The European Commission adopted the final 51-page guidelines to Article 50 of the AI Act on July 20, 2026, handing chatbot operators, synthetic-media providers, and deepfake deployers exactly 13 days to interpret them before enforcement begins on August 2. Market surveillance authorities across the member states gain their powers that day, with fines reaching €15 million or 3% of worldwide annual turnover, and a floor near €750,000.

The compression is the story. Brussels spent much of the past year telegraphing flexibility, and the Digital Omnibus duly pushed the high-risk tier out by 16 months, to December 2027 and August 2028. Transparency wasn’t touched. Enterprises that read the delay as a general reprieve have discovered, with days to spare, that it wasn’t one.

The obligations themselves are structural. Providers must tell users they’re talking to a machine and embed machine-readable marks in AI-generated audio, image, video, and text. Deployers must disclose deepfakes, emotion recognition, biometric categorization, and unreviewed AI-generated content on matters of public interest. Bird & Bird notes the final text drops retroactive labelling under Article 50(4): for synthetic media, the cut-off is the date of generation, a meaningful shift from the May draft. Text on public-interest matters is governed by date of publication, which means pre-August drafts published after the deadline still need labels unless an editorial-control exception applies.

Two details reshape the compliance perimeter. Data Protection Report flags that in-house branded chatbots are captured: apply your trademark to a third-party model and you’re the provider. And the Commission’s Code of Practice on Transparency of AI-Generated Content has been assessed as adequate, so signatories get a presumption of compliance. A four-month grace period for pre-existing systems runs to December 2, 2026.

The familiar pattern: publish the manual after the exam has started, then judge who studied.

Sources

Sources