What the Anthropic fight says about AI regulation

Fables of the deconstruction of American power

Below, a post on the Trump-Anthropic fight over Fable, focusing on the generalities (which are more important than you might think) rather than the particulars, which I don’t know any better than anyone who has been following events. If you haven’t been following events, skip down to the first footnote for a brief outline of the facts that don’t seem to be in dispute.¹

I’m less interested in the fight itself than the politics behind it. One group of commenters sees the blocking of Fable as an arbitrary exercise of power by an administration that is using export controls to punish Anthropic rather than cracking down on exports of AI chips to China. Another blames Anthropic, saying that the company pressed for regulation of what it itself has described as a dangerous technology, and is hypocritically complaining now that it has gotten what it was looking for.

These two perspectives loosely map onto two broader approaches to AI regulation which are currently struggling for dominance. The first is associated with the past Biden administration and Anthropic itself, and argues that geopolitics and US-China relations make it urgently necessary to regulate the diffusion of AI to make sure that America stays in control. The other is connected to the informal alliance between the Trump administration and the pro-AI, anti-DEI coalition in Silicon Valley, and argues that impeding the development of a transformative technology will carry massive geopolitical and human costs. Both assume that AI will likely have world-changing consequences, but disagree on what to do next.

There is a third approach that better gets at the underlying problems. It is best articulated in a piece that Alondra Nelson wrote a few months back for Science. Alondra argued that it was a big mistake to think that the Trump administration wanted to deregulate AI. As she put it:

What distinguishes the current [Trump administration] approach is … its character: a greater reliance on executive discretion than on deliberative, public processes. … what it has advanced is not the absence of AI regulation but its rearrangement, often by caprice: intensive state intervention operating through industrial policy, trade restrictions, immigration controls, equity stakes in private firms (selected by the state), the redirection of research funding, and the strategic preemption of state authority.

Perhaps then, instead of shoehorning debate into stylized pro- and anti- regulation positions, we should ask what particular forms of regulation actually involve, and what degree of discretion is associated with them.

Specifically, given what has happened with Fable (and in earlier Trump administration moves against Anthropic), we should pay attention to the tools of national security policy as a mode of regulation. Export controls and related national security tools provide the administration with great discretion in regulating tech. This was true before Trump’s second term: the Biden administration proposed a vast system that would control who had access to AI across the entire planet, which would not be founded on treaties, or even unilateral legislation, but on export controls and executive agreements. It is even more true now.

In the past, this discretion has been deployed in pursuit of a vision of global American technological hegemony. Now, it is being used for very different purposes. What the Trump administration is doing to Anthropic is bad, in the same way that other Trumpian exercises of arbitrary regulatory authority to punish enemies have been bad. Equally, it is the past vision of national security via American power, which Anthropic openly embraces, that makes the Trump administration’s sweeping claims of authority over technology possible. Anthropic is trapped by the assumptions it shares with the administration. It cannot readily push back against arbitrary exercises of export control powers, without recognizing how this arbitrariness is built into the broader framework of national security.

********

Alondra’s article highlights the ambiguities of the word ‘regulation.’ People often think about regulation as narrowly involving governing through the predictable application of specific rules. As Alondra suggests, there are many means through which government regulates private sector behavior. My own personal list would also include informal connections, antitrust and other such legal interventions, spending decisions and a variety of other threats and inducements, whether public or private; perhaps that goes too far.

Under any broader understanding of regulation, the Trump administration is not opposed to regulating AI or other private sector activities. It is emphatically interested in controlling what private actors do and say. It just wants to move away from regulatory approaches that emphasize predictability and public or democratic accountability and towards approaches that maximize its flexibility and discretion. Such a shift (a) reflects a notional theory of the unitary executive, under which the president uniquely represents the interests of the American people, (b) allows the administration to undermine and punish actors whose politics they dislike, and (c) allows people in the administration or their friends to pursue private benefit at the public expense while minimizing the opportunity for outside pushback.

That differs in very important ways from how other modern administrations have viewed their role. But - and this is a big but - there are genuine continuities with existing institutional frameworks. Some policy areas are much more open to executive discretion than others. In particular, the zone where national security, economics and technological knowledge intersect has always allowed a great deal of governmental discretion, because such discretion was seen as necessary to the untrammeled pursuit of the national interest.

That applies, for example, to financial sanctions. Abe’s and my book, Underground Empire (obligatory plug: do go buy ) explains how the US employed strategic ambiguity to scare foreign banks and businesses into advance compliance with administration wishes. As Stuart Levey, who used to run the relevant bits of Treasury, described banks’ incentives:

Keeping a few customers that we have identified as terrorists or proliferators is not worth the risk of . . . a regulatory action that may impact on [sic] their ability to do business with the United States or the responsible international financial community.

A major US law firm said that US sanctions rules “at times seem almost purposefully confusing” so that “the strongest impact of these sanctions may be their mere existence rather than their exercise, as highly-regulated and risk-averse financial institutions steer well clear of the line.” The result, as economic sociologists Gregoire Mallard and Jin Sun have described it, is that banks have developed extraordinarily cumbersome systems of internal controls to satisfy U.S. demands for compliance.

Export controls have similarly become a cornerstone of US power. Kevin Wolf, who ran the export controls bureau at the US Department of Commerce, told us how the first Trump administration built up the so-called Foreign Direct Product Rule (FDPR), which would apply U.S. export controls to a much broader range of products made across the world, like financial sanctions. As he put it, “trillions of dollars of transactions, trillions with a ‘T’ ” were affected by a technical rule “that was buried in 9 point font in a footnote at the bottom of a 320 page Entity List.” Weirdly for Abe and me, our ideas were used by a senior official in the first Trump administration to justify this rule change.

Export controls are often specified in much greater technical detail than sanctions, sometimes allowing them to be gamed. Equally, crafting and applying them involves difficult judgment calls. As Mario Daniels and John Krige note in their recent history, Knowledge Regulation and National Security in Postwar America, “to assess how dangerous a technology, if exported, could be, is incredibly challenging.” The result is an incredibly detailed and complex body of regulation, which is based on tricky judgments both of how the technology might be used against the interests of the United States, and of the actors most likely to use it in these ways.

Daniels and Krige emphasize that export controls stem from America’s sweeping vision of national security, which posits that:

the control of intangible knowledge and know-how was the most important function of export controls. They target objects as carriers of “embodied” knowledge that can, for example, be extracted by means of reverse engineering or actively transferred in face-to-face interactions between an American donor and a foreign receiver. Moreover, these targets are assessed not in isolation, but in relation to the technological system in which they are, or will be incorporated. The overriding question is: “would this export help the foreign country to close a gap it cannot fill with its own knowledge?”

In a response, Karen Alter argues that it’s more complicated than this: some US officials recognize that it can be counterproductive to try too hard to control the spread of knowledge between the United States and other countries. But all these authors agree that export controls have the purpose of ensuring continued US technological leadership. Furthermore, Karen notes some agreement with Daniels and Krige:

The far reaches the US government asserts via the deemed export category alarmed the authors, and it will also surprise most readers because the rules apply to all US citizens and firms, regardless of where the sharing occurs. Sharing knowledge with graduate students or foreign colleagues in a US university or academic meeting; international research collaboration with foreign labs or a company subsidiary; working with new foreign owners to help them understand what a US company does; setting up a joint production or licensing arrangement in the US or abroad; even the non-commercial sharing of the ideas and expertise that budding entrepreneurs or researchers generate and own within their very person may be controlled, disallowed, and even criminalized.

To be clear: there are limits. Such controls over conversation are generally understood to be restricted to engineering and technical knowledge: mere social scientific knowledge and ordinary conversation aren’t covered. Daniels and Krige describe how the first Trump administration’s National Security Strategy:

continue[d] to stress the need for deregulation to stimulate exports. But it combines the use of export controls to deny sensitive technology to rivals with a raft of new measures to erect a protective wall around sharing emerging critical technologies and related knowledge with China, in particular.

When the Biden administration came into office, it built upon this approach, creating what then-National Security Advisor Jake Sullivan described in a speech as a “Small Yard, High Fence” strategy, under which China would be denied access to the high-end semiconductors needed to train powerful AI, and other “foundational technologies” as needed. The fence that was supposed to deny China access was mostly made out of export controls.

As Trump was preparing to come into power, the Biden administration proposed an extraordinary plan to use export controls on semiconductors and AI model ‘weights’ to control the development and use of AI across the planet. The entire world would be divided into three different zones - close friends and allies, which would have more or less unrestricted access to AI, competitors and enemies who would have none, and a large intermediary zone of countries with initially limited access, who could negotiate with the US for better arrangements.

Like many other exercises of US technological power, this plan depended on a complex mixture of detailed technocracy and high discretion. Biden administration officials hoped that extraterritorial export controls would allow the US to leverage the US head-start in AI, building an enduring system of global control over future development of technology, while providing officials with substantial discretion to cut specific deals with other countries, such as the Gulf States, which wanted more access to AI.

The Trump administration, however, declined to take up this detailed technocratic plan to impose a system on the rest of the world, in favor of its own vaguer but equally sweeping ambitions for AI “dominance.” It has no principled objection to sanctions and other tools of economic coercion, although officials express their contempt for the moralism that has inspired their use in past administrations. Instead, the current administration cares more about short-term deal-making and breaking, and raw exercises of power. It seems at least as interested in subduing its domestic enemies as in coercing adversaries abroad. Both are means towards the end of remaking a world so that the administration and its friends are in control, and others either have to supplicate them or suffer the consequences.

Hence, the administration is less interested in further cutting off China from AI, especially after China has demonstrated its ability to cut off the supply of rare earths that US manufacturers care about, than in figuring out deals. The Biden administration wanted to use AI to bind the Gulf States closer to the US, and further from China. The Trump administration has looked to cut them into the bargain, in return for general and particular concessions. Externally, AI is another means of leverage to exploit as the administration tries to make one-sided bargains. Internally, it would like to adopt AI to replace bureaucracy with all its human inconveniences. It worries about the AI labs as an alternative power base, and is especially concerned about Anthropic because of its connection to the past administration and loosely liberal politics.

*****

Thus, the Fable fight is part of a bigger struggle. Both of the two dominant approaches to regulating AI believe in continued American technological hegemony. Both tend to be bullish about the potential of AI to radically transform society. They disagree on whether it is better to build global power through painstaking technocracy or through deals based on self interest, and on what should happen on the home front.

As already mentioned, the Biden administration enthusiastically embraced the first Trump administration’s expansion of export controls to cover actors outside the US, and deployed it systematically to stifle China’s AI ambitions, through restrictions on high end semiconductors. As Graham Webster describes it, officials feared that authoritarian governments might use AI as a tool of widespread control. In part, they worried that whoever got first to ‘recursive self-improvement’ (AI that could make itself better in a feedback loop of self-reinforcing awesome), would be able to shape the destiny of the world.

The stakes were especially high, some believed, because a powerful AI system might be used to develop even more capable AI systems, compounding a country’s lead and leaving competitors not just incrementally behind, but wallowing in a bygone paradigm. If some of these fears sound vague or far out, that’s because they are. Policymakers, however, are forced to take action in the face of uncertainty, and here, the uncertainty itself was cause for alarm.

That led them to embrace export controls as a tool of global management, hoping that they could keep China behind for just long enough to make the necessary difference.

the US government found the FDPR irresistible. It would … eventually wield it to constrain high-powered computing in China. “Obviously we started using it like candy,” says Estevez. “Certainly threatening to use it if not actually using it.”

The Biden administration’s belated sweeping plan to remake the world, using export controls to subordinate other countries to America’s technological hegemony, was a culmination of this approach.

Now, the Trump administration is more interested in deals with side-payments than in constructing a global empire of technocracy. It wants instead to use national security tools to subordinate American companies. First, the US Department of Defense designated Anthropic as a supply chain risk - a tool that has hitherto been deployed against non-US companies - because it refused to revise a contract that prevented its models from being used for autonomous weapons or for domestic mass surveillance in the US. As Dean Ball (who was the White House AI advisor for the first part of this administration) described what happened:

Anthropic was the company most enthusiastic about these national security uses, and they came to an agreement with the Biden administration … the Trump administration agreed to those restrictions, as well. Then in the fall of 2025 … Emil Michael as under secretary of war for research and engineering…. comes to the conclusion that no, we cannot be bound by these usage restrictions. The objection is not so much to the substance of the restrictions but to the idea of usage restrictions in general. So that conflict actually began several months ago.

That is how and why the US government has gone after Fable. Supply chain risk designation and export controls authorities most certainly cover technologies like AI - indeed it is hard to imagine a technology that is more precisely a “carrier of embodied knowledge” than a frontier model - but they are not supposed to be used for domestic control. However, because national security authorities are quite open-ended, they provide the executive branch with a lot of discretion to go after domestic actors whom it doesn’t like. The result is that export controls and related measures are being used to shape domestic politics in Trump-friendly ways.

People clustered around the Biden administration-Anthropic axis argue that something has gone wrong when the Trump administration goes after Anthropic, while apparently slow-walking many of the Biden era’s measures against China. Under their understanding, the US is misusing these tools, neglecting actual national security threats in favor of notional ones.

Others such as Dean - who has just moved to a political/strategic position at OpenAI - do not fully embrace the administration’s approach but say that it is understandable, since they believe that AI is a fundamentally world changing technology.

Everything the U.S. government was communicating, in policy and in rhetoric, seemed to suggest “go ahead, release your model!” And yet common sense would dictate otherwise. Anthropic is still in the midst of a heated dispute with the Department of War about that agency’s decision to label Anthropic a supply-chain risk. Bitter disputes about policy and politics between the Administration and Anthropic remain unresolved, among them export controls, federal preemption, and the general reality that Anthropic supports Democratic candidates for office while Republicans occupy the seat of power.

Of course they needed to tread carefully. What the law says does not matter. What Administration officials argue on one day does not matter. Anthropic is a political enemy of this Administration, in part because they have explicitly chosen to make themselves one. …

The stark reality is that making superintelligence is a profoundly political act even in the healthiest of societies, to say nothing of the filthily political world we Americans currently inhabit. A model like Mythos goes beyond being a mere political act and implicates the sovereignty of the state itself. No company gets to shake the foundation of state sovereignty while staying blithely above the raw reality of politics.

It is unusual for a sincere classical liberal to defend an administration exploiting regulatory ambiguity to go after its “political enemy” on the grounds that “[w]hat the law says does not matter. “ The presumable justification is that AI is in that small class of technologies (e.g. nuclear weapons) that are so powerful that private companies, especially one that “supports Democratic candidates,” cannot be allowed control.

That seems a questionable way of thinking about the politics of AI, but it is startling that so many of the underlying premises are shared with the arguments of a blogpost that Amodei put up shortly before the launch of Fable. AI is set to transform national security.

Mythos Preview scrambled the global cybersecurity landscape. But its broader significance is that it proves beyond doubt that AI models are now tools of global and national strategic consequence.

And it needs to be regulated.

I believe the best analogy, at least at the current stage of the exponential, is to cars, airplanes, or drugs—powerful technologies essential to the modern economy, but capable of killing large numbers of people if designed or operated poorly. I therefore believe we should model AI regulation on agencies like the Federal Aviation Administration (FAA). Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety [emphasis in original]. I am grateful to see the Trump administration’s Executive Order move incrementally towards a greater role for government in AI, though Anthropic’s proposal recommends even further action.

To be clear: it would be completely unfair to claim that Amodei was begging to be export controlled - he was calling for the careful construction of ordinary bureaucratic institutions rather than for sweeping seat-of-the-pants national security decisions by the executive. But when Amodei writes that:

US export controls on frontier chips and SME to China have been a major contributor to the US’s overall lead in AI, and these policies need to be expanded, tightened, and coordinated with other likeminded states.

he seems to be suggesting that export controls and other national security authorities are absolutely fine for China. It’s when they are applied to Anthropic (or, more generously: American and likely allied companies that are not looking to cooperate with the adversary) that something has gone wrong.

And this gets to what I see as the heart of the matter. National security tools like export controls and supply chain risk designations are relatively unbounded. That makes them very attractive to the current administration which, as Alondra says, wants to regulate in ways that maximize its discretion. And these tools are likely to become more attractive soon. If the Republicans lose one or both houses of Congress in November, as seems likely, the administration will find some of the means that it is using harder to employ. The president’s foreign policy powers as commander-in-chief - and willingness of courts to defer to them - have always been substantial, and have grown over the last couple of decades. That suggests that we are likely to see the Trump administration using national security powers more, and possibly much more, after November, to pursue domestic goals and punish domestic enemies.

Such overreach can be fought - possibly successfully - in the courts. On the one hand, courts tend to defer to the president on many national security issues. On the other, of the Achilles’ heels of the whole economic security apparatus is that large parts of it are built upon authorities that have been stretched, and stretched, and stretched some more. The foreign actors who have been the typical targets have often not wanted to appeal, because they haven’t fancied their chances in US courts. That might change with plaintiffs like Anthropic (and in so doing, reshape what the US government can and can’t do). But courts tend to work slowly, especially on sensitive matters. Victories won through the legal system may come too late to make a real difference.

The use of such powers puts both sides of the existing debate in an awkward position. I believe that Dean (who I know slightly) is sincere in his arguments, but they depend both on his enthusiasm for unitary executive theory, and his belief that current AI is, or is about to become, so powerful that it’s going to sweep traditional forms of government away. I think that Dean is straightforwardly wrong on both - but even if he is not, does he really believe that the Trump administration will confine its use of these tools to Anthropic and other AI companies? The logic of his position that perceived enemies of the administration, including companies that support “Democratic candidates for office,” can place no faith in the law, goes in obviously problematic directions for classical liberalism, which has typically been understood to imply the need to defend a genuinely plural society against the incursions of the state.

The implications for the other side are more subtle, but quite as damaging. The standard approach - as Amodei’s own writings repeatedly make clear - is that you can have two great things at once. You can build up a thriving democracy at home, regulating AI carefully to avoid existential risk and support democratic institutions, and you can use your awesome national security powers abroad to stunt the technological development of authoritarian adversaries.

I think that this formulation was already in trouble, in a world where China is clearly capable of building up a large array of technological advantages regardless of US coercion. But it is now being undermined from inside too, as the second Trump administration is beginning to turn the national security state against domestic enemies.

What is now becoming obvious is that the Amodei and Biden administration position that you could have your cake and eat it too, depended on a semi-explicit consensus on what US national security was, and how sweeping powers with ambiguous constraints could or could not be used to pursue it. That consensus has now shattered, and it isn’t plausibly going to be rebuilt. Strengthening powers such as export controls without very careful design (and I am not sure that they can be so designed) will strengthen the present and future abilities of illiberal and antidemocratic officials to deploy these powers against domestic as well as foreign adversaries.

These and other related tradeoffs are likely to get sharper over time. The United States has a set of national security institutions that are purpose-built for a dominant power, but the United States itself is now clearly in relative decline. Its ability to use these institutions against foreign adversaries will dwindle. Its ability to deploy them against perceived domestic enemies will not shrink to nearly the same extent. We need to think more seriously about the problems that this will create, and indeed is already creating. I’m not seeing much of that thinking yet.²

Addendum: This, from Ganesh Sitaraman and Tim Meyer is extremely useful:

To date, AI companies have wanted an unregulated federal regime – and they have even pushed for a ban on any state regulation of AI. But the demand for regulation is like water: it has to go somewhere. Frontier technologies like advanced AI models raise too many policy considerations–to say nothing of the economic competition between technology firms both within and outside the United States–for policymakers to simply turn a blind eye.

In practice, the failure to regulate economically and politically significant developments means that we will get presidential regulation, an idiosyncratic and non-transparent or procedurally irregular system of regulation that will oscillate based on who is president, which company is at issue, and the particular issue. To the average American, it will seem much more like rule by fiat than the rule of law.

1. Anthropic warned some weeks ago that its new Mythos model could discover cybersecurity vulnerabilities at scale (there is some disagreement among outsiders about how effective it was) and declined to share it beyond a small number of trusted partners. Those outside the limited circle of trust got Fable, a version of Mythos which was supposed to have substantial built-in safeguards that would limit its usefulness for finding vulnerabilities. Soon after its launch, Andy Jassy, the CEO of Amazon, expressed concerns both to Anthropic and the Trump administration that the Fable model could be jailbroken so as to circumvent these safeguards. After a brief back-and-forth with Dario Amodei, who leads Anthropic, the Trump administration quickly used its export controls authority to block Fable from being shared with non-US persons. This effectively meant that Anthropic had to stop everyone from accessing Fable, since Anthropic has no way of knowing whether e.g. a French person living in the US is using the model. Anthropic very likely had to prevent any of its many non-US citizen employees from having access either.

2. A general stylistic/acknowledgment of relationship note that I’ve been meaning to explain for a while. As with many of the topics that I talk about, I am drawing on the work and ideas of some people I know, and some whom I don’t. When I am on first name terms with someone, I will typically provide their full name the first time that I mention them, and afterwards will just use their first name. When I am not, I will use their full names or surnames throughout. When there is something highly germane to the post about my relationship with them, I will say so. I’ve evolved towards that style over time as a way of signaling social context without being too obtrusive about it, and it is about time to explain it explicitly.