The European Commission published its final Article 50 transparency guidelines on July 20, a 51-page document arriving 13 days before the obligations it explains become enforceable. From August 2, providers and deployers of AI systems inside the EU face fines up to €15 million or 3% of worldwide annual turnover, whichever is higher, for failures the rulebook only just finished defining.

The guidelines are non-binding, but they’ll frame supervisory expectations from day one. Providers must ensure chatbots disclose their artificial nature and mark synthetic audio, image, video, and text in machine-readable form. Deployers must label deep fakes, disclose AI-generated content on matters of public interest that lacks human editorial control, and flag emotion-recognition and biometric-categorisation systems. AI agents get a decisively agentic reading: they must disclose both that they’re artificial and the person on whose behalf they act.

The final text softened one edge. Retroactive labelling is out: for deep fakes the trigger is the date of generation, not publication, so anything made before August 2 escapes marking. Text on public-interest matters is treated differently, with publication as the trigger, meaning pre-deadline copy published after can still require a label unless editorial control applies. Generative systems already on the market get until December 2, 2026 to implement machine-readable marking under Article 50(2).

June’s Digital Omnibus pushed the AI Act’s high-risk Annex III obligations, covering employment, creditworthiness, education, and essential services, to December 2, 2027, and regulated-product high-risk systems to August 2, 2028. Transparency wasn’t delayed. Brussels held the line where the compliance burden was widest and the political optics cleanest.

Enforcement now passes to national market surveillance authorities. In Ireland, the AI Office of Ireland coordinates sectoral regulators under the Regulation of Artificial Intelligence Bill 2026. Signatories to the Code of Practice on Transparency of AI-Generated Content, published June 10 and assessed by the Commission as adequate, get supervision focused on code implementation, with adherence potentially mitigating fines. The rulebook arrived late; the meter starts anyway.

Sources

Sources