As generative AI improves, the line between real and synthetic content blurs. Images, voices and videos can now be generated so that they are barely distinguishable from authentic material. The EU AI Act responds with a dedicated transparency duty in Article 50: certain AI-generated or manipulated content must be labelled as such. For any company that creates content with generative AI, from marketing to media to product development, this is a concrete operational duty.
Who the labelling duty applies to
Article 50 distinguishes two addressees. Providers of AI systems that generate synthetic content must ensure outputs are marked in a machine-readable format as artificially generated, for example through watermarks or metadata. Deployers have complementary duties: anyone using an AI system to create a deepfake, meaning image, audio or video content that appreciably resembles real people, objects or events, must disclose that the content was artificially generated or manipulated. Similar rules apply to AI-generated text published to inform the public on matters of public interest, unless it has undergone human editorial review and someone holds editorial responsibility.
Understanding the exceptions
The duty is not absolute. For evidently artistic, creative, satirical or fictional works, a lighter form applies: disclosure must not hamper the display or enjoyment of the work, for example an unobtrusive note rather than an intrusive overlay. Determining the applicable variant correctly is the craft: a satirical video is subject to a different requirement than a deceptively real advertising clip with the synthetic voice of a real person.
Machine-readable and human-readable
A point often misunderstood: Article 50 essentially requires two levels. Technically, outputs should be marked so platforms and downstream systems can detect synthetic content automatically; content provenance standards such as C2PA and robust watermarking are the technical answer. For people, it must be recognisable in the case of deepfakes that the content is artificial. Both levels are needed: a watermark alone does not fulfil the disclosure duty towards humans, and a visible note alone does not meet the machine-readable requirement.
Technical implementation and timeline
Anyone using generative AI for publicly visible content needs a process that applies labelling reliably. Provenance metadata and watermarks that survive editing serve the machine-readable side; a caption, overlay or note often suffices for the human-readable side. Robustness matters: a label that disappears at the first crop or re-export does not serve its purpose, so anchor labelling as early as possible in the generation process. The transparency duties have applied since 2 August 2026; for existing generative systems, the transitional period for machine-readable marking ends on 2 December 2026.
Beyond the duty: trust and adjacent rules
Labelling is also a credibility signal in a media landscape where trust is scarce. At the same time, Article 50 does not stand alone: if a generative system produces personal data such as the face or voice of a real person, the GDPR applies in parallel, and deepfakes of real people touch personality rights and the right to one’s own image. Treat labelling as one building block of a broader responsibility framework.
What to do
Inventory your generative AI use: where do you create or edit image, audio, video or relevant text content with AI, and which of it is published? For each use case, determine the applicable labelling variant, define the responsible process and implement both machine-readable and human-readable marking. Document the decision. The related chatbot duties are covered in our article on Article 50 for chatbots and in the glossary entry on Article 50 transparency duties.