Tomorrow Brussels Stops Asking Nicely: The EU AI Act Grows Teeth

For a year, the toughest rules in the EU AI Act existed only on paper — binding in law, but with no one able to enforce them. Tomorrow that changes. Here is what actually becomes enforceable on 2 August, why the quiet transparency clock matters more than the fines, and the one line most weekend coverage will get wrong.

What Actually Switches On

On 2 August 2026, the European Commission's AI Office gains real enforcement power over providers of general-purpose AI models for the first time. The obligations themselves are not new — they have applied in law since 2 August 2025. What arrives tomorrow is consequence. From this date the AI Office can request technical documentation under Article 91, run its own technical evaluations of a model under Article 92, order compliance and risk-mitigation measures under Article 93, restrict or withdraw a model from the EU market, and impose fines of up to 3 percent of global annual turnover or 15 million euros, whichever is higher, under Article 101.

For twelve months this was a dormant toolkit. Brussels could publish guidance and it could persuade, but it could not compel, and a gap between a rule that exists and a rule that can be enforced is a gap every legal department learns to live inside. Tomorrow that gap closes. The switch moves from persuasion to compulsion, and the models that came to market on or after 2 August 2025 are fully in scope. Older models get until 2 August 2027 to fall in line — a two-tier reality that quietly splits the market into the already-accountable and the not-yet.

The Part Most Coverage Misses: Article 50

The GPAI enforcement powers take the headlines, but a second clock starts on the very same day and reaches far more of ordinary life. Article 50 becomes applicable on 2 August 2026. At its core it requires that anyone deploying a chatbot or conversational assistant tell you, in plain and accessible language and at the start of the interaction, that you are dealing with a machine. It requires that synthetic and manipulated media be marked as artificially generated.

Read the fine print, though, because the Omnibus adjusted the timing. The duty to inform people they are talking to an AI is live tomorrow. But the machine-readable marking of synthetic content for systems already on the market before that date carries a grace period to 2 December 2026. So the honest framing is not that every deepfake in Europe gets labelled overnight. It is that the legal obligation to disclose and to watermark now formally exists, with the watermarking of legacy systems phased in over the following four months. That is still a profound shift — it touches every customer-facing assistant, every AI voice line, every generated image passed off as real — and it carries the same penalty ceiling as the GPAI rules.

The Omnibus Confusion

Expect a wave of "it all got delayed" commentary this weekend. Read it carefully, because it conflates two different things. The Digital Omnibus on AI, given final Council approval on 29 June 2026 and published in the Official Journal on 24 July, does defer the heavy high-risk obligations under Annex III out to 2 December 2027, and high-risk systems embedded in regulated products out to 2 August 2028. That delay is real, and for anyone building education, hiring, credit-scoring or biometric systems it is a genuine sixteen-month planning window.

But the Omnibus did nothing to the GPAI enforcement powers or the Article 50 transparency duties landing tomorrow. It also added new prohibitions — on AI systems designed to generate non-consensual intimate imagery and child sexual abuse material — with compliance required by 2 December 2026. Anyone who read "high-risk delayed" and filed the whole Act under "problem for 2027" has read the wrong line, and tomorrow is when that misreading starts to cost money.

Claim and Counter-Claim

The optimistic reading, voiced loudly in Brussels, is that Europe now holds the one lever the United States has set down: the ability to inspect a model, demand its documentation, and pull it from the market. Where Washington shifted from regulation toward ownership — a state stake in OpenAI, access-gating debated down to the power grid — Europe kept the referee's whistle and just learned to blow it.

The sceptical reading is equally grounded. Power on paper is not power in practice. The AI Office is a young institution, the models are vast and opaque, and evaluating a frontier system under Article 92 is a real technical undertaking, not a filing exercise. Signatories of the voluntary GPAI Code of Practice are to be treated as acting in good faith through the transition, which softens the first year further and gives the largest labs a ready shield. The calibrated verdict: tomorrow changes the Commission's legal standing far more than it changes any lab's behaviour by breakfast. The teeth are real. Whether Brussels bites, against whom, and how hard, is the story of the next eighteen months — not of tomorrow morning.