Four categories, one core idea
The core idea behind Art. 50 is simple: people should be able to tell when they're interacting with AI or looking at AI-generated content. This duty is entirely independent of the risk classification – it applies on top, the moment one of the following four situations occurs, regardless of whether the system otherwise counts as minimal or limited risk.
The four categories in detail
- Chatbots and conversational systems: people must be able to tell they're interacting with an AI rather than a human – unless that's obvious anyway.
- Emotion recognition and biometric categorisation: affected people must be informed that such a system is being used.
- Deepfakes (synthetic images, audio, video): content that deceptively replicates or alters real people, objects, or events must be labelled as AI-generated or AI-manipulated.
- AI-generated text on matters of public interest: if a text is published to inform the public, it must be disclosed that it's AI-generated.
Deadlines: not everything starts at the same time
Most Art. 50 obligations apply as originally scheduled on 2 August 2026. One exception: the labelling duty specifically for synthetic media (deepfakes) was pushed back four months to 2 December 2026 under the AI Omnibus – a small but practically relevant difference from the other three categories. Also important is the split by role: providers of generative systems must mark their outputs in a machine-readable format as AI-generated; deployers who publish a deepfake must additionally label it visibly for humans.
Practice section: concrete labelling for your chatbot or content
For a website chatbot, a clear, visible notice at the start of the conversation (“You're chatting with an AI assistant”) is usually enough – no elaborate certification procedure required. For AI-generated marketing content depicting real people or events, a visible, unmissable label is required, not a note buried in the imprint's small print. Purely internal use – a chat assistant only staff can see – typically doesn't trigger these public-facing obligations, since nobody outside the company interacts with the system or consumes the content. This classification doesn't replace case-by-case legal advice.