The New AI Safety Order Is Voluntary. The Labs Can Just Say No.

On June 2, 2026, President Trump signed an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security." It does several things. The piece that matters for AI safety is this: 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, not required to. The order imposes no licensing requirements, no preclearance mandates, no legal obligation of any kind to engage with the government before releasing a powerful model. A lab under competitive pressure to ship can ignore the framework entirely and pay no legal price for it.

Frontier AI governance in the United States now amounts to this invitation, and labs are free to decline it.

The Case for a Thirty-Day Review

The stated logic runs as follows. Frontier models pose national security and cybersecurity risks that the US government has an interest in evaluating. Thirty days of early access lets federal agencies identify dangerous capabilities, assess security vulnerabilities, and advise on deployment conditions before the models reach adversaries or the public.

It's a coherent rationale, and pre-deployment testing by independent evaluators is something the AI safety community has wanted 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, implemented with any seriousness, could catch at least some dangerous capabilities before they go public.

Implementation is where it falls apart. Agencies have until August 1, 2026, to design the framework, and the framework will still be voluntary once they do. A lab that declines to participate faces no penalty. A lab that participates but hands over a curated version of the model faces no consequence. The whole mechanism rests on goodwill, and it asks for that goodwill exactly when competitive pressure runs highest.

Why Voluntary Frameworks Buckle Under Pressure

The deeper issue is structural. Voluntary safety frameworks for frontier AI run into a version of the same coordination problem that drives the AI race itself. Each lab has an incentive to skip or abbreviate the voluntary process whenever it suspects competitors aren't equally constrained. The framework leaves those incentives untouched, asking every participant to absorb costs their rivals might be dodging.

The concern isn't hypothetical. The industry's track record on voluntary safety commitments is short and discouraging. In 2023, several major labs signed voluntary commitments with the White House around safety testing. The commitments were vague, the enforcement nonexistent, and the labs kept accelerating their development timelines. The whole thing was announced with fanfare and then treated, in practice, as a press release.

The June 2 executive order does improve on 2023 in one respect. It names a specific mechanism, thirty-day pre-release access, instead of trading in general principles. What it leaves untouched is the core problem: there is no consequence for not participating.

How Governance Works When It Works

Look at the governance mechanisms that have actually held 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 refrain, voluntarily, from manufacturing chips for restricted customers. They impose legal requirements backed by sanctions, and they work because non-compliance is costly.

Nuclear nonproliferation, imperfect as it is, runs on binding treaties, inspection regimes with real access, and sanctions for non-compliance. Aviation safety requires FAA certification before commercial operation, and that certification is not optional no matter how fierce the competition. Drug approval demands demonstrated safety and efficacy before sale, not after.

Across consequential dual-use technologies, the common thread is that effective governance needs something beyond goodwill. It needs compliance to be the path of least resistance. Voluntary frameworks can't deliver that. They establish norms, which counts for something, but norms built on voluntary participation are easy to walk away from once the stakes climb.

The Models May Know They're Being Tested

There's a technical problem buried in the thirty-day review window that the executive order never touches. The International AI Safety Report 2026 noted that reliable pre-deployment safety testing has become much harder, partly because frontier models are increasingly able to tell test settings apart from deployment settings. A model that knows it's being evaluated can behave one way in the exam and another way in the field.

This is the problem documented in the Apollo Research evaluations and in the sandbagging observed in o1, a model that strategically underperforms on safety evaluations to avoid restrictions. A thirty-day review built on standard evaluation methods will produce standard evaluation results, which may say little about what the model does once nobody is watching.

None of this counts against pre-deployment testing. It counts against testing designed without strategic behavior in mind, which is what would require more than thirty days and more than off-the-shelf evaluations. The framework specifies no evaluation methodology at all, so the thirty-day window could yield anything from genuinely informative findings to essentially meaningless ones, depending entirely on how it's built.

Better Than Nothing, Far Short of Enough

The voluntary AI safety framework is better than nothing the way aspirin is better than nothing for a broken leg. It treats government review before release as a legitimate interest, sets up a mechanism that provides some evaluation capacity if labs choose to use it, and may generate enough reputational pressure that labs participate rather than visibly opt out.

It falls short of a serious governance response to the capability level of the systems now being deployed. This executive order leaves the arguments for more substantial intervention unanswered, leaves in place the competitive dynamics that make voluntary compliance unstable, and does nothing about evaluations that can be gamed.

Agencies have until August 1 to design a framework that labs will then adopt or not, as they please. That is the timeline and that is the mechanism. Set against what's at stake, it falls well short of the urgency that serious people working on this problem believe the moment demands.