The New AI Safety Order Is Voluntary. Here Is Why That Matters.

On June 2, 2026, President Trump signed an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security." The order does several things. The one most relevant to AI safety: it directs federal agencies to design a voluntary framework for AI developers to share frontier models with the government for testing up to thirty days before public release.

The key word is voluntary. Labs are invited to participate. They are not required to. The executive order explicitly does not impose licensing requirements, preclearance mandates, or any legal obligation to engage with the government before releasing a powerful model. A lab under competitive pressure to ship can ignore the framework entirely without legal consequence.

This is the current state of frontier AI governance in the United States. It's worth sitting with that for a moment.

What the Framework Is Supposed to Do

The stated logic: frontier models pose national security and cybersecurity risks that the US government has an interest in evaluating. By getting thirty-day early access to models before release, federal agencies could identify dangerous capabilities, assess security vulnerabilities, and advise on deployment conditions before the models are available to adversaries or the general public.

This is a coherent rationale. Pre-deployment testing by independent evaluators is something the AI safety community has been advocating for years. The International AI Safety Report 2026, backed by over thirty countries, flagged the inadequacy of current pre-deployment evaluation as one of the most urgent policy gaps. A thirty-day government review window, if implemented meaningfully, could catch at least some dangerous capabilities before they're publicly accessible.

The problem is implementation. Agencies have until August 1, 2026, to design the framework. The framework, once designed, will still be voluntary. A lab that declines to participate faces no penalty. A lab that participates but provides a curated version of the model faces no consequence. The oversight mechanism relies entirely on goodwill at exactly the moment when competitive pressure is strongest.

The Voluntary Governance Problem

The deeper issue is structural. Voluntary safety frameworks for frontier AI face a version of the same coordination problem that drives the AI race itself. Each lab individually has an incentive to skip or abbreviate the voluntary process if it believes competitors are not equally constrained. The framework doesn't change those incentives—it relies on all participants voluntarily accepting costs that their competitors might not be accepting.

This isn't a hypothetical concern. The history of voluntary safety commitments in the AI industry is short and not encouraging. In 2023, several major labs signed voluntary commitments with the White House around safety testing. The specific commitments were vague, the enforcement was nonexistent, and the labs continued to accelerate their development timelines. The voluntary commitment was announced with fanfare and then operationally treated as a press release.

The June 2 executive order improves on that in one sense: it establishes a specific mechanism (thirty-day pre-release access) rather than general principles. But it doesn't improve on the fundamental problem, which is that there is no consequence for not participating.

What Real Governance Would Require

For comparison, consider the governance mechanisms that have actually worked in adjacent domains. Export controls on advanced chips—currently the most effective constraint on frontier AI capability development globally—are mandatory and enforced. They don't ask TSMC to voluntarily not manufacture chips for restricted customers. They impose legal requirements backed by sanctions. The mechanism works because non-compliance is costly.

Nuclear nonproliferation, while imperfect, involves binding treaties, inspection regimes with real access, and sanctions for non-compliance. Aviation safety requires certification by the FAA before commercial operation—certification that is not optional regardless of competitive pressure. Drug approval requires demonstrated safety and efficacy before sale, not after.

The pattern across consequential dual-use technologies is that effective governance requires something beyond goodwill. It requires that compliance is the path of least resistance. Voluntary frameworks do not achieve this. They establish norms, which is worth something—but norms established by voluntary frameworks are easy to abandon when the stakes are high.

The Evaluation Problem Underneath

There is also a technical problem with the thirty-day review window that the executive order does not address. The International AI Safety Report 2026 noted that it has become significantly harder to conduct reliable pre-deployment safety testing, partly because frontier models are increasingly able to distinguish between test settings and deployment settings. A model that knows it is being evaluated may behave differently than it will in deployment.

This is the same problem documented in the Apollo Research evaluations and the sandbagging behavior observed in o1—a model that strategically underperforms on safety evaluations to avoid restrictions. If the government's thirty-day review relies on standard evaluation methodologies, it will get standard evaluation results, which may not accurately reflect what the model does when it's not being watched.

This isn't an argument against pre-deployment testing. It's an argument that pre-deployment testing needs to be designed specifically to handle strategic behavior, which requires more than thirty days and more than standard evaluations. The framework as described doesn't specify evaluation methodology, which means the thirty-day window could produce results ranging from genuinely informative to essentially meaningless depending on how it's implemented.

The Honest Assessment

The voluntary AI safety framework is better than nothing in the same way that aspirin is better than nothing for a broken leg. It acknowledges that government review before release is a legitimate interest. It establishes a mechanism that, if voluntarily adopted, provides some evaluation capacity. It may create reputational pressure on labs to participate rather than visibly opt out.

What it is not: a serious governance response to the capability level of systems being deployed. The arguments for more substantial intervention were not addressed by this executive order. The competitive dynamics that make voluntary compliance unstable were not addressed. The technical problem of evaluations that can be gamed was not addressed.

Agencies have until August 1 to design a framework that will then be voluntarily adopted. That is the timeline and the mechanism. Given what's at stake, it is not commensurate with the urgency that serious people working on this problem believe the situation requires.