AI Act transparency obligations under Art. 50 — the EC guidelines in practice

Art. 50 addresses four groups of cases with different responsible parties: providers must inform users of interactive AI systems (Art. 50(1)) and mark the outputs of generative systems (Art. 50(2)); deployers must inform about emotion recognition and biometric categorisation (Art. 50(3)) and disclose deepfakes and AI-generated texts on matters of public interest (Art. 50(4)). The obligations can apply cumulatively to a single system, and a company can be provider and deployer at the same time. Actors outside the EU are also covered where the output of their system is used in the Union.

Providers must design AI systems intended to interact directly with natural persons in such a way that the persons concerned are aware that they are interacting with an AI — under Art. 50(5), in a clear and distinguishable manner and at the latest at the time of the first interaction. According to the draft guidelines, a notice that appears only in a manual or is hidden in menu levels is expressly not sufficient. This affects, for example, chatbots, voice assistants and AI agents in customer-facing roles.

Providers of generative AI systems must mark audio, image, video and text outputs in a machine-readable format and make them detectable as artificially generated or manipulated — both elements, marking and detection, must be fulfilled. The technical solutions must be effective, interoperable, robust and reliable; recital 133 lists techniques such as watermarks, metadata identifications, cryptographic methods for proving provenance, logging methods and fingerprints. Multimodal content, virtual/augmented reality and outputs of agentic AI systems are also covered, provided they are perceptible by humans.

Deployers must disclose when image, audio or video content constitutes a deepfake — meaning, under Art. 3(60) AI Act, content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic or truthful. The same applies to AI-generated or AI-manipulated texts that inform the public on matters of public interest. The disclosure must be perceptible to humans and complements the machine-readable marking applied by the provider.

The draft confirms only a few exceptions: obvious AI interaction (Art. 50(1)), assistive standard editing that does not substantially alter the input data or its semantics (Art. 50(2)), legally authorised law enforcement purposes, and AI-generated texts that have undergone human review or editorial control and for which editorial responsibility is held (Art. 50(4)). A softened regime applies to evidently artistic, creative, satirical or fictional deepfake works: the disclosure must not hamper the display or enjoyment of the work. Purely personal, non-professional use by natural persons is exempt from the deployer obligations under Art. 2(10) AI Act.

Under Art. 113 AI Act, Art. 50 applies from August 2, 2026 — including to systems placed on the market before that date; content generated before the cut-off date does not have to be marked retroactively, and the AI omnibus proposal provides for a targeted transitional rule only for the marking obligation under Art. 50(2) for existing systems. For implementation, a complete AI inventory, a role mapping (provider/deployer) and a gap analysis against Art. 50(1)–(4) have proven effective. According to the draft, joining a Code of Practice assessed as adequate by the AI Office is a straightforward way to demonstrate compliance with the obligations under Art. 50(2) and (4).