the advice that they need as they legislate in the next session. So: federal preemption; make sure that we’re boosting the researchers and the advisors we already have throughout the federal government; and don’t create laws that are going to look outdated in 20 minutes, because the tech’s moving too quickly, just because we need to get to the end of the session quick and get something to show that we worked hard on something.
Ashley Baker: Okay, Kevin.
Kevin Frazier: Okay, very quickly, I’ll just echo what we heard earlier: we need to make sure that the techlash doesn’t become a lawlash. The longer we see this status quo of a sort of arbitrary, ad hoc, secret approach to AI governance, the more likely we’re going to see backlash against governing institutions themselves. So first and foremost, Congress needs to invest in expertise, and needs to invest in democratizing access to AI, so fund NAIRR, and make sure that we’re supporting CAISI and supporting our public universities. What Congress should not do is continue to allow, and acquiesce to, the administration ruling by a voluntary framework. Again, this is so antithetical to the rule of law that you don’t need a law degree to realize that this is not the way public policy is meant to be done.
Ashley Baker: Paul, you have the last word.
Paul Steidler: Yeah, so the one thing I’d like to see Congress do is enact the Frontier Act, which has been proposed by Congressman Jay Obernolte and Lori Trahan of Massachusetts. It’s bipartisan legislation. It would increase funding for CAISI. It would put in place mechanisms for identifying catastrophic risk in models and requiring that.
The thing that Congress should not do is anything that Bernie Sanders says. It’s been amazing to me that over the past couple of months, Bernie Sanders, who at the beginning of this year put out a lot of ideas that were just out there as far as the Democratic leadership in the House and Senate was concerned, is now getting traction on AI proposals. They’re not going to pass and become law. But he’s talking about a moratorium on AI development, he’s talking about a moratorium on data centers being developed, and that’s really an unfortunate result of how he’s been able to fill a vacuum, because I think a lot of legitimate issues have not been addressed, or addressed as fully as they should have been.
Ashley Baker: Okay, thank you for that. That is all the time that we have today. Thank you to our panelists. A recording will go up on YouTube later and will be sent to registrants. Have a good afternoon.
Paul Steidler: Thank you, Ashley. Thank you, everyone.
