Juan Londoño
After weeks of being shut down due to safety concerns (via questionable export controls), Anthropic’s Claude Fable is once again accessible to the public with new guardrails that limit its capabilities and revert to inferior models more frequently. OpenAI’s ChatGPT‑5.6 has had a delayed rollout to fulfill a reported request from the White House. These models have been evaluated, withheld from the market, and re-released. But no one knows what made them too dangerous to release, and no one knows what would make them safe enough to release. The obscure and seemingly arbitrary standards of the current evaluation process established by the White House’s executive order on Frontier Artificial Intelligence (AI) show that, while it did not formally introduce an “FDA for AI,” it might have done so in practice. In fact, this regime shares most of the common pitfalls of a vetting regime and, in some cases, might worsen them.
The pitfalls of a vetting regime for AI are clear. Granting a new regulatory agency or an existing agency in the executive branch the power to decide when a model is considered safe enough for release will likely lead to regulatory capture, politicization, and legal uncertainty. In a similar way to how the Food and Drug Administration (FDA) could randomly impose additional reviews on drugs for clearly political motives, its AI equivalent would likely punish politically disfavored companies, hurting their employees and investors. The risk of retaliation is particularly high for a product that could be used to create expressive content, such as AI.
The White House’s executive order on frontier AI supposedly set up a voluntary evaluation process for frontier AI models that, according to the text, should not be interpreted as a pre-release vetting or licensing regime. However, once Anthropic decided to expand access to its Fable and Mythos models, the White House responded by invoking export control powers that would prevent foreign nationals from accessing these models, even if they resided in the US. As Anthropic itself admitted, enforcing this mandate was so complex that there was no choice but to shut down the models. Before this incident, there was no indication that these export controls could be invoked, and Anthropic and the AI industry at large were caught off guard.
This episode clearly draws out the issues with the current evaluation regime: It is seemingly arbitrary; the industry has no clear and public standard over what would make a model “safe” or “dangerous”; and there is no clarity over what the executive can and cannot do to models it deems to be too dangerous to release. This uncertainty, as seen in recent examples, can disrupt not only AI companies but also businesses worldwide that rely on these products.
The Grave Pitfalls of the White House’s Approach
To be clear, heavy-handed pre-approval processes, like those used by the FDA, are deeply flawed and ultimately an undesirable approach for governing fast-moving AI. However, it is an established process that must make decisions on technical grounds tied to specific risks rather than broad concerns about perceived capacity or computing power. This guarantees due process, provides a degree of predictability in outcomes, and ensures appropriate congressional oversight. All of these are currently lacking in the White House’s ad hoc approach.
Another major concern with the existing White House approach is the lack of guardrails to curtail the significant potential for abuse and lack of oversight that have manifested in recent situations with frontier models. Even under a prescriptive (but congressionally established) vetting regime, decisions to enact AI export controls or other restrictions on frontier models would have to be accompanied by publicly available reasoning and appropriate opportunities for industry and scientific critique and appeal. This provides additional mechanisms to scrutinize the process and limit potential abuse. Such guardrails incentivize at least some degree of objective and factually sound review. Additionally, what can and cannot be done is detailed in the congressional delegation that governs it.
The current White House AI evaluation process has little to none of that. Anthropic needed to hold private talks and negotiations with the White House to have the export controls lifted. The claims about the Fable jailbreak were raised by another private entity. And the public only learned about them through media reporting.
Anthropic claimed to have worked with government officials to test the model and to have received assurances that it was safe, but once export controls were imposed, the government’s justification was based on verbal evidence, with no report of an expert-led evaluation process. The government’s punishment (the export controls) was not previously considered a potential policy tool by the White House, and the industry could not reasonably have expected that they would be invoked.
The current frontier AI evaluation system, while technically voluntary and touted as “light touch,” has resulted in a licensing system in disguise. The Fable restriction and reintroduction have shown that companies that do not submit to these pseudo-voluntary tests will likely face retaliatory action from the government. To make matters worse, the informal nature of this scheme makes the process unnecessarily obscure, untraceable, and essentially undisputable, making it somehow worse than a formal licensing process.
In the current status quo, the administration needs only a legal or policy tool—such as a supply chain risk designation or export controls—that it can invoke to retaliate against an AI company when it wishes. Supported by spuriously defined criteria and few or no checks on these powers, the White House essentially holds free rein over the frontier AI industry with no mechanism or recourse for other entities or branches to question the White House’s decisions. This grants the executive branch a tremendous amount of power that will be ripe for abuse or subjective cronyism.
Conclusion
This state of affairs makes it more urgent for Congress to step in and establish a statutorily bound regime that provides clarity for all parties involved, respects the rule of law, and maintains the flexibility to allow innovative, industry-led development. For the US to be a global leader in frontier AI, it necessarily requires that governance of security concerns be guided by predictability, clarity, and objectivity. Doing so requires both a clear set of parameters that apply only to significant risks and clear constitutional checks that subject regulators to oversight and limit their power. Congress is already considering an alternative regime. But recent events show that it is perhaps more urgent to establish a process that can respond to legitimate concerns appropriately without burdening the industry or undermining the competitiveness of US technology.
