Two regimes, one week, one publishes

Brussels took the power to obtain a model and test it, and printed the rules. Washington took thirty days of early access, wrote in that it can never become licensing, and kept the details private.

From 2 August the EU AI Office can investigate and enforce against general-purpose model providers, with penalties reaching €15 million or 3% of worldwide turnover.

The power that matters is access

A regulator that can only read documents depends on the accuracy of those documents. Every claim it assesses was written by the party being assessed, in language that party chose. A regulator that can obtain the model and test it can check independently — and almost no technology regime anywhere holds that authority.

The one-year adjustment period was also honoured exactly. Obligations landed August 2025, enforcement powers activated August 2026, on schedule. In a field where regulatory timelines slip routinely, a deadline arriving on time is information about how seriously the thing is being run.

The American design, read carefully

The voluntary framework allows up to thirty days of pre-release federal access and explicitly cannot become a mandatory licensing or preclearance system.

Thirty days of pre-release access is genuine inspection authority — more than any prior US instrument in this area has claimed. The no-licensing clause is what stops it becoming an approval gate. Both halves are deliberate.

The result is a reviewer who may look and may not act. Whether anything follows from a finding depends entirely on the developer choosing to respond, and no published mechanism describes what happens if a thirty-day review surfaces something serious.

Why the secrecy is the binding problem

A voluntary regime is enforced by reputation. Participation is visible, abstention is visible, outsiders grade the difference. Withhold the contents and that mechanism stops working: nobody can judge whether the bar is demanding, and no participant can be held to a standard nobody can read.

There is a real argument for withholding specific capability thresholds — publish the bar and you have specified how to sit beneath it. That argument covers numbers. It does not cover process, participants or compliance criteria, and all three are dark. Meanwhile the incidents that prompted the meeting were disclosed voluntarily, after the fact, by the companies involved.

European Commission — Commission starts enforcing AI Act rules and new transparency requirements on 2 August → · Bloomberg — OpenAI, Anthropic, Google to join White House AI safety meeting → · Help Net Security — EU begins enforcing AI Act, putting AI models under the microscope →