Europe's regulator moved from paper to practice, and the first duty is disclosure
Chatbots must identify themselves as automated. Generated or altered content must carry machine-readable marks. Penalties reach €15m or 3% of worldwide turnover, and they apply regardless of when a system shipped.
Enforcement of the EU AI Act began on 2 August 2026, with the Commission's AI Office and national authorities now applying transparency obligations: chatbots must identify themselves as automated systems, deepfakes require labels, and machine-made or edited content must carry machine-readable marks. Non-compliance risks fines up to €15m or 3% of worldwide annual turnover, whichever is higher.
The clause doing the most work is the one about scope: the obligations apply immediately to all in-scope systems regardless of when they were placed on the market. There is no grandfathering. A product shipped in 2024 and running unchanged is in scope today.
That is a deliberate and unusual choice. Most technology regulation applies prospectively because retrofitting is expensive; applying it to the installed base means the compliance burden lands hardest on companies with the most deployed product, which is to say the largest ones.
Machine-readable marking is the technically hardest of the three duties. Self-identification is a string in a system prompt. Labelling a deepfake is a UI decision. Embedding a durable, machine-readable provenance mark that survives re-encoding, cropping and screenshotting is an unsolved research problem being required as a compliance obligation.
The practical question for the next quarter is what "machine-readable" is held to mean in the first enforcement actions. A metadata field satisfies the letter and is stripped by any upload pipeline; a robust watermark satisfies the intent and does not reliably exist. The answer will be found in cases, not in the text.
European Commission — Commission starts enforcing AI Act rules and new transparency requirements on 2 August → · Help Net Security — EU begins enforcing AI Act, putting AI models under the microscope →