As of August 2, the European Commission’s AI Office can request information from any general-purpose AI provider, demand model access for evaluation, order corrective measures, and issue fines up to €15 million or 3% of global turnover. The obligations existed before. The regulator empowered to police them didn’t.

Article 50 flipped on the same day. Chatbots have to disclose they’re machines, deepfakes have to be labelled, and AI-generated content has to carry machine-readable provenance marks. Generative systems already deployed have until December 2, 2026 to implement marking and detection. GPAI models placed on the market before August 2, 2025 get until August 2, 2027 to comply. The AI Omnibus pushed out several high-risk deadlines earlier this year but left the August 2 GPAI switch intact, which itself tells you which lobbying arguments Brussels found persuasive and which it didn’t.

The early targets are legible. Reuters reports the EU is already in talks with OpenAI and Anthropic following recent cyber incidents involving their models; OpenAI has confirmed contact with the AI Office. Signatories to the voluntary Code of Practice on Transparency get a presumption of conformity and a gentler posture. Non-signatories don’t.

“A U.S. address does not put a lab outside the EU regulator’s reach,” said Elisabetta Righini, a partner at Sidley Austin, who noted that refusing an information request or blocking a model evaluation is “fineable on its own.” The AI Office says “technical compliance dialogues” remain its preferred first tool.

The political overlay is already visible. MEP Michael McNamara warned that Washington may treat any high-profile action against a U.S. lab as an attack on American commercial interests. Brussels built the instrument during a period of transatlantic regulatory alignment. It’s switching it on during a period that no longer resembles that one.

Sources

Sources