Labelling AI-generated content: the machine-readable marking teams get wrong
Article 50 gets summarised as "label your AI content," and that summary is exactly why teams get it wrong. The transparency obligations that apply in full on 2 December 2026 include a machine-readable marking requirement that a visible watermark alone doesn't satisfy. Here's the distinction that matters.
Two different audiences: people and machines
Article 50 covers several transparency situations, and the one teams underestimate is synthetic content. The obligation isn't only to tell a human "this was AI-generated" — it's to mark AI-generated or manipulated audio, image, video and text in a format that machines can detect, so the provenance travels with the content. A visible "made with AI" caption is good practice, but it's not the same as the machine-readable marking the Act asks providers to build in.
Deepfakes and manipulated media get explicit treatment
Content that appreciably resembles real people, objects or events — deepfakes and manipulated media — carries its own disclosure duty: it must be labelled as artificially generated or manipulated. There are narrow carve-outs (for example, clearly artistic or satirical work, handled proportionately), but the default is disclosure. If your product can generate realistic media of people, this is squarely your obligation.
Chatbots: tell people they're talking to AI
The simpler, near-universal duty is interaction transparency: if a person is interacting with an AI system (a chatbot, a voice agent), they must be informed, unless it's obvious from the context. This one is cheap to implement and easy to forget — a single clear line at the start of the conversation usually does it. It's also the obligation most mid-market companies actually have, because so many run a support or sales chatbot.
Who owes what
Roughly: providers who build generative systems carry the marking-at-source duty; deployers who use them carry disclosure duties in how they present the output. If you both build and deploy, you carry both. The practical move is to map your content-generating and conversational systems now and assign each obligation, rather than discovering on 1 December that the machine-readable part needed engineering work you didn't budget for.
Educational orientation, not legal advice. EU AI Act references current as of September 2026. For binding interpretation of your obligations, consult qualified counsel.