Key takeaways
- 01.Labeling obligationAI-generated content must be marked in a machine-readable format and detectable as synthetic by providers.
- 02.Disclosure obligationDeployers of AI systems that generate or manipulate content resembling real persons or events must disclose the manipulation to users.
- 03.Training data summaryGPAI providers must publish a sufficiently detailed summary of copyrighted content used for training.
Background
I · ContextThe evolution of the AI Act reflects a shift from a product-safety approach to a fundamental rights-based framework. While initial drafts focused on high-risk industrial applications, the rapid commercial deployment of general-purpose AI necessitated a specific regulatory tier targeting generative outputs.
Article 50 is the Act's transparency backbone for content. It sits alongside the broader high-risk system regime but applies horizontally: any provider or deployer generating synthetic content in the EU market is potentially covered, regardless of whether the underlying system is classified as high-risk.
Detailed provisions
II · The obligationsLabeling & watermarking
Article 50(2) requires providers to ensure that outputs of AI systems generating synthetic audio, image, video, or text content are marked in a machine-readable format and detectable as artificially generated or manipulated. The marking must be technically feasible, effective, interoperable, robust, and reliable as far as technically feasible.
Deepfake disclosure
Article 50(4) requires deployers of AI systems that generate deepfakes — image, audio, or video content that appreciably resembles existing persons, objects, places, or events — to disclose that the content is artificially generated or manipulated. Exemptions apply to artistic, creative, satirical, or fictional works, provided the disclosure is not disproportionate to the creative purpose.
Transparency on AI interaction
Article 50(1) requires deployers of AI systems that interact with natural persons to ensure the affected persons are informed they are interacting with an AI system, unless obvious from context or required by law.
GPAI training data summary
Article 53(1)(d) requires providers of general-purpose AI models to publish a sufficiently detailed summary about the content used for training, using a template to be provided by the AI Office.
Specific exemptions apply for artistic, creative, satirical, or fictional works, provided they do not infringe the rights of others. The Act delegates the exact technical specifications for watermarking and labeling to a forthcoming implementing act, expected in late 2026.