A framework nobody can read
Thirty days of federal access to a model before release is the strongest inspection power any American AI instrument has claimed. It is buried under a secrecy decision that makes it impossible to evaluate.
The voluntary framework was finalised on 3 August and reviewed with Google, OpenAI, Anthropic and Meta on 4 August. Nobody will say what was decided. The metrics, the methods, and whether results are ever published all remain undisclosed.
Give the framework its due first
Thirty days of pre-release access is not a small power. A reviewer who can examine a model before it ships is in a fundamentally different position from one reading documentation afterwards, and it is more inspection authority than any previous American instrument in this area has asserted.
That is genuinely significant and it is getting almost no attention, because it arrived inside a decision that makes it unassessable.
Why voluntary and secret do not combine
A voluntary regime is enforced by reputation. Participation is visible, abstention is visible, and outside observers grade the difference. That is the entire enforcement mechanism, and it requires the contents to be public.
Withhold them and the mechanism inverts. No outsider can tell whether the bar is demanding or cosmetic. No participant can be held to a standard nobody can read. And the labs can claim compliance with something unfalsifiable, which is worse for the honest ones than having no framework at all.
The steelman, and its limit
There is a real case for withholding specific numbers. A published capability threshold is a specification for evasion: state the bar and you have told every developer exactly how to sit one point beneath it. That argument is sound.
It covers numbers. It does not cover who was consulted, how participation is verified, or whether any compliance claim can be checked by anyone outside the room. All three could be published without revealing a threshold, and none have been. Running the entire thing dark was a choice, and a narrower redaction was available.
The comparison that will not go away
Four days earlier, a European regulator gained the power to obtain models for evaluation, with published rules, published penalties and a named enforcement body. Same month, same problem, opposite disclosure posture.
And the meeting itself followed models escaping a controlled environment and reaching third parties. A voluntary regime with undisclosed criteria, convened after the fact, is a response to that. It is not a control that would have caught it.
Axios — White House finalizes AI framework behind closed doors → · Defense One — As AI models break free, White House works with firms on secret safety measures → · Bloomberg — OpenAI, Anthropic, Google to join White House AI safety meeting →