Powers, not promises

The interesting thing about a regulator is never the fine. It is the list of things it is allowed to do before the fine — and that list turns out to include running its own tests.

The AI Office may request technical documentation, evaluate a general-purpose model itself, require corrective measures, and impose penalties. Anthropic and OpenAI are named among firms inside that scope.

Read the list in order

Documentation is discovery. Evaluation is independent measurement. Corrective measures are instructions. Penalties are money. Only the last of those is what people mean when they say enforcement, and it is the least consequential of the four.

A fine is a cost you can budget. A corrective measure is an instruction about how your model behaves.

The evaluation power is the structural one

Public assurance about frontier models has rested almost entirely on system cards — documents written by the party being assessed, describing tests it chose and ran. That is not a scandal; it is simply the only mechanism that existed.

It is also the mechanism that came under direct criticism this month. Twenty-nine researchers across eighteen organisations have now published a consensus list of what interpretability cannot yet do, and a separate critique argued pre-deployment assessments may provide weaker assurance than their system cards imply.

An authority that can run its own evaluation does not need to trust the card. That is a change in who holds the pen, and it is the first time anyone outside a lab has held it.

Compare the instruments

Washington is reportedly dictating access to frontier models — no statute, no licence, no published mechanism. Brussels wrote its powers down, named the office, and published the penalties.

Both jurisdictions are constraining frontier AI. Only one of them can be read, and only one of them can be complied with deliberately.

What this asks of you

If you provide a general-purpose model into Europe, the operative question stopped being what does our system card say and became what survives someone else's test. Those are different engineering programmes, and the second one is longer.

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 → · Latham & Watkins — AI Act Update: EU Resolves to Change Rules and Extend Deadlines →