With July 4th looming, it’s tragically poetic that the country is again facing the question of how to constrain the exercise of arbitrary power. The Trump administration’s decision — based on questionable justifications — to block foreign nationals from using two Anthropic models indicates that AI governance is being shaped by actors who wield incredible influence over this critical domain and are yet subject to few effective constraints. This episode strengthens the case for Congress to step up and pass an AI governance framework that will reduce the frequency with which AI policy is breaking news.
What Happened
A brief review of the known facts (of which there are not many—itself a troubling sign from a rule of law perspective) surfaces the many ways in which highly consequential AI policy decisions seem disconnected from a reasoned, transparent, and thorough process. Commerce Secretary Howard Lutnick sent Anthropic a letter on Friday detailing that their latest models—Fable 5 and Mythos—are now subject to export restrictions.
Mythos drew popular attention following documented instances of its being able to detect and exploit cyber vulnerabilities. Anthropic decided to share that model with a select set of trusted partners rather than risk bad actors leveraging the technology for nefarious ends. Fable 5 is a version of Mythos with several additional safeguards and was therefore made popularly available. In fact, Anthropic reports that Fable 5’s safeguards are “substantially more effective than those of any previously deployed model.”
The administration’s imposition of an export restriction followed from evidence of a potential narrow, non-universal jailbreak of Fable 5. Amazon researchers probing the model determined that it could provide users with information about cyber vulnerabilities when directed at websites. Yet, as Katie Moussouris, chief executive with the cybersecurity firm Luta Security, told the Wall Street Journal, such information is far more valuable to defenders of Internet systems than attackers.
Nevertheless, the administration’s directive means that Anthropic can no longer make those models available to foreign nationals, theoretically including their own employees. Anthropic responded by blocking everyone’s access to Fable 5 and Mythos to ensure it does not violate the administration’s expansive directive.
A Failure to Adhere to the Rule of Law
In conjunction with my frequent co-author Alan Rozenshtein, I have previously reported on the ways in which AI governance by ad hoc executive action (or, in many cases, inaction) and by exclusively private efforts raise rule-of-law concerns. The rule of law can be defined in many ways by myriad actors, but most would agree that its core tenets include publicity, prospectivity, and legality. Publicity mandates that the affected party know the rules and laws that govern their conduct. Prospectivity ensures that the affected party is able to align their behavior with any applicable rules and laws, rather than be held accountable for prior acts. Legality confines every actor to only operate where they have clear authority to do so.
Failure to adhere to those tenets undermines the extent to which individuals can live an ordered life and monitor those with extensive influence over their well-being. Government officials opting not to adhere to the rule of law also imposes significant costs on our economy. “The rule of law is the most influential factor for long-term economic growth and societal wellbeing,” as Annie (Yu-Lin) Lee and Joseph Lemoine summarized in an Atlantic Council report. Perpetuating an AI governance approach in which enforcement decisions are publicly announced on a Friday during the middle of a World Cup game is unlikely to safeguard individual liberty or encourage innovation and entrepreneurialism.
That’s precisely why the administration’s recent action requires scrutiny. And that’s precisely why Congress must act so that such consequential governance decisions are not disclosed via X posts and are instead made in alignment with the Rule of Law.
On publicity, the administration’s decision to take extraordinary action against Anthropic was disclosed by Anthropic—not the administration. Of course, the government necessarily withholds information or delays public announcements of enforcement actions in many cases. When it comes to decisions that may deny millions of people access to AI technology that is already shaping our economy, such opacity is in immediate conflict with the rule of law.
On prospectivity, a scroll of X makes clear that no one saw this coming. That’s not a good thing. While some AI scholars have flagged that export control authorities may one day be used to block individuals from accessing certain AI models, it has not been the topic of popular discussion. In turn, it has not been a potentiality actively shaping the decisions of AI labs. This surprising turn to export restrictions—even after Anthropic and other labs have engaged with the administration on numerous occasions about the cyber risks posed by their models and the safeguards taken to mitigate those concerns—instead suggests that labs will have to constantly be looking over their shoulder for what ‘creative’ enforcement decisions may come down the pike.
On legality, research on the export restrictions raises significant doubts about whether these provisions are intended to reach AI models. As Joe Khawam and Tim Schnabel wrote in Just Security, “[T]hese frameworks, designed for discrete transfers of static information between known parties, are ill-suited to govern AI systems that generate unlimited, dynamic outputs on demand for potentially anonymous users worldwide.” Perhaps the administration has a clearer explanation as to why the legal authority it’s relying on in this case fits the facts. However, that explanation has not yet been made available.
What has been made clear is that if the administration thinks such restrictions are legally available and necessary, it ought to limit access to many more such models offered by other labs. Anthropic is not alone in offering models with advanced cyber capabilities, which were purportedly the basis of the administration’s decision, as former AI czar David Sacks spelled out on X. The related rule-of-law concept of generality mandates that like cases be treated alike, which implies that the administration should be in the process of sending letters to other labs.
Conclusion
This is the second time in as many weeks that I’ve written about wild swings in AI policy. It should not be the case that the government needs to be reminded by private entities about the rule of law. Yet, here’s what Anthropic had to say on Friday: “[W]e believe the government should have the ability to block unsafe deployments as part of a statutory process that is transparent, fair, clear, and grounded in technical facts. This action does not adhere to those principles.”
Transparency, fairness, clarity, and evidence-driven enforcement should be core components of AI governance. So long as AI policy decisions are driven by a few actors taking action behind closed doors in response to non-public reports, we will be far short of that standard.