AI now has to be recognisable: Europe’s transparency rules become operational

Since 2 August, a central part of Europe’s AI rulebook is no longer only future law. The transparency duties in Article 50 of the AI Act apply, and the European Commission and national authorities are moving into enforcement. For users, the core idea is simple: in relevant situations they should be able to tell whether the counterpart is an AI system and whether an image, video, audio clip or text has been artificially generated or materially manipulated.

The deeper change is technical. Providers of generative systems are expected not only to display visible notices but also to mark outputs in a machine-readable way where technically feasible. Provenance is therefore meant to be detectable automatically as well. Transparency becomes infrastructure, not merely interface design.

For interactive systems the rule is relatively intuitive: people should be told when they are directly interacting with AI unless that is already obvious. This affects customer service, assistants and other conversational products. Companies therefore need to review product copy, onboarding and interface behaviour across many touchpoints rather than waiting for an enforcement request.

The obligations are especially important for synthetic or manipulated image, audio and video. Deepfakes generally have to be disclosed as artificially generated or manipulated. For artistic, satirical or fictional content, the regulation allows a more proportionate form of disclosure so that transparency does not automatically undermine the work itself. The implementation challenge is therefore less about whether to label and more about how.

Article 50 also addresses AI-generated or manipulated text published on matters of public interest. But the regulation includes a significant distinction where content has undergone human review or editorial control and a natural or legal person takes editorial responsibility. For news organisations, that makes documented accountability central. The decisive issue is not simply whether a tool was used, but what editorial process surrounded it and who is responsible for publication.

The Commission has published guidelines and a voluntary code of practice to make implementation more consistent. Technical questions remain: which marks survive compression, screenshots or later editing? How interoperable will standards be? How can labels themselves resist manipulation? Europe has set the principle. The competitive and regulatory challenge now moves to who can prove provenance and responsibility robustly.