A First Amendment Approach to Regulating AI Model Weights
The world is in a critical window for the regulation of artificial intelligence (AI) model weights, the crystallized intelligence that allows AIs to converse, write code, or generate images. In the United States, these weights have traditionally been kept private by the frontier AI labs, which run the models on their own servers and sell access through application programming interfaces (APIs). Open-weight models, in contrast, have publicly available weights and can be run directly by their users. They are surging in popularity in the United States and elsewhere.
This poses a dilemma: A U.S. AI ecosystem with more open-weight models would diminish concentration of power within the AI industry, but it would also limit the government’s ability to control AI—it is easier to oversee a small number of labs providing model intelligence than it is to tell millions of companies and private individuals to stop running model weights. As models demonstrate more dangerous capabilities, regulators may be compelled to address the dilemma by licensing the publication of capable AI model weights.
Yet the government risks regulating model weights poorly. The export control regime is currently its only realistic tool to license model weights, and it would allow the government to restrict their transfer with few procedural checks; that lack could inflict heavy damage on the AI ecosystem, and in particular on open-weight model providers.
This underscores the importance of the chief constraint on the export control regime’s use against AI models: the First Amendment. Model weights are trained on speech and may generate speech; they too may be speech. And if AI model weights are speech, the government will be obligated to institute far greater protections for producers and consumers of AI than it would otherwise. Given that, regardless of whether model weights are speech, we should regulate them as though they are. A regulatory regime for model weights that adheres to the guardrails of the First Amendment will be legal if they are speech and will inoculate this nascent technology against abuse even if they aren’t.
The Export Control Regime Is the Tool That Is Likely to Be Used for Major Regulation of Model Weights
It is plausible that the government will soon impose significant regulation on the dissemination of AI model weights in response to alarming increases in AI capabilities. If so, that is likely to be through a licensing regime, in which the government evaluates and approves AI models for release. The Commerce Department’s use of export control licenses in its June 2026 effective ban of Anthropic’s capable closed-weight models, Mythos 5 and Fable 5, shows there is already a template for the hasty licensing of AI models.
Indeed, the export control regime, encoded in the Export Administration Regulations (EAR) and administered by the Bureau of Industry and Security (BIS) within the Commerce Department, is the tool through which future licensing would likely also occur. With a few small adjustments, the export control regime could be configured to require prepublication licensing of model weights, allowing regulators to evaluate the safety of the weights before they can be published. (The EAR currently exempts published information from its controls, but BIS could carve model weights out of that exemption.)
Although nominally meant to restrict the access of foreign persons to U.S. technology, the export control regime could also regulate the access of Americans to open AI model weights: Releasing a model’s weights to a foreign person—whether abroad or within the United States—is an export under the regulations, and publishing weights on the open internet makes them accessible to foreign persons everywhere. In other words, even if the intended recipients of open weights are Americans, any lab considering open sourcing its weights would have a nearly impossible time preventing foreigners from accessing them. Open model weights could thus require a license prior to their publication if BIS said so. And even if model weights were being sold rather than open sourced, transfers to discrete American companies could trigger licensing: Under the deemed-export rules, the sharing of controlled technical information with foreign-person employees is itself an export.
The reach of the export control regime does not stop there. A lab could lease its weights to a foreign data center operator, transfer them to an overseas subsidiary, or sell them outright as part of an acquisition. All of these actions could be subject to licensing under the export control regime. The Foreign Direct Product Rule under the EAR, moreover, has allowed the Commerce Department to restrict the export of goods made using American parts, components, or software. That could be extended to the release of foreign model weights trained with American algorithms, chips, or data, as the rescinded AI Diffusion Rule briefly did for weights not yet published. While it is unlikely that China would abide by such rules, many U.S. allies might.
To be sure, there are risks associated with employing the export control regime. Many judges—particularly in the U.S. Court of Appeals for the Fifth Circuit—are skeptical of expansive executive uses of economic statutes. An adverse ruling against export controls on model weights could be written broadly enough to undermine the basis for the regime’s wider technology controls, such as the deemed-export rules. But those litigation risks may not be an effective deterrent: If the government reaches for the export control regime to restrict model weights in the wake of a crisis, it may not give much thought to an adverse court decision that lies years away.
The use of the export control regime specifically is made likelier by the fact that beyond it, there are few extant statutory authorities that adequately allow for AI model weight regulation. While the International Traffic in Arms Regulations (ITAR) under the State Department and the International Emergency Economic Powers Act (IEEPA) grant the government vast powers to regulate economic activity, they fall short. ITAR is built for defense articles, and although general-purpose model weights could in principle be designated on the Munitions List, their dual-use character makes ITAR an awkward fit; IEEPA, via the Berman Amendment, exempts “information or informational materials,” and whether that exemption would cover AI model weights remains an open question.
In short, the capabilities of AI models may be such that the government feels compelled to institute a regime for the regulatory licensing of AI model weights. If so, it would hastily reach for the only tool now available to do so: the export control regime.
The First Amendment and Model Weights
If model weights are licensed through the export control regime, it would likely be—like any other regulatory regime for model weights—challenged in court on the grounds that it is restricting speech. A First Amendment challenge would boil down to three questions: whether the weights themselves are speech, whether the restriction is content-based or content-neutral, and how the restriction operates, specifically whether it is a prior restraint and provides sufficient protections for speech if so.
The first question—whether the weights are speech—is a binary. If the item being controlled is speech, it will be entitled to First Amendment protections and it will be necessary to address the second and third questions. If it isn’t, the later analysis is moot.
But the speech status of model weights is unclear. Some argue that weights, which no human can read, are not speech and are therefore not protected by the First Amendment; others respond that weights encode editorial judgment, invoking the code-is-speech cases of the encryption era. Courts have yet to resolve the question. Drawing from the doctrine of constitutional avoidance, a court will likely initially address the second and third questions. It will discern the weights’ speech status only if its analysis reveals the regime would be illegal if the weights were speech.
On the second question, a content-based restriction—aimed at a topic, an idea, or a message, for instance a law banning signs criticizing foreign governments near embassies—is subject to strict scrutiny, which the government fails unless it can prove that the restriction on speech is the least restrictive way for it to achieve a compelling government interest.
A content-neutral restriction—aimed at what the regulated item does rather than what it says, for instance an ordinance limiting the volume on loudspeakers—faces only intermediate scrutiny. That standard is far more forgiving; the government usually prevails under it, though it can still fail if the restriction operates in a manner that is overly burdensome. Such a regime would not be toothless. For example, a restriction on the publication of AI model weights that were trained with more than a set number of computational operations (flops) is plausibly content-neutral.
It is difficult to predict whether a court will say a regime is content-based or content-neutral, as the doctrine often does not map cleanly to complex fact patterns. In the case of AI model weights, a regulation triggered by a model’s capabilities could be deemed content-based, since it involves what the model expresses. A regulation triggered by objective criteria such as training compute or cost is more plausibly content-neutral.
Last, regardless of whether the regime is content-based or content-neutral, it is necessary to answer the third question: how the restriction operates. Licensing schemes that require government permission before speech is released are prior restraints, the most disfavored form of speech regulation. To survive, they must carry procedural safeguards. When the prior restraint is content-based, Freedman v. Maryland requires deadlines for the issuance of a license, a default-granting rule, and prompt judicial review with the government bearing the burden. When the restriction is speech targeted but relatively content-neutral, these guardrails become looser: FW/PBS, for example, requires that it provide a decision deadline and the opportunity for prompt judicial review. And when the restriction is content-neutral and not speech targeted, it simply must apply objective standards of coverage without delay.
Evaluating Export Controls on Model Weights
The employment of the export control regime to license model weights is wanting from both a legal and policy perspective. As noted, if the above First Amendment framework is applied to export controls on model weights, the latter two questions will initially be most important: a court will consider the content-neutrality and procedural safeguards of the export control regime before determining whether model weights are speech.
Lacking these safeguards would put the regime on rotten legal foundations if model weights are speech—while the regime would likely prevail on the content neutrality question, it would still be likely to fail as an unconstitutional prior restraint.
Indeed, export controls could straightforwardly be content-neutral if the government covers models through objective criteria—for instance, a compute threshold, as it did for closed-weight models under the now-defunct AI Diffusion Rule. Intermediate scrutiny would be straightforward for the government to weather, and a compute threshold could also offer the export control regime jurisdiction over many of the models that could present the greatest risks. However, export controls could also be subject to strict scrutiny if—for example—they were based on an evaluation of what the model could tell the person using it. But while it would be much more challenging for the government to prevail under strict scrutiny, even there the government might if it could demonstrate catastrophic risk from release of the model weights.
That means the export control regime is far more likely to fail for lacking the procedural safeguards required of prior restraints on speech. Specifically, the export control regime lacks statutory deadlines and the right to judicial review, even as arbitrary and capricious under the Administrative Procedure Act. The regime’s one nominal deadline proves that point: applications must be “resolved or referred to the President” within 90 days, but there is no further deadline if the application is referred—the precise loophole the U.S. Court of Appeals for the Ninth Circuit condemned in Bernstein v. United States Department of Justice, the famous, albeit only persuasive, “code is speech” case which held both that encryption source code is speech and that BIS licensing of it was an unconstitutional prior restraint.
If model weights are speech, the absence of these safeguards could cause this regime to collapse. This is especially likely to occur if the export control regime is being subjected to strict scrutiny, requiring heightened safeguards, but could even occur under intermediate scrutiny as the export control regime lacks even the basic safeguards—objective standards of coverage and enforceable limits on when a decision must be made—that even content-neutral licensing schemes must provide.
If the prior restraints on model weights through the export control regime were unconstitutional, BIS could—as in the aftermath of Bernstein—remove the prior restraints by requiring that an AI model provider merely notify it of a release of a model, rather than asking it for a prepublication license. However, a notification regime would be practically toothless, allowing model weights to be released and saddling the government with the difficult task of trying to recall them. A ruling that export controls on model weights are an unconstitutional prior restraint could therefore effectively nullify the whole of the regime.
In short, if the export control regime is used to license the release of open-weight AI models, its legality will likely hinge on whether model weights are speech.
That is dangerous territory. As a matter of policy, does the government want to risk losing? And does the public want to risk the government winning? Neither outcome is appealing. A loss would remove the government’s ability to regulate the release of model weights at all in the near term and might smother any energy Congress had to create a regime that can survive a First Amendment challenge.
A win for the export control regime on the grounds that model weights are not speech would be worse. Formally, the government would have won only the power to require a license for model weights. But in a practical sense, the guardrails of the First Amendment are meant to protect the public against the unbridled exercise of government authority. They are a pillar of liberal values, and speech is just one instance in which those values should be applied. The regulation of AI model weights, an emergent technology of great promise and peril, will sorely miss those guardrails regardless of the answer to the legal question of whether the weights are speech.
Indeed, the export control regime—if legal—would allow restrictions on model weights to be imposed quickly, without published standards, stated reasons, or judicial check. A lack of judicial review means that it will be difficult for AI consumers or model providers to challenge the government’s rulings. No enforceable deadlines for review mean that the government could hold up the release of AI model weights indefinitely. And no requirements for justification mean that the government could arbitrarily restrict the release of model weights. The government will not necessarily abuse this lack of guardrails. But the growing calls for AI regulation, and uncertainty about the optimal level of it, mean the government might.
Abuse of the export control regime would have consequences: It could tilt the regulatory scales in favor of closed-weight model providers who—because they never distribute their weights—will not be subject to restrictions as readily. That in turn would increase the concentration of market power in the hands of the extant closed-weight AI labs. If for even a year or two export controls effectively block model weight distribution while closed-weight model providers can continue providing inference, the momentum in the U.S. behind open-weight models could be crippled and the dominance of closed-weight model providers would be entrenched. And some users would turn not to closed models at all, but to foreign, likely Chinese, open-weight alternatives beyond the regime’s initial reach. There is also the risk of regulatory capture. The export control regime being upheld as a tool of model weight regulation would reduce near- and medium-term pressure to develop alternatives. As noted, closed-weight model providers may become the economic beneficiaries of such a model weight licensing structure. They will be in a strong position to discourage the development of any new regime that would replace the existing, legal, one.
As a tool of foreign policy, the export control regime understandably lacks the procedural safeguards that the First Amendment demands of regimes that license speech, allowing the government to take quick and often adversarial actions against foreign adversaries. Yet employing that same regime to determine whether Americans themselves can access one of the most important technologies of our time is perverse.
A Better Way
The regulators who may soon license model weights run the risk that their regime will be struck down by a court as illegal. And the opponents of regulation—for example, the more than 270 signatories to a recent industry open letter in favor of open-weight models—risk being saddled with the potentially oppressive exercise of the export control regime if they fail to articulate a better vision.
Incorporating the guidance of the First Amendment could resolve each of these dilemmas. Adherence to its principles will lower downside legal risk in the event that model weights are speech. It would also place guardrails around the exercise of regulatory authority over model weights. It would require the government to provide greater transparency, issue determinations in a timely manner, and offer judicial recourse to those model developers who feel their license applications have been unjustly denied.
While it is undeniable that even such moderated regulation may be burdensome, it will diffuse immediate pressure for hasty and damaging government action. And just because such authorities exist does not mean they must be employed to their maximal extent. A regulatory agency could tighten or loosen them as needed after they are created.
How might the government incorporate these First Amendment principles? One option would be to build First Amendment safeguards into the export control regime itself. BIS could for example amend the EAR for model-weight licenses by sunsetting denials, requiring written criteria for them, imposing deadlines on internal appeals, and closing § 750.4’s referral-to-the-president loophole.
However, § 4821(a) of the Export Control Reform Act (ECRA)—the statutory source of the EAR—exempts EAR functions from “sections 551, 553 through 559, and 701 through 706 of title 5,” that is, the federal rulemaking process and judicial review. That is legally unfortunate, as judicial review is a requirement of Freedman and FW/PBS. That also has unfortunate policy implications. It means that BIS would not face any judicially cognizable consequences if it were to breach its self-imposed safeguards, absent action so “blatantly lawless” as to contravene a clear and specific statutory command. That is a high standard to meet. The U.S. Court of Appeals for the D.C. Circuit confirmed as much in FedEx Corp. v. United States Department of Commerce, holding that § 4821 forecloses ordinary judicial review of BIS licensing actions and leaves only an ultra vires channel so narrow that the court described it as “essentially a Hail Mary pass.” Self-imposed safeguards by BIS for AI model weight licensing would certainly be welcome, but BIS’s structural limitations mean they arguably do not go far enough.
An alternative would be to build statutory infrastructure to license model weights. But how would Congress create this authority? To be legally robust, the coverage of the regime would need to be content-neutral and would need to incorporate the guardrails described above, up to Freedman: A covered developer would file a prepublication notice, and the reviewing agency would be required to render a decision by a deadline. If it failed to act before the deadline, the license would be issued automatically. Crucially, there would be judicial review; unlike the export control regime, the entire apparatus would be subject to the APA. The denial of a license would obligate the government to file suit in federal court within a set number of days and to bear the burden of justifying the restraint there. Expedited scheduling and a prompt judicial decision would be a requirement. The new regime would also publish a redacted justification of its decision, and its denials would sunset absent fresh findings.
A regime of this nature would stand on firmer legal footing, with procedural guardrails that exceed those required for content-neutral regimes. But more importantly, these guardrails would serve to meet calls for regulation while being light enough to reduce its harmful effects.
There are of course objections to this statutory solution. The clearest is China’s own lead in the provision of model weights: Any truly effective licensing regime will need to identify ways to address the model weights emanating from the other side of the Pacific. A more disquieting objection is to the licensing model of AI regulation itself: It may be too brittle a medium for a technology as transmissible as artificial intelligence; better mechanisms will take time to develop. A final objection is realism: Will Congress marshal the energy to create this new authority? It may not, leaving the export control regime as the most plausible means by which to quickly impose regulations on model weights.
Yet these objections are to the vehicle of regulation. They do not undermine the spirit of a First Amendment-consistent approach. Recognizing that, if BIS licenses model weights it should at least adopt the more modest First Amendment safeguards that are within its remit, imperfect as they may be.
Conclusion
How the U.S. will regulate AI model weights is one of the most important questions of our time. Yet the regime that is readily available to license AI model weights—the export control regime—has major legal and practical pitfalls. These pitfalls are uniquely cured by adherence to the principles of the First Amendment. It would be wise to at least incorporate those principles into the export controls that may be imposed on model weights. And it would be even wiser to incorporate them into the regulatory regimes for AI that are yet to be conceived.
