Article 50 AI Act: labelling AI avatars and synthetic content

Article 50 concerns AI systems that do not carry high risk but where people could be deceived: conversations with a machine that seems human, and media that look real but are generated. The regulation answers this with transparency rather than bans. People who know they are talking to AI or watching a generated video can put what they hear in context.

The duties apply under the general rule of Article 113 from 2 August 2026. They sit alongside other obligations: if a system is also high-risk, the requirements of Chapter III apply in addition (Article 50(6)). Breaches of Article 50 can be fined under Article 99(4) up to EUR 15 million or 3 percent of worldwide annual turnover; for SMEs the lower amount applies.

Article 50 deliberately splits duties between providers and deployers: those who build a system must make labelling technically possible; those who use it must disclose in certain cases.

Article 3(60) defines a deep fake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Two features matter: resemblance to something real and the capacity to be taken as genuine.

In training this becomes relevant as soon as a digital twin of a real person speaks — for example an avatar built from photos and voice recordings of a lecturer, saying new sentences she never spoke. Such a video resembles a real person and could be taken as a genuine recording. Under Article 50(4) the deployer must disclose that the content is artificially generated. A clearly drawn learning character that resembles no one is not a deep fake; paragraph 1 may still apply if it interacts with learners.

Article 50(5) requires information to be clear and distinguishable, provided at the latest at the time of first interaction or exposure, and to conform to accessibility requirements. A sentence hidden in the terms of use hardly meets that. Visible notices right next to the content, also readable by screen readers, work well.

For machine-readable marking under paragraph 2 the regulation prescribes no particular technique; it requires solutions that are effective, interoperable, robust and reliable as far as technically feasible. Under paragraph 7 the Commission's AI Office is to encourage codes of practice on detection and labelling, and the Commission may adopt implementing acts. Anyone generating content should follow these developments.

Education providers are often both: deployers of third-party systems and — if they offer their own AI tools under their own name — providers. The first step is to clarify which role applies to which tool. Then duties can be assigned: technical marking with the provider, visible disclosure with the deployer.

Labelling pays off didactically regardless of the legal duty. Learners who know an avatar is generated judge its statements more realistically and are more likely to ask questions. A label that also points to sources and subject-matter review strengthens trust in the course.

In autumn 2025 the Commission proposed amendments to the AI Act that could also affect transition periods for marking under paragraph 2. The text published in the Official Journal is what counts.

This article explains the law in general terms and is not legal advice. For any individual case, the official text in its current version prevails; if in doubt, consult a lawyer, your data protection officer or the competent supervisory authority.