Trump Signs Mythos-Inspired Executive Order After Delay and Reps. Obernolte and Trahan Release Draft Federal Framework for AI

A

Foreign.

B

Welcome back to the AI Policy Podcast. I'm Alok Mehta, director of the Wadhwani AI Center.

A

And I'm Matt Mand, a researcher with the Center. It's been a big few weeks in AI policy. There's a lot going on, both from the administration and Congress. Today we'll be covering the new executive order from the Trump administration, a bipartisan draft bill on AI development, and recent talk of the government taking stakes in AI. But first, an update on who's leading AI policy in the White House. Sriram Krishnan, a senior policy advisor for AI, has announced that he will be leading the administration this month. Krishnan was a key architect of the administration's AI Action Plan, a national AI policy framework for AI, among other documents. And before joining the White House, he was a general partner at the venture capital firm Andreessen Horowitz, also known as a 16Z. Alok, what would you say is notable about his departure?

B

Well, this isn't the first time we've seen senior AI advisors leave this White House. So in the past, we've seen both Lynn Parker and Dean Ball also depart the administration. This happens appointee jobs, they're very stressful, and so they tend to have some turnover. The Trump administration, at least the first one, was notorious for its high levels of turnover in senior positions. And so we're seeing some a little of that repeat of that pattern in this administration. I think the key thing to note here is that a lot of what the Trump administration wants to do in the AI space involves bringing in and retaining technical talent to be able to execute on a lot of the programs that they're thinking about contemplating implementing. And so the real key here is can they fill their senior ranks with people who have the right level of expertise and can execute on their agenda?

A

Yeah. Well, we'll be keeping an eye out to see who, if anyone, ends up replacing Sriram Krishnan as a senior AI policy advisor in the administration. But moving on to our first main topic of the episode. On June 2, President Trump signed an executive order on cybersecurity risk from AI. The executive order was titled Promoting Advanced Artificial Intelligence, Innovation and Security, and comes after weeks of internal debate in the administration for following the release of Anthropic's highly capable cyber model, Mythos. We covered this debate in our last news roundup. But before we get into the content of the eo, do you mind giving our readers a little refresher on what we talked about last time in the context before Trump ended up signing something?

B

Yeah. So the Release of this executive order is a little bit of a surprise because it had been Nixon is when it was on hiatus for a couple of weeks before it was actually released. So basically, the timeline here is that Anthropic developed this really powerful model called Mythos. It was really capable, particularly in the realm of detecting cyber vulnerabilities. And so Anthropic released it in a limited fashion only to a set of trusted partners. And it really spooked the administration. They were really worried about the implications this might have for the security of government systems. And so that spurred this debate within the administration. At one point in time, the administration was even thinking about something like a regulation, sort of a regulatory model model on the fda. So something akin to licensing. There is definitely a robust internal debate within the administration about how to approach this issue. It led to the planned announcement of the executive order on May 21, and then just hours before the EO was slated to be signed, and there was a ceremony with tech executives and other sort of people slated to go. It was on the calendar and just hours before the President canceled it. Supposedly this was due to a call from David Sachs, the former aizar, who had in turn been contacted by tech executives who had raised concerns about the measure to Trump. And so Trump pulled it down and pulling it down, he mentioned that his concerns were how this might harm the US AI industry's ability to compete with China. So he said, I think specifically, I think it gets in the way of, we're leading China, we're leading everybody, and I don't want to do anything that's going to get in the way of that.

A

Yeah. So you're painting a picture of a pretty volatile few weeks. And anyone who wants to get a more in depth overview of those few weeks should tune into our previous news roundup, where we go more in depth. But what actually ended up in the executive order that Trump signed, and he did so, I think pretty discreetly, without any major publicity, which is pretty unusual, I think, for his administration.

B

Yeah. So there are two main elements to this executive order. The first focuses on shoring up federal and critical infrastructure defense. So it basically directs a bunch of different parts of the government. So the Department of Defense, or the Department of War, the Department of Homeland Security, and other parts of the government to bolster the security of federal IT systems and critical infrastructure. It also creates a AI cybersecurity clearinghouse organized by the Department of the treasury that brings together the AI industry and operators of critical infrastructure to coordinate on scanning and patching software, VULNERABILITIES the other part of the executive order really focuses on a voluntary early access framework for government. So this is a framework in which AI developers could provide the federal government with access to powerful models for up to 30 days before they release it to other trusted partners. Alongside this, the EO mandated development of a benchmarking process to assess the advanced cyber capabilities of AI models and determine which models should be subject to this framework in the first place. And the most notable thing about the version that came out versus some of the drafts we'd seen of earlier versions is that the advanced access period dropped from 90 days to 30 days. So this is, I think, a reflection of some of the concerns that tech executives had about the length of the review period. Even though it's voluntary.

A

Yeah, and that's a substantial change. But other than that, as you mentioned, there wasn't really much else major that changed between the drafts. I think as a result, some people have come out and said that for accelerationists, people who want fewer regulations and guardrails on the development of these tools, that this executive order is a loss and it's more of a win for those within the administration at least pushing for more guardrails. But on June 5, you and your colleague Lauren Williams published an article arguing that the EO really does represent a continuation of the administration's laissez faire approach to AI regulation and fails to address public concerns. Could you walk us through your reasoning here?

B

Yeah. So I think what is important to note here is that most of the major AI developers, so that includes Google, DeepMind, Microsoft, Xai, OpenAI, and Anthropic. Prior to the release of this EO, they'd already entered into agreements with the center on AI Standards and Innovation, which is part of the Department of Commerce, for advanced testing of their models for national security risks. And so, in large part, this involuntary framework is A, it's voluntary and B, it is largely an reflection of existing industry practices, things that the industry is already doing. There's some changes in sort of how the interagency manages this and what parts of the government are involved, but it's not a substantial shift from what is already happening. And so this EO is kind of reflecting what's already happening. It's not pulling the industry in a new direction. And the issue here is that if you look at public polling, the American public is really skeptical about the benefits of AI. They're concerned about the impacts of AI on their jobs, their local communities, on the job market as a whole. And they really do want the federal government to step in and regulate in a more robust way. And I don't think that this EO really answers that kind of call. It's not really meeting the level of interest and regulation from the American public.

A

Yeah. And I want to go back to something you said about which parts of the federal government are carrying out what's within the executive order. There's one specific line from the piece that you and Lauren wrote that stood out to me, which is agencies like the Office of the National Cyber Director and the Department of Defense and Homeland Security, where much of the government's cyber policy and technical expertise resides, are placed in consulting roles. Why do you think that is?

B

You know, I. I think this whole EO can be read as sort of a set of compromises between different parts of the administration. So the administration is not a unitary actor. It doesn't have a single view on this issue, and it's. It has multiple views. And so the various pieces we see in the executive order reflect the internal debates we see in the administration. So some of those have to do with concerns about the impact of the executive order slowing down the industry. Others have to do with what parts of the administration will be in charge of certain elements. And that debate is often specifically about should there be more centralized White House control and more control by sort of political appointees, or should we rely more on technical expertise from experts, often experts in the civil service? And so some of what you see in the executive order is a decision to prioritize sort of parts of the White House and other political decision making over technical experts in the government.

A

Interesting. Well, shifting from the internal debates within the administration to the debate happening externally with experts in the AI policy space, what have been some of the reactions to the executive order?

B

We've seen mixed reactions. I think that some people are skeptical that the CEO will change anything. Some people want the more transparency into the. Into models. Right. So they, they want us to be able to see what these models are capable of, not to have the government sort of have early expertise and conduct testing in secret. And others who think that the wrong parts of the government are, are in charge here and that we should really have technical organizations within the government like, like Casey, the center I mentioned earlier, approach us from a technical perspective and really try to focus on how to accurately test and evaluate models and then provide feedback to companies about how to design appropriate safeguards to mitigate any issues that come up. I think one interesting point, which is something that I've said before, is that there's often an argument that you can't regulate and have innovation at the same time. And we see. Matt Shaheen, who's a senior fellow at the Carnegie Endowment, wrote that that's not really the case. He says for the past four years, China has had the world's most extensive and burdensome AI regulations. And during the same time, China has largely caught up with the United States and AI technology. I'm not sure that it's completely caught up with the US in terms of developing the most powerful models, but it has made some pretty significant inroads. And I think that what, what we see is that particularly if you worry about adoption in the US and you think that's key to US Success, then you have to convince the public. And I think regulation is part of the answer here. If you regulate AI technology in the right way, in a very considered way, you can help bridge that trust gap, and then you can drive adoption. And that is also something that facilitates AI innovation, because as more people adopt technology, you have more money to invest in it, and so you create this sort of positive flywheel.

A

Yeah, well, I think our next topic, which is the new federal framework that two congresspeople proposed, will sort of get into that space. But one last thing. The Executive Order wasn't the only major White House AI release. Last week, the administration also released its National Security Presidential Memorandum on AI and the National Security Enterprise. We're not going to go as in depth on this document, but. But what is it trying to do and what are some of its major elements?

B

Yeah, it really is focused on accelerating AI adoption across the military and intelligence functions of the US Government and trying to bring in more technical talent to support those efforts. I think there are a couple of interesting things about this National Security Memorandum. So one is that it explicitly talks about adopting open source AI. This is something that was mentioned in the AI Action Plan, but we haven't seen a lot of concrete action from the government since. And so I think that's pretty interesting. But the other thing is that it's really focused on a diverse set of models and having access to models from a bunch of different providers. And I think this is a recognition that in the past, the government has often become reliant on a single provider for critical technology, and that's caused various issues. And so they're aware of this. And so the memorandum tries to address that issue.

A

Yeah, absolutely. Right. Well, let's move on to the federal AI framework that was recently proposed on June 4th. Representatives Jay Obernolte, who is a Republican from California and Laurie Trahan, who is a Democrat from Massachusetts, released together a discussion draft of what they're calling the Great American AI act, which is a 269 page national framework for AI. Could you give us an overview of what the draft contains and some of its most important provisions?

B

Yeah, it's a pretty ambitious document. So it has lots of different components in it. I don't think we have time to talk about all of those. But it does have major sections on AI research and development and international cooperation, on cybersecurity, on workforce, which touches on things like AI education, collecting data on the labor market, exploring ways to protect workers. But I think a lot of the meat and a lot of where we're seeing attention focus is on the first part of the bill, which is around frontier AI governance. And so within that you see things like formally establishing the Casey within the Department of Commerce and providing it, authorizing a pretty significant budget for it. You see mandated transparency requirements for frontier developers, specifically that they have to implement and publicly post a frontier AI framework that covers risk thresholds, assessment procedures for catastrophic risk, how they protect model weights, and some transparency into deployment decisions. It also touches on the issue of IBOs or independent verification organizations that can help audit how large frontier developers comply with various parts of the bill. There are parts about whistleblower protections and then something I'm sure we'll dive into in a second. A three year preemption of state laws that regulate AI development.

A

Right, that's one of the most controversial parts.

B

Yeah, definitely.

A

Yeah. So speaking of which, there's been a lot of discourse among AI policy experts following the release of this discussion draft. What are some of the main points of agreement and disagreement that you're seeing?

B

So I think one of the major points here is that if you look at how this discussion, the discussion around frontier AI regulation has developed over time, this idea of frontier AI companies having some sort of framework laying out how they assess risk and how they make deployment decisions, a kind of safety framework or frontier framework has been sort of a point of emerging consensus for a while. Almost all the frontier AI labs do this voluntary, voluntarily. I think it provides very useful information to the public. But some of the issues we run into are that each of these frameworks are done sort of idiosyncratically by each company. And so it's often difficult to compare across them. They sometimes have different definitions of risks, they're written in different ways. Some of the sections don't line up, some of the level of details don't line up. And so Having the government step in and provide some scaffolding for that, I think is really useful if you want to engage in sort of cross company comparisons. And so I think that this bill does a lot of good things in terms of looking at that emerging industry consensus and taking steps to codify it in a way that makes sense. The transparency protections building out. Casey, I think Kasey is done a lot of good work and finding ways to bulk up the organization, allow it to hire good people and do more technical research, I think is the right way to go in terms of taking a technically minded approach to these very technical AI governance questions. So I do think it does a lot of things correctly.

A

Yeah. And I think, I mean, those are the points of agreement. I think many of the points of disagreement center around whether those provisions are enough to justify preempting state AI laws that regulate AI development for three years. And the text is pretty specific. It says that the law would preempt any state or local law or regulation specifically targeting the development of AI models. And then a caveat expressly does not preempt laws of general applicability, common law remedies, or laws regulating AI user deployment and sunset's three years post enactment. But AI like laws, state laws targeting the development of AI models is a broad category of state laws. And so what have you seen in terms of how people are thinking about whether what is in the bill for transparency and governance justifies three years of state preemption?

B

I think in this case the safest thing to do is to preempt narrowly in the areas that the bill directly addresses. So in this case you would, you would, you could envision a preemption that's just, you know, we are mandating this kind of frontier safety transparency around model development. And so states can't implement their own laws that touch in that specific area. But I think we are seeing some concern that the preemption here is broader. It touches on every aspect of model development. And so it touches in on areas that are not directly addressed in, in the bill that's been put out. So, you know, some commentators have talked specifically about children's safety. So there's very little on children's safety specifically in this bill. But you could read the preemption as saying that a state couldn't pass a bill related to model development and children's safety. And so that that is a concern. And in the past when people have discussed preemption, they've often explicitly carved out areas like children's safety and said the states can still regulate in that area because the politics of opposing children's safety is very bad.

A

I think I saw some pretty hostile Twitter threads touching on this subject. I won't name any names, but I'm sure other people have seen similar discourse around that. And one other thing I want to bring into this is OpenAI's blueprint for a federal framework. I think they had. They released this on June 2nd and had a lot of overlap with what ended up in this national framework, but also diverged on some components, particularly preemption. So I'm curious if you could just talk a little bit about where that overlap was and where the divergence was.

B

Yeah. So, you know, if you look at their blueprint includes many of the similar elements around frontier transparency that we've discussed. So things like independent assessment and auditing, whistleblower protections, authorizing and bolstering. Casey. And that, I think, is because, like I said, this is a point of emerging industry consensus. And so not only OpenAI, but I think a lot of labs would agree on these types of points. I think the difference is that OpenAI's blueprint in specific reads that policymakers should preempt state laws that seek to regulate the same frontier safety risks. And so I interpret that as them being in favor of a more narrow preemption versus this broader preemption. And possibly that could be because that is safer and less controversial position to take when it comes to preemption.

A

Yeah. Well, the bill is currently just a discussion draft, as I've said, meaning its sponsors are looking for feedback, and its text may ultimately change before it's introduced and certainly before it would pass. But as it stands, what does the bill's future look like?

B

I mean, the fact is that we're about to enter midterm elections in the U.S. and so we have a few more weeks before campaigning is going to take up a lot of oxygen, especially on the House side, but. But for a significant part of the Senate. And so there's a very narrow window for anything significant to happen on Capitol Hill. And this bill is too big and too broad to see significant action in that time. So I think it's good that they've put this bill out and they're gathering comment. But, you know, in terms of prognosis, it's unlikely to see anything serious happen this year. And I think we've also seen that there are concerns about some of the specific content of the bill. Some of those concerns, say, on the Democratic side, come from concern about the preemption being too broad or political concerns about not engaging in certain kinds of bipartisan agreements because it might hand the other party a win right before the terms. And so I think what we're going to see is that this, this debate will, or this bill will spur a lot of debate. We'll get lots of interesting, they'll receive lots of interesting feedback, and that will likely lead to some sort of revised bill that comes out maybe early next year.

A

Sure. And I have to mention that we actually had Congressman Ober Nolte on the AI Policy podcast last October. So anyone who's wanting to learn more about his background or his views on AI governance, I definitely recommend going back and checking out that episode. But I want to move us on to our last topic of today, which is that last Friday, Trump told reporters that he was planning to meet with AI industry executives to discuss the possibility of the US Government acquiring shares of their companies, which was reportedly news to many industry leaders. So what exactly did Trump share about the administration's plans?

B

Not a lot of specifics. So mostly we have a sense of the broad idea, which is that the, the president seems to want the government to have some sort of equity stake in these companies. We've seen this before. Right. He's come to arrangements with other types of companies. Yeah. Where there's some sort of revenue sharing arrangement or some equity piece that the government gets in return for investing federal dollars in companies. In this case, again, lots of details lacking, but we've heard or seen some reporting that this might revolve around having firms voluntarily provide shares to the government. We've seen discussion that some sort of model for this could be Alaska sovereign wealth fund. So the way that works is basically Alaska receives some portion of oil revenues from activities in the state and then is able to redistribute those to people in the state. And so this is not a lot of money, I think, but it is significant and it is a way in which we see the profits from a specific industry being redistributed to residents of that particular state.

A

Yeah. Well, it's interesting because I think Sam Altman of OpenAI originally pitched the idea to Trump last year. Right.

B

Yeah. I mean, you know, OpenAI has put out some, a few papers that sort of talk about this idea. There's one in April 2026 talking about the creation of a public wealth fund. So I think some of this is predicated on the idea that AI is going to be super transformative, but also extremely disruptive and that we need to figure out a way to prevent some of the concentration of wealth that might accumulate, if that's what happens. And Sam Altman and OpenAI, they've both been long interested in this idea of universal basic income. And I think either OpenAI or SAM directly has funded research into UBI and how that might work and what the effects are. And so this isn't something completely new. There's definitely been interest in this type of idea before. Other AI companies, I think anthropic and particular, have also talked about the importance of sharing some of the economic wealth, particularly if you see a lot of concentration happening because of developments in the AI sector.

A

Well, one other person talking about this is Bernie Sanders. In fact, Trump's comments came just days after Sanders in an op ed published in the New York Times called for the creation of an AI sovereign wealth fund through a 50% tax on American AI companies. Can you tell us a little bit about what Sanders vision for redistributing AI wealth looks like and what differences it might have from the future that the Trump administration might envision?

B

Yeah, so what Bernie Sanders envisions is very different from what Trump envisions. And the primary difference here is that. That's a surprise. Yeah, definitely not a surprise. The big difference here is that what a Sanders wants is something that is mandatory, not voluntary. And so he wrote in an op ed in the New York Times, this legislation would give the public a direct ownership stake in the largest AI companies in our country. How it would create a sovereign wealth front through a one time 50% tax, not on the profits of these companies, but copaid with something far more valuable than that, the stock. And so he envisions a sovereign wealth fund which is basically sort of a government run investment fund that is used to benefit the country's citizens. We've seen this, we've seen other countries do this. So Norway, Kuwait and Saudi Arabia, they all have these types of wealth funds. They make investments and the returns of provide a benefit to the, to the government that runs the wealth funds. Oftentimes those returns are used to fund government programs or other types of benefits for the residents of those countries.

A

Yeah, and I think that's an important distinction that the returns go to the government and not directly to people. And so that's notable. I mean there's, there's ways in which government can run programs that benefits benefit people with the money they make, but that's not necessarily what ends up happening.

B

Well, it could, I mean there are different ways to run it. Sometimes, sometimes a sovereign wealth fund might, might be used to fund government programs, but other Times, you know, like if you look at the Alaska example we talked about earlier, you could distribute those directly to citizens. And if that's what happens, then you're essentially. Then what essentially you're creating is a form of universal basic income. So kind of like the discussion has come full circle. It's the idea of taking stock in these companies and then returning the profits to people and essentially supplementing their income through the AI trade. And so a little bit of this is essentially the infrastructure for creating universal basic income, potentially.

A

Yeah. Well, both of the proposals have received a fair amount of criticism. Not everyone is a fan of this sort of industrial policy. What are some people saying in response to both Senator Sanders as well as President Trump's proposals?

B

Yeah, I've mostly seen negative commentary here and I think a lot of this revolves around the fact that this type of proposal is just, it goes in the opposite direction as the kinds of industrial policies, trends we've seen over the past century, which is a movement away from public ownership in key sectors and, and towards privatization of industries. And this is happening all around the world. The UK did this, the US has done this where the preference is really for the free market to manage a lot of our critical industries, including things like some of our utilities, our telecommunications infrastructure, et cetera. And so the reason that we move in that direction is, are some of the same reasons people cite when they talk about why they don't like this idea. So it's the idea that governments can't run organizations well because they don't have the same profit seeking motives, they don't run things efficiently and this leads to wealth destruction, not creation. Then you see other kinds of criticism around whether this is even legal. Would it violate the fifth Amendment's prohibition on the government taking property without just compensation? Other people who say there's no reason for the government to do this. We have a stock market. Anthropic and OpenAI have both filed for IPOs. And so it means it's possible that we will see both of them go public this year. And so if you're really want to make a bet on how transformative AI will be on the economy, you can just go and purchase that stock. So why have the government do it when you can just do it directly?

A

Yeah, and I think like that point is interesting to me. I guess what I hear people who support this sort of industrial policy saying is that, well, not everyone can go and purchase stocks in these companies because they just simply don't have the capital.

B

I mean, I think that's fair. One of the things we hear over over again is that over time our economy has been become more concentrated, that the average worker is receiving less and less a share of the growth of the economy. And there's a concern that AI will supercharge this. And so if the idea is that you're going to rely on people buying stocks, well, your ability to buy stocks is is dependent on your ability to have capital to buy those stocks. And our economy has moved away from that direction in large part for many people in this country. And so politically it just might not be very satisfying to tell people you don't need to worry about it, just buy some shares and open anthropic and you will also share in this wealth.

A

Yeah, no, I think that makes sense. Well, I think that does it for this week's AI Policy Podcast. Thank you Alok for your insights and thanks to our audience for tuning in.

B

Thanks thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform. Subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker and Matt Mand. See you next.