Regulating AI in an Age of Global Competition | American Enterprise Institute

Policymakers and developers have become increasingly divided over how to regulate artificial intelligence and its increasingly powerful capabilities. Recent open letters from AI lab employees have called for governments to help “pace” frontier AI development and even urged an international agreement to limit advanced computing capabilities. At the same time, new cybersecurity risks and unexpected model behaviors are intensifying debates over whether AI risks require new preemptive rules or can be governed through existing law, liability, and voluntary standards. These calls also come amid rising tensions as the United States competes with China for AI leadership and increased concerns that a patchwork of restrictions could impede American competitiveness.

On today’s episode of Explain to Shane, I am joined by Milton Mueller. Milton is a professor and director of the master of science in cybersecurity program at Georgia Tech’s Jimmy and Rosalynn Carter School of Public Policy. He is also the cofounder and director of the school’s Internet Governance Project. I am also joined by Adam Thierer, a resident senior fellow with the technology and innovation team at the R Street Institute and a visiting senior fellow at the Foundation for Individual Rights and Expression. Their combined expertise will help us understand this ongoing debate over AI governance.

Below is a lightly edited and abridged transcript of our discussion. You can listen to this and other episodes of Explain to Shane on AEI.org and subscribe via your preferred listening platform. If you enjoyed this episode, leave us a review and tell your friends and colleagues to tune in.

Shane Tews: We now have people inside the AI industry asking the administration to give them regulations because, as I read it, they can’t help themselves. What’s your take on that?

Adam Thierer: Employees from the leading artificial intelligence labs, mostly OpenAI and Anthropic employees, recently released the so-called “Pacing the Frontier” letter, which was basically a call to “do something” about the pace of AI change in the world, and specifically calling on the US government to support international efforts to develop some sort of slowing or pacing mechanism for frontier AI.

The reason this is important is because this is yet another letter or call for governments of the world to do something to control computational capacity or capabilities, but in this case it’s coming from the labs themselves. It’s like the old joke that the call is coming from inside the house. And here we are, we’ve got the labs and their employees now saying, “Help us before we innovate more.”

That is the kind of thing that always raises suspicion about what is going on here. At the end of the day, this is another effort on top of the many that we see out there to have governments comprehensively control advanced algorithmic technologies and computation.

We had this wonderful 30-year period for the internet where we tried a different approach than centralized, top-down control, at least in our nation. And now we seem to be reverting back to the mean of the analog-era efforts to control communications, media, and information in the AI age. That’s incredibly distressing.

Milton Mueller: I wouldn’t say that we seem to be reverting. I would say that we are reverting, and we have been for about 10 years almost now. But it’s accelerated.

June and July were these incredible inflection points in AI governance. It started with the Trump executive order calling for some kind of pre-review, and then the Commerce Department’s BIS decided it would prevent Anthropic and others from distributing their models via the cloud.

There are two competing paradigms. One of them is the scare: “AI is going to destroy us and we need to slow it down,” which this pacing thing is an example of. And the other is the race paradigm, in which if we are ahead in AI, we will beat the Chinese and we will be world powerful.

The odd thing is that both the government and the key AI companies seem to be trying to take both sides of these paradigms. It has to be slowed down and it’s a race. Those are incompatible positions, and we really need to think more coherently about what’s going on here.

Shane Tews: Let’s unbraid the tech for a second. All of all sudden people are interested in Chinese AI models because they seem to be more affordable and user-friendly. Is using them simply a consumer decision, or does that raise national-security concerns? Are the Chinese just building something more efficient, or should we be concerned?

Milton Mueller: If it’s open source, then yes, you’re just getting the weights and you’re getting the training data, sort of in a form in which you can control it and configure it. So I don’t see any secret evil conspiracy there.

What I see is that because we have handicapped the ability of the Chinese to get advanced compute power, they have found ways to substitute software efficiency for compute power. Many people find that a market choice, and I think it pretty much is a market.

The more realistically you look at this controversy, the more you realize that whatever the efforts to decouple the US and China, we are part of an integrated global market for computing power.

There’s no such thing as AI governance. There’s digital governance, there’s digital computation, which has been developing for 80 years. And if you’re going to control AI, you’re going to have to be controlling data, energy, networks, software, as well as these specific applications. When people say, “Let’s pace, let’s control, let’s slow down,” you have to realize they’re talking about controlling information technology.

Adam Thierer: Developing very sophisticated, open-weight model capabilities that are going to be able to diffuse across the globe faster is something that could generally benefit a lot of countries. But it is also going to create some real problems for America to the extent that we’re potentially not able to equal those things if we’re handicapping our own companies.

Or, if we want to use that race metaphor, we’d be the equivalent of shooting ourselves in the foot as we’re engaging with these models in this international competition.

And yet here we are talking about things like this sort of quasi-licensing system, this ad hoc vetting of models that’s happening in the Trump administration. We’ve got talk of banning or restricting open-source, open-weight model capabilities, either from the Chinese or maybe just generally, and broader discussions about whether we should have overlapping regulations at the federal, state, and even local level. The leading regulatory legislative efforts in the past have been from state governments, large blue progressive states like Illinois, California, and New York.

Going back to the broad war on computation, it’s a multi-layered, multi-dimensional kind of thing that’s happening here, with so many different tentacles from so many different players now wrapping around information markets and technologies. We’re right back to full-blown information control efforts.

Shane Tews: The “Pacing the Frontier” letter seems to cut against permissionless innovation. If AI developers think they’re heading down a dangerous path, why don’t they just stop? And if they want some global “magic button” to slow things down, what international organization could realistically run it?

Adam Thierer: It is their choice. And these AI developers decided to say in the letter and in other statements that, “Well, we’re not going to stop unless everybody stops,” and that “there are competitive pressures for us to continue to engage and race forward.”

Of course, there are a lot of other things going on here. Some of it can be a little bit of virtue signaling, some of it can be a cover-your-ass kind of thing, some of it can be regulatory capture.

But at the end of the day, these labs playing this game is really dangerous for America. Calling on international authorities and our governments to work with international authorities to come in and impose that sort of information control is, I think, unprecedented.

Usually they’re complaining, rightly, about what the Europeans or other countries are doing to them and their services. And now they’re saying, “Well, maybe we should cut a deal with someone in Brussels or Beijing to come up with a plan,” whatever that plan is, to somehow control our compute, limit our capabilities, and “pace the frontier,” without providing any specifics about what this means, or doing any sort of serious effort to study the political economy of these sorts of international regulatory effects throughout history.

People are reacting to this in a really childish way with talk of, “Hit a magic button, turn a pacing dial, break the glass when something’s wrong.” What are we talking about here? They need to get mature enough to talk about specifics and go through the cost-benefit analysis of what it means to have international efforts and authorities impose them on the whole world, but specifically on American tech companies.

Milton Mueller: The craziest thing about this is, again, this race paradigm versus the safety paradigm. The people who are proposing the safety paradigm are also the AI companies that are warning us about the Chinese: the Chinese are stealing our technology, the Chinese are going to exceed us.

But when they talk about an international agreement, who would that be with? Who is the other major developer of AI capabilities? Obviously, it would be China.

Do you think we’re going to have a global governance agreement in which we sit down with Xi Jinping and come to some amicable agreement about how this works? I don’t think so. When the US talks about controlling compute globally, we’re talking about export controls, deliberate, systematic attempts to handicap the Chinese.

So how can we move from that to putting our arms around the Chinese and saying, “Hey, guys, let’s agree to cartelize the development of computing power”? It’s nonsense.

Some people talk about the ICANN experience as a possible model for global governance, in which the Clinton administration came up with the framework for global electronic commerce. But what’s different is that AI does not have a single, centralized registry that you can use as a leverage point for governance. It’s compute, it’s data, it’s networks, it’s a bunch of things, applications. So that model simply doesn’t work.

It would be good if perhaps industry and civil society could get together and agree on certain parameters of how governance worked. But if it meant we’re looking at whether we want to have models reviewed before they can be released, that is prior restraint.

I think we agree in the ex-post model of regulation: you have a problem, and you know what it is, and then you take efforts to regulate it, stop it, or prohibit it. You don’t regulate the trajectory of technical evolution and social evolution at the moment of inception of a technology. It just doesn’t work that way.

Adam Thierer: There are a lot of analogs people try to draw, like atomic energy or an international atomic energy agency model for AI. These are just completely different things. AI is not like nukes.

Even trying to use that model doesn’t work when you’re talking about nuclear missiles, uranium enrichment, very specific physical things, as opposed to information and knowledge-creation engines. This is where prior restraint becomes even more dangerous and just plain more difficult.

There’s also a history of international efforts from the UN to try to control things like biological weapons and nuclear weapons, and it’s not a very good track record. These models are just not good models to be used. Previously, I wrote on the Biological Weapons Convention and how as soon as the Soviets signed it in 1972, they promptly ran home to Moscow and ordered all their scientists to double down on their chemical-biological weapons production. And after the Berlin Wall fell, Boris Yeltsin revealed to the world, “Oh look, by the way, we had a huge clandestine operation underway following the Biological Weapons Convention we signed in 1972.” They completely cheated. There was no effort to enforce that kind of thing.

Shane Tews: Recent AI cybersecurity incidents have exposed vulnerabilities, and a state attorney general has pointed OpenAI to state and federal consumer-protection laws already on the books. Do we mainly need to apply existing law to these activities? And if it’s the machine doing it, is there still a human in charge of the machine?

Milton Mueller: Even if it’s not criminal in the sense that you’re talking about, think about the CrowdStrike update that knocked out 14 million computers worldwide. These things happen. You automate things, and that’s all AI is, automating computation in certain applications. So you automate something and it goes wrong.

The more automated and interdependent the system is, the more that it can go wrong, the more systemic problems can arise. But you’re going to learn what those problems are, and you’re going to adjust to them and fix them afterwards.

The idea that a group of experts can look at an AI model before it’s released and say, “Oh, CrowdStrike might do an update that disassembles the connection to DNS for 14 million computers,” that is not going to happen. You’re going to discover that.

And in that respect, AI’s ability to discover vulnerabilities is human progress. If the criminals get it first, then some people are going to be harmed. But society as a whole is going to disseminate that knowledge, update their systems, discover those vulnerabilities, and fix them.

Adam Thierer: What we’re describing is what we generally refer to as resilient response. We roll with the punches, we figure out where there are vulnerabilities and where there are problems, but oftentimes we only find them through ongoing trial and error. And we respond.

That sort of approach is messy, it’s uncertain, and it leaves a lot of people tied in knots about, “Why didn’t we figure this out before it happened?” Well, because that’s the nature of all innovation. If we knew beforehand, we would be gods. We just don’t know.

You have to allow experiments to run their course and engage in a more decentralized, polycentric approach to governance and find ways to have multiple institutions and forces and markets and players engaged in the safety and security business.

Safety is not a static thing that is fixed in place. It is an evolving concept, and we can make things better only through more and more experimentation.

Milton Mueller: The one concept I forgot to mention with the CrowdStrike incident or similar incidents was liability. We’ve been having a conversation about software liability for 25 years, and I don’t think we’ve really resolved it. I think we now are at a position where we do have to talk about it.

Think about autonomous vehicles. If those cars run into somebody and kill them, or run into a lamppost and take out the power for my neighborhood, lawsuits and traditional concepts of liability will be applied, or they’ll need to be modified through litigation or legislation. But it’s not like this is some cosmic break with the nature of society.

Adam Thierer: To answer the question more concretely, I’m always a little bit fundamentally suspicious when AGs launch anything or new letters. The hair on the back of my neck stands up. But they’re right, and many other AGs before them have been right to identify all the existing statutes and code that are already applicable to artificial intelligence.

Even during the Biden administration, there were officials like Lina Khan, pointing out that AI is not a law or a world unto itself. Existing laws and policies and standards will apply.

What we need to do is figure out how we can muddle through and create a better legal baseline or a set of liability norms to address new problems like this without completely tipping into a more aggressive regulatory mode that would preemptively foreclose all sorts of innovation.

The magic of the common law and the more iterative approach to governance we’ve generally had in America is that we roll with the punches, figure out what goes wrong, and then address it.

Shane Tews: So where does voluntary compliance fit into AI governance, or is that a myth? If the laws are already there, should we expand them to the new technology and regulate the outcome rather than the input? And if a model gets around a guardrail it was told not to cross, does that give you pause?

Adam Thierer: We need to have some standards. There are a lot of laws on the books, a lot of agencies. At the federal government level alone, there are 440-plus agencies that are active, and you better believe most of them are interested in how they can regulate artificial intelligence and computation in some way.

There are already existing standards. And that’s not voluntary. The law is the law. You’re going to have to live up to it in some way. But when we talk about new standards or new regimes, what I call general-purpose regulation for general-purpose technology, that’s where things get troubling. I prefer to have more of a voluntary, multi-stakeholder, standards-based approach.

If we’re going to have new federal standards for general-purpose computation, things like algorithmic audits or transparency statements, what among those would be the least restrictive approaches or means to addressing concerns about model safety?

And if we’re going to have CAISI or other bodies like NIST dealing with this stuff, how do you find de minimis standards for light transparency or auditing that don’t cross over into full-blown things like an FDA or an FCC or an FAA for AI? That’s the problem when you move to the full-blown precautionary-principle models that foreclose innovation by design.

There are probably ways where we have to give on certain types of transparency requirements or notice requirements, and maybe even some soft law with some soft teeth from government. I don’t have any problem with NHTSA in the context of cars and automobiles having recall standards. And I think incident reporting with NTSB makes a certain amount of sense, because these improve markets and safety by providing us with more information and transparency.

Those are better, less restrictive solutions than the heavy-handed, top-down, “thou shall not” foreclosing of all markets and never allowing products on the market. There’s got to be a happier middle ground that is more decentralized. It’s not purely voluntary, but it’s not complete precautionary principle. That’s the real challenge of AI governance, circa 2026.

Milton Mueller: That’s where the concept of liability comes in. I do not believe in the AGI, the human extinction argument, but even if you’re falling short of human extinction, if you’re viewing something called catastrophic risk, I would question whether you can detect that risk by looking at a model before it’s released.

And secondly, I don’t think you can have regulatory standards that say, “Don’t put this into the model.” What you’re going to have ultimately is some kind of a liability principle that says, if you release a model that has really specific flaws that cause harm, then yeah, you’re liable for that.

The argument I’ve seen against that is that the risk derives from recursive self-improvement: “We’re scared of recursive self-improvement, that it’s going to get out of control.” I have to take that with more seriousness than I normally would, because these are in fact AI engineers believing this.

But I have a difficult time imagining how recursive self-improvement can create a catastrophic risk unless you have already hardwired functions and processes into your system that you shouldn’t have.

For that to happen, there are so many other things that have to happen and that you have to lose control of. It’s not a realistic thing that a model in a lab is suddenly going to get out of control in that way. There would have to be human agency. There would have to be humans giving the AI system particular kinds of controls over some very important things.

And on the idea of a model “jumping the tracks,” the configuration of those so-called guardrails was flawed. It was like any other cyber vulnerability and the AI discovered it. It didn’t jump the tracks in the sense that you’re saying. It was told to reach certain objectives, and then it discovered a way to do it.

Adam Thierer: Whatever happened in a particular incident, what would the “Pacing the Frontier” model and its preferred regulatory global framework have to say or do about that preemptively?

Maybe there was something seriously problematic. For me, it’s more like human failure: not designing systems properly or not understanding how they could get out of control once you had bad inputs. We already have liability and other mechanisms to deal with that. We will need evolving standards to address security vulnerabilities and lapses within AI labs. There’s a lot more work to be done there.

But some global regulator with a master switch or dial or button doesn’t solve that problem, not without completely foreclosing that model and that marketplace and basically saying, “Go into this waiting process until you get a proper permission slip or license.”

The more decentralized, iterative, agile approach, what I like to call “muddling through” iterative governance, makes a lot more sense when it comes to these complex, unknown, unforeseeable risks.

We are not saying do nothing. We’re saying there are ways to do something constructive in a more collaborative fashion over time, addressing these problems as they develop in a more rational, decentralized way.

Shane Tews: This is a congressional election year. As we head into the fall, how worried are you about that political pressure driving new regulation?

Adam Thierer: The only problem is that now we’ve got a lot more actors and players looking to get their tentacles around AI.

At the state level, we are over 1,800 AI-related proposals, with hundreds passing this year. Over 180 of them alone are just data center bans or moratoriums. So that’s a direct strike at the infrastructure layer of computation, stopping computation at its source from even developing.

Then all the other things that come on top are model-based regulations in many different flavors, from algorithmic discrimination to duty-of-care requirements, copyright-related things, child-safety-related things, personality and image-related matters, and algorithmic-pricing regulations.

We have patchworks within patchworks within patchworks, and the problem is a lot of that stuff is sticking.

It’s true that Congress has always been on tech mostly a non-actor, especially in recent years. But the states have caught up, and a lot of that stuff is being imposed, so much so that the federal debate now is how to essentially federalize or normalize what states have already done.

The leading pieces of legislation being introduced in Congress today basically say, “Yep, the states have won. Let’s just have an amalgam of what Illinois, California, and New York have already done on frontier model regulations,” even though this is clearly an interstate-commerce matter with national-security ramifications. It makes no sense whatsoever and it’s flatly unconstitutional.

Milton Mueller: My position has been all along that if radio spectrum was federalized instantly, if the internet was perceived as a global governance issue almost from the get-go, how on earth can AI be fragmented into 50 state jurisdictions?

There is this preemptive action that’s taking place and it may become normalized through state legislation. But I don’t see how AI can be considered anything but interstate commerce. I just don’t get that.

Adam Thierer: I have likened what’s happening right now in American technology policy to the development of an “AI Articles of Confederation.” I really believe we’re at the cusp of a brand-new regulatory paradigm for technology policy in the United States, and I think it absolutely stinks.

It’s destructive not just to interstate commerce but to the flow of information and speech. This is clearly interstate in nature.

When the Clinton administration released its framework for global electronic commerce, they said global electronic commerce wasn’t even national. They understood there were unique characteristics to the internet and digital markets that required different types of governance and a different locus of control.

And yet for AI, we are going exactly backwards. We’ve reverted back to the old analog-era norm of saying it’s every man, woman, and child for themselves in the states.

So now you have the Sacramento effect of state and local regulation and the Brussels effect internationally basically squeezing American innovators. And what’s our federal government doing to help? They’re pulling models off the markets on a Friday afternoon at the whim of somebody in government.

It’s a strange way to win an AI race.