EU AI Act Enforcement Has Started

On 2 August 2026 the EU AI Act got its referee. Most of its rules were already law. The powers to investigate them, order changes, and issue fines were not. Authorities can now act on breaches for the first time. And the Commission has put a public complaint form online, so anyone who thinks your AI system breaks the rules has an easy, free route to telling a regulator about it.

If your current exposure is Article 50 transparency, 2 August is when the disclosure duties started applying, and when the machinery to enforce them switched on.

What actually changed

Article 113 says the Regulation “shall apply from 2 August 2026”, with a short list of chapters carved out to arrive earlier. Chapter IX is not on that list, and Chapter IX is the market surveillance chapter: investigation powers, corrective-measure powers, the AI Office’s supervision of general-purpose models under Articles 88 to 94. Article 101, which lets the Commission fine a general-purpose model provider up to €15 million or 3% of worldwide annual turnover, is carved out of the earlier date explicitly. Both started applying on 2 August 2026.

Obligation	Applied from	Enforceable from
Article 5 prohibitions	2 February 2025	2 August 2026
Article 4 AI literacy	2 February 2025	2 August 2026
General-purpose AI model rules	2 August 2025	2 August 2026
Article 50 transparency	2 August 2026	2 August 2026
Annex III high-risk obligations	2 December 2027	2 December 2027

Until now, the Article 5 prohibitions had been law with no market surveillance framework behind them. Nothing the rules require changed on 2 August; what changed is that a regulator can now act on them, using penalty ceilings that were always in the text.

Article 50 affects most businesses now

The high-risk obligations moved to 2 December 2027. Article 50 didn’t move, and it is the piece that catches ordinary customer-facing software. From 2 August 2026 you must tell people when they are dealing with an AI system rather than a person, disclose deep fakes, disclose exposure to emotion recognition or biometric categorisation, and mark synthetic output in a machine-readable way.

2 December 2026 is not a general marking deadline, though it is widely reported as one. Article 111(4), added by the Digital Omnibus, is a transitional provision for providers of generative systems “placed on the market before 2 August 2026”, who get until 2 December 2026 to comply with Article 50(2). Anything placed on the market from 2 August 2026 onward has been in scope since that day, so a new generative feature shipping this quarter has no grace period.

On 20 July 2026, the Commission adopted its final guidelines on Article 50, and roughly 190 organisations have signed the Code of Practice on Transparency of AI-Generated Content. Adopting the Code is voluntary, and remains the cheapest way to show an authority you have thought about Article 50.

Who enforces what

Three bodies split enforcement, and which one you deal with depends on what you built.

Who	What they cover	What it means for you
AI Office (inside the Commission)	General-purpose AI models; systems where the system and its underlying model come from the same provider; AI built into platforms and search engines designated very large under the Digital Services Act (a Digital Omnibus addition)	Centralised, resourced, already operating. If you train your own model and ship a product on it, this is your regulator.
National market surveillance authorities	Every other AI system	If you build on someone else’s model, this is your regulator, in every Member State your system reaches. Intensity will vary by country.
European Data Protection Supervisor	AI used by EU institutions, bodies and agencies	Relevant if you sell into the institutions rather than the market.

The boundary that catches people is Article 75(1). Where an AI system is based on a general-purpose AI model and “the model and the system are developed by the same provider”, the AI Office holds the powers of a market surveillance authority over that system. So OpenAI’s own consumer products and Google’s own assistant features answer to Brussels. A company using those models through an API does not: it answers to its national authority.

The three ways a complaint reaches a regulator

Enforcement no longer depends on a regulator noticing you. There are three live channels, and two can be used by people you deal with every day.

1. The Complaint Tool, under Article 85. Article 85 gives “any natural or legal person having grounds to consider that there has been an infringement” the right to complain to the relevant market surveillance authority. The AI Office has now put that on the web as a form.

• It is not anonymous. Complainants must give identification and contact details.
• Out-of-scope complaints are referred on, with the complainant’s consent, to the relevant national authority or to one supervising fundamental rights.
• It does not cover matters outside the AI Act, other EU or national law, general-purpose model obligations under Articles 53 to 55, or Article 89(2) matters.

If your product is an ordinary customer-facing AI system built on a third-party model, the AI Office is not your supervisor, so a complaint about you gets routed to your national authority instead.

2. The Whistleblower Tool, under Article 87. Article 87 applies Directive (EU) 2019/1937, the EU whistleblower protection directive, to AI Act infringements and to the people who report them. The AI Office has built a secure channel for anyone professionally connected to a provider or deployer.

Read that as an internal risk. Your engineers know whether the chatbot discloses that it is a chatbot, and whether the watermarking claim on your sales page is true. They now have a protected channel to Brussels, and retaliating against them is unlawful.

3. The channel for companies building on someone else’s model, under Article 89(2). “Downstream providers shall have the right to lodge a complaint”, identifying the model provider and the provisions they say were breached. If you build on a third-party foundation model and that provider will not give you what Article 53 requires, this is your escalation route. If you are the model provider, it is how your customers can complain about you.

What this means for you

If you run a customer-facing chatbot or assistant. Your Article 50(1) disclosure is now enforceable, and it is the easiest breach in the Act for anyone to spot. A competitor can check your AI chatbot in ten seconds and file in five minutes.

If you publish AI-generated content. Deep fake disclosure and the labelling of AI-written text published to inform the public are live. The free EU icons published with the Code of Practice are the path of least resistance, and the editorial-control exception for reviewed text needs a named person holding editorial responsibility.

If you provide a general-purpose AI model. Article 101 is now available to the Commission. Your training-content summary and your copyright policy are the documents most likely to be asked for first.

If you deploy someone else’s AI. Your Article 26 duties are not enforceable until December 2027 for high-risk systems, but your Article 50 duties as a deployer are enforceable as of now. Deep fakes, emotion recognition notice and public-interest text all sit with you, not the provider. Article 86, the right to an explanation, sits in the same chapter but only binds you once the Annex III obligations land.

The delays nobody at the Commission wants to dwell on

The press release announcing all this contains one carefully worded sentence: effective enforcement “will also depend on Member States ensuring that national competent authorities are properly designated and adequately resourced.”

That is diplomatic language for a real hole. Article 70 required every Member State to designate a notifying authority and a market surveillance authority by 2 August 2025. Public tracking in June 2026 put the number that had designated both at 9 of 27. Six had designated nothing at all.

So your enforcement risk has two halves: the AI Office, staffed and directly competent over model providers and the platforms, and national enforcement of everything else, which is uneven and in some Member States non-existent.

Your obligation is not suspended by your Member State’s delay, and a complaint filed in a state without an authority does not evaporate. It sits in a queue that will get picked up when the authority appears.

What to do now

1. Check your disclosures against the final guidelines. The 20 July 2026 version is the one authorities will use, so if your Article 50 work was scoped against the March or May drafts, some of what you relied on has moved.

2. Find out who your market surveillance authority is in every Member State your system reaches. If there isn’t one yet, write down that you checked and the date.

3. Name your internal contact point. One person who receives regulatory correspondence, knows where the documentation lives, and can respond. A request for information with a deadline is a bad time to discover nobody owns this.

4. Test the obvious things a complainant would test. Does the chatbot disclose? Is the deep fake labelled? Does the disclosure appear at first interaction rather than three clicks into a policy page? Is it accessible to a screen-reader user?

5. Make your internal escalation route work before someone uses the external one. A concern raised internally and answered properly does not become a whistleblower report.

The Act has spent two years as a planning exercise judged against dates in the future. This month it is a live instrument with three complaint channels, a resourced central regulator, and fines that can be issued. The businesses caught out first will not be the ones with imperfect technical documentation. They will be the ones whose chatbot never said it was a chatbot.