EU AI Act Enforcement Begins: Fines Now on the Table
2 min readThe European Union’s AI Act crossed a hard line on August 2. The Commission can now investigate and fine companies over general purpose AI models, and the law’s full penalty regime is live. The rules have existed on paper since 2024. As of this week, they have teeth.
What the EU AI Act enforcement date actually turns on
Three things switched on at once. First, the Commission gained direct enforcement powers over general purpose AI, the category that covers frontier models from OpenAI, Google, Anthropic, Meta and anyone else shipping a model into the European market. Those labs have carried documentation, copyright policy and training data summary duties since August 2025, but Brussels had no mechanism to punish a failure. Now it does.
Second, the full penalty regime took effect, giving member state authorities the ability to levy fines rather than issue warnings. Third, the transparency obligations landed and were not postponed: chatbots have to tell users they are machines, AI generated media has to carry machine readable markings, and deepfakes have to be labeled.
The part that got pushed back
The picture is more complicated than a clean deadline. On June 16, 2026, the European Parliament approved amendments delaying the Act’s high risk obligations by 12 to 16 months. Standalone high risk systems listed in Annex III, which cover uses like hiring, credit scoring, education and biometric identification, now face a December 2, 2027 deadline instead of this week’s.
That is why analysts have been careful to frame August 2 as the day enforcement powers switch on, not the day most enterprises inherit new compliance work. Companies deploying high risk systems got breathing room. Companies building or shipping foundation models did not.
Why it matters
This is the first time any major jurisdiction has attached a working penalty mechanism to frontier model governance. Every other framework in play, including the emerging United States approach, is still voluntary or under negotiation. Whether the Commission actually opens a GPAI investigation in the coming months is the signal to watch. Enforcement authority that goes unused sends a very different message than a first case.
The practical near term test is the transparency rules, since they are visible to ordinary users. Expect labeling changes on European versions of consumer AI products, and expect at least some providers to argue the markings are technically infeasible at scale.