The AI Act is live, and nobody knows who checks first

The AI Act's enforcement date arrived on 2 August with remarkably little drama, which is itself the story. Two years of lobbying, redrafting and trade-press panic produced a set of obligations that, on the day they became enforceable, most serious vendors had already implemented. The disclosure requirements are not the hard part. The enforcement architecture is.

Chatbots must say they are chatbots. Deepfakes must be labelled. Generated or altered content must carry machine-readable marks. Any competent product team can ship all three in a sprint, and most already had — the disclosure line at the bottom of a chat window has been standard practice since well before it was law.

The one with teeth is the machine-readable marking. A visible label survives until someone screenshots it; an embedded provenance signal is supposed to survive the pipeline. Making that work across formats, platforms and re-encoding is an unsolved engineering problem dressed as a compliance requirement, and 'reasonable effort' is doing a great deal of quiet work in how vendors intend to satisfy it.

Enforcement runs through national market surveillance authorities in every member state, plus the European AI Office for general-purpose AI, plus the European Data Protection Supervisor for EU institutions. That is twenty-seven regulators with different budgets, different backlogs and different views on what adequate looks like, coordinating through a central office that has just hired around forty people.

GDPR is the precedent worth studying, and it is not encouraging on timelines. Years of variable enforcement, forum shopping towards the friendliest regulator, and a long wait before anybody could tell you what the rules meant in practice. The AI Act's designers clearly learned something — the central AI Office exists precisely to hold the line on general-purpose models — but the same asymmetry of resources applies.

More than the quiet launch suggested. In September the AI Office and national data protection authorities began initiating technical audits on Article 11 technical files. By 15 September, providers of general-purpose models trained above 10^25 FLOPs were due to file their first formal systemic risk evaluations, covering red-teaming methods, energy disclosures and copyright training summaries.

That copyright training summary line deserves attention it has not received. Every frontier lab operating in Europe now has a filing obligation about training data provenance, at a moment when several are in active litigation over exactly that question elsewhere. Documents produced to satisfy a European regulator are documents that exist, and discovery is a thing that happens in other jurisdictions.

There is also a competitive angle the compliance conversation keeps missing. Filing obligations scale badly for small providers and barely at all for large ones. A frontier lab with a policy team absorbs a systemic risk evaluation as routine work; a twelve-person European startup shipping a general-purpose model treats the same filing as a quarter of its engineering capacity. Every regime like this quietly advantages incumbents, and the SME proportionality language in the penalty regime addresses fines rather than the paperwork that precedes them. Whether the AI Office calibrates for that in practice will matter more to the European AI industry than the fine ceiling ever does.

The honest summary: the compliance burden of 2 August is modest, the enforcement risk is unevenly distributed and mostly unknown, and the real exposure sits in the disclosure obligations around training data rather than the chatbot labels everyone wrote about. The next date is 2 December 2026, when new prohibited practices land — including a nudifier ban drawn broadly enough to catch general-purpose tools marketed without reasonable safeguards. That one will generate the first real fight.