The AI Office can demand the documentation, evaluate the model, and fine you
The enforcement powers over general-purpose models are broader than the first fines suggested: request technical documentation, run its own evaluations, require corrective measures, issue penalties. Anthropic and OpenAI are named among the firms now inside that scope.
Yesterday the interesting fact was that money had changed hands. Today the interesting fact is what the AI Office is actually permitted to do, and it is a longer list than a fine.
Over general-purpose AI models the Office may request technical documentation, evaluate the model itself, require corrective measures, and impose penalties for non-compliance. Read those four powers in order and the third one is the heavy one. A fine is a cost. A corrective measure is an instruction about how your model behaves.
Reporting names Anthropic and OpenAI among the firms facing new scrutiny under these powers. That matters because general-purpose obligations were the part of the Act that applied to model providers rather than to deployers, and it settles an open question about whether a US lab serving Europe is inside the regime. On this reading it is.
The evaluation power is the one to watch technically. An authority that can run its own tests does not depend on a provider's system card, and system cards have been the entire basis of public assurance about frontier models to date. A critique published this month argued those assessments may provide weaker assurance than they imply. An independent evaluator changes who gets to decide.
Set against the American position and the divergence is now about instruments rather than intent. Washington built a review with no gate. Brussels built an office that can ask for your documents and then test the thing itself.
CNBC — Anthropic, OpenAI among firms facing new scrutiny under EU AI Act enforcement powers → · European Commission — Commission starts enforcing AI Act rules and new transparency requirements on 2 August → · European Commission — Safer and more transparent AI →