While AI companies and President Trump continue to ponder a pause in developing frontier AI models, a much more difficult problem looms on the horizon.
In the near future, pause or no pause, it’s highly possible that a dangerous open-weight AI model will be released in a form that current federal regulation cannot meaningfully contain. Today, the federal government’s regulatory power derives from the president’s orchestration of various financial and legal threats against U.S. frontier AI companies to regulate deployment of their closed-weight models (i.e., models that cannot be directly copied and modified by users). But once open-weight models that users can freely copy and modify catch up to the frontier—and experts think that will come in a matter of months—that regulatory structure will prove inadequate. There will be too many companies and people that can replicate the model, and they won’t all have federal contracts or other valuable relationships the president can credibly threaten. The president cannot jawbone them all.
When that happens, someone in the federal government will need to embark on serious regulation of existential AI risk, whether by having an existing federal agency take on a dramatically expanded role or creating a new agency to take on the task. But the Supreme Court has inadvertently thrown down a formidable gauntlet for would-be federal regulators. A series of rulings dating back to the 1990s has left the federal government with a narrow, possibly politically untenable path to regulating the open-weight models that make existential AI risk hardest to contain.
The Jawboning Model and Its Downfall
The current federal AI regulatory paradigm involves President Donald Trump jawboning closed frontier AI companies when something dangerous may (or may not) be going on with their models. Trump’s executive power maximalism highlights the danger of centralizing power so aggressively in one official. But it is not totally irrational to find existential AI risk sufficiently daunting that one would reluctantly rather have someone who can regulate AI, even if they are not the optimal choice.
The specific legal mechanism for Trump’s recent game of red light/green light with Claude Fable and ChatGPT 5.6 was export controls. The Commerce Department may or may not have actually had legal authority to use export controls in this way. But legality was mostly irrelevant to the situation. As the dispute over the Defense Department’s supply-chain-risk designation of Anthropic illustrated, Trump is willing to use retaliatory leverage unrelated to model safety.
That system sort of worked. It certainly slowed deployment of frontier models. More broadly, the jawboning model works right now because there are so few closed-weight frontier-AI companies: OpenAI, Anthropic, xAI, and Google. And really, the two that matter right now are OpenAI and Anthropic. Like many major U.S. corporations, Anthropic and OpenAI are largely trying to curry favor with Trump, not pick fights that invite escalation.
The jawboning model doesn’t work with open-weight models. Once the weights are released, there are too many companies that can set up their own model based on the weights, and each company is too small to have a particularly significant stake in maintaining federal government access. Trump probably can’t threaten, say, Fireworks AI (a U.S. company offering China’s Kimi K3) with the loss of Defense Department contracts because they likely don’t have such contracts. If frontier AI will have dangerous intelligence at some point in the future, and open-weight models are only a few months behind, all the jawboning antics in the world will only postpone the reckoning by a few months. What is the plan when another Chinese company releases an open-weight model similar in capability to Anthropic’s Mythos?
The impending decline of the jawbone model means it will soon be necessary for the federal government to figure out a new, more effective regulatory system. But this is where the Supreme Court’s gauntlet will soon confound would-be regulators.
Bipartisan Commissions with Regulatory Continuity—Eliminated by Trump v. Slaughter
With open-weights models, regulation would have to be both more flexible and potentially more intrusive. AI is complicated, changes fast, and poses potentially significant dangers. That is not an unprecedented set of attributes. Consider securities regulation and monetary policy—both areas that share these qualities, and which Congress entrusted to independent agencies (the Securities and Exchange Commission and the Federal Reserve).
A multi-member independent commission along these lines could be structured in a bipartisan way to build trust with the public at large and Congress. That bipartisanship can allow for continuity across administrations. An independent commission may also more effectively retain technical experts long-term as compared with agencies whose political officials (and in certain cases, some civil servants) are easily removed. It can be granted enforcement power to move quickly and decisively. It can also be given licensing and inspection authority. This kind of model could have been highly effective for handling the complexities of regulating AI.
But the feasibility of this model is now seriously complicated because of Trump v. Slaughter, which held that the president must be able to remove the officers of an independent commission at will. Although it may be too early to characterize the full extent of what the elimination of bipartisan commissions will mean in practice and where any creative space might be found, there is plenty of reason to be concerned. With presidential removal power, the benefits of a commission largely evaporate. A regulatory framework can instill a panoply of commissioners but if they all answer to the president, then the situation is no different than any ordinary department of the federal government led by political appointees. Nonpartisanship is functionally nonexistent, as is continuity across administrations. What each president wants, each president gets.
Why the President Cannot Close the Open-Weight Gap
In sum, neutral technocrats likely cannot be insulated from presidential control, and the current jawboning model is not sustainable. But perhaps a wise president could use the powers of the office to issue formal, legal regulations governing open-weight AI?
The problem is that the Supreme Court has hemmed in executive power in ways that make this approach unreliable.
Arguably the signature doctrinal innovation of the Court under Chief Justice Roberts’ leadership is the major questions doctrine. The doctrine requires Congress to clearly delegate authority to regulate a fundamental sector of the economy.
Although it’s not impossible that the Court will turn out to be more deferential on AI-specific matters than recent precedent would indicate, there’s a readily available argument that the major questions doctrine would seem to preclude the executive branch from regulating AI, particularly open-weight models, under existing authority. A regulatory scheme capable of reaching open-weight models would necessarily be broad and intrusive. It would need to regulate the possession and running of an open-weight model, even if the output is for research rather than commercial purposes. And AI generally represents an increasingly vital part of the economy. Without clear delegation, it’s reasonable to anticipate that courts may not conclude that Congress granted authority for such consequential regulation.
Relatedly, Loper Bright Enterprises v. Raimondo, the case that ended Chevron deference in 2024, will make it more difficult for agencies to use existing ambiguous or broad authorities to regulate AI. Even if agencies could construe a broadly-worded statute to survive major questions scrutiny, the agency interpretation is less likely to stand as a result of this precedent. A court, now required to apply its own independent construction of the authorizing statute, could not defer to a strained agency interpretation in the manner that might have been previously available.
But what of national security deference? The president has some power (maybe statutory, maybe constitutional, maybe both) to protect the country from national security threats. It is not hard to make the case that existential AI risk is a national security threat. So, perhaps the president can decide which models are safe or unsafe for deployment under this construction?
Not quite. The Supreme Court’s recent decision in Learning Resources, Inc. v. Trump suggests that the Court will not necessarily read national security and foreign affairs statutes broadly in favor of the executive branch. And the statute at issue in Learning Resources, the International Emergency Economic Powers Act (IEEPA), would be even less likely to help the president regulate open-weight models. IEEPA withholds authority to regulate the import or export of “information or informational materials,” regardless of format or transmission medium. Whether machine-generated model weights count is contestable. The government has argued in an adjacent context that the exception covers expressive material, not every functional dataset. And IEEPA is not a clean authorization for a domestic licensing and possession regime, and an agency effort to make it one would invite both statutory and First Amendment litigation.
Indeed, regulating open-weight models to mitigate existential AI risk is a poor fit for national security deference. Consider the leading case on the limitation on unilateral presidential power, the 1952 Supreme Court case Youngstown Sheet & Tube Co. v. Sawyer. The Court struck down President Truman’s seizure of steel plants during the Korean War in the face of a potentially imminent supply emergency.
AI regulation would present a much less compelling case for deference than Truman’s seizure. The president would have to engage in permanent domestic regulation and the promulgation of new rules without any relevant ongoing armed conflict. An executive order requiring private parties to obtain federal permission before storing or using an open-weight model would create domestic law rather than execute it.
Although it’s possible the Court may surprise us and rule on AI-specific matters in a manner that is not wholly consistent with these precedents, it is also entirely possible (perhaps even likely, in my view) that it will not. And that will certainly leave those interested in regulating this sector with limited options.
Commerce Clause Problems for Open-Weight AI Regulation
Congress is currently unified under Republican control and yet is barely able to skate from one shutdown threat to the next. Under these conditions, it is hard to imagine Congress actually passing meaningful AI regulatory legislation of any kind, let alone something reaching existential AI risk. And, of course, for the next two-and-a-half years, it is doubtful even Republicans want to give some new major grant of regulatory power to Donald Trump.
But the less appreciated danger is that even if a modern-day Disraeli produces a legislative miracle and gets a bill passed that a president signs, the resulting law could face serious constitutional challenges.
Let’s start with the widely assumed constitutional authority for federal AI regulation, the Commerce Clause. It can certainly provide the basis for regulating U.S. frontier AI companies because their operations just about always involve multiple states. But does the Commerce Clause empower Congress to prohibit every person in the United States from possessing specified weights or locally operating models built from them, even when that person is not engaged in sale or transmission of data?
Congress can regulate a U.S. developer’s commercial release, repository and cloud-hosting services, and cross-border downloads. Congress also has express power over foreign commerce, so the fact that a model originates in China does not disable regulation of its import into the United States. But the practical problem is that if the weights are the only thing one needs to effectively recreate the model, then it is difficult to actually stop importation of the weights.
Realistically, regulating open-weight models requires a law that dispenses with any transactional nexus and simply prohibits specified weights or models wherever found. Before the 1990s, such a law would likely have survived under New Deal-era precedents allowing Congress to reach intrastate activity whose exemption would undermine a federal regulatory scheme.
But United States v. Lopez (1995) and United States v. Morrison (2000) present a significant problem for open weights regulation. In those two cases, the Court held that noneconomic conduct (possession of a firearm or violence against women) could not be regulated under the Commerce Clause even if the activity in question could have broad economic consequences. Noncommercial local operation of open weight models could similarly have vast economic consequences, but is not itself economic activity.
The 2005 case Gonzales v. Raich supplies arguments to both sides. Raich involved federal regulation of locally grown marijuana intended for use rather than sale. The Court reasoned that the Controlled Substances Act comprehensively regulated an interstate market in a fungible commodity, homegrown marijuana could be diverted into that market, and an intrastate exemption could leave a “gaping hole” in the federal scheme.
The problem for open-weight model regulation is that the conduct Congress ultimately cares about is local operation, not participation in an interstate market. Dangerous weights may be given away rather than sold. Running them offline does not necessarily alter the supply or demand for an interstate product. The catastrophic harm Congress seeks to prevent is not itself economic. Those points likely make Lopez and Morrison more relevant than Raich.
The Open-Weight Delegation Dilemma
Even if AI legislation survives a Commerce Clause review, it would almost certainly also face a nondelegation challenge. Much like the Commerce Clause, the nondelegation doctrine was dormant for much of the second half of the 20th century, but has been revitalized by conservative members of the Supreme Court.
All of the current Republican-appointed justices have at one time or another expressed support for stricter enforcement of the nondelegation doctrine. Justice Barrett wrote a law review article in 2014 that explicitly called the intelligible principle test (the heart of nondelegation doctrine, requiring only some guidance to cabin executive branch discretion in using a congressionally delegated power) “notoriously lax.” The major recent nondelegation case, the 2025 decision FCC v. Consumers’ Research, saw Roberts, Kavanaugh, and Barrett joining the three Democratic-appointed justices in reaffirming and applying the intelligible principle test to uphold Congress’s delegation of authority to the FCC.
But the grounds of that decision are probably cold comfort for would-be AI legislation drafters. The majority upheld the scheme because Congress itself had made the important policy judgments on a relatively straightforward issue. The statute at issue told the FCC whom to benefit, which services to support, and supplied standards limiting how much revenue the agency could raise. That system had been in place for decades. A statute regulating open-weight models would realistically be unable to recreate those features because they could be used by anyone, and the form of the models and computing power necessary to run them might change dramatically in the future. In other words, the technology is evolving so rapidly that it is doubtful whether Congress could truly specify substantive constraints that are even somewhat future-proof. As a result, the application of the FCC precedent to AI regulatory efforts is uncertain.
Congress’s Narrow Path
Given the gauntlet the Supreme Court has thrown down, the narrow constitutional path left to Congress is something like this:
- Congress passes a detailed statute deciding which actors and activities are regulated; what factual findings justify intervention; and what prophylactic measures a federal agency can impose (e.g., testing, security requirements, delayed release, prohibition).
- It delegates implementation of this system to a new or existing agency. There is no immediately apparent way around the fact that this regulator would be presidentially controlled, vulnerable to politicization, and subject to partisan discontinuities.
- Given the rapid change in AI models and capabilities, Congress would likely have to update the legislation frequently to add new systems of regulation—maybe even more than once per year.
This is all theoretically possible but does not seem politically feasible. There is no recent evidence of Congress actually having the institutional capacity to regulate in this manner—see, e.g., Section 702 pertaining to foreign intelligence collection. The idea that it could do so about an issue where so much lobbying money and effort will be at play is not realistic.
And, of course, even if this framework was achievable, it is still effectively powerless to meaningfully regulate open-weight models developed abroad. This hypothetical congressional regulatory regime could not prevent a foreign developer from publishing model weights or recovering copies after they are first published. At most, Congress could regulate commercial deployment and require stronger resilience measures for critical infrastructure.
What Can Anyone Actually Do About This?
I do not have a hopeful answer here. The Supreme Court could overturn Trump v. Slaughter. Given the 6-3 party-line outcome, if two conservative justices are replaced by two liberal justices, then a return to Humphrey’s Executor could pave the way to a technocratic independent commission. Congress could also propose, and three-fourths of states could ratify, a constitutional amendment overruling Trump v. Slaughter and allowing for independent commissions (either generally or in the specific context of AI). Similarly, the Supreme Court or constitutional amendment could remove some other part of the gauntlet, like modern commerce clause doctrine.
Constitutional amendment is, of course, an unrealistic hope. Although it is early to fully evaluate the long-term implications of these rulings, the Supreme Court’s gauntlet has likely made durable, expert, democratically authorized regulation much harder to build, while leaving in place a system of ad hoc presidential pressure. But if we do not start working toward a solution soon, we will not be ready for the dangers of powerful open-weight models.
Disclosure: The author has done contract work for Anthropic supporting the training of Claude on analyzing federal law.
