Brussels Spells Out Its AI Labelling Rules Before They Bite

The European Commission has laid out, in detail, what its new AI transparency rules will and will not require. On July 20 it published guidelines on Article 50 of the AI Act, the section that governs when people have to be told they are dealing with AI. The timing is deliberate. Those obligations start to apply on August 2, and companies have spent months asking exactly where the lines fall.

What Article 50 actually demands

Two duties sit at the heart of it. Systems that interact directly with people, such as chatbots and voice agents, have to make clear they are AI rather than a human, unless that is already obvious. And content generated or meaningfully altered by AI, including synthetic images, audio, and video, has to be marked as such in a machine-readable way. The guidelines clarify who counts as a provider and who counts as a deployer, and which of them carries each obligation.

The document is not itself law. It is the Commission's reading of the law, meant to support what it calls the "smooth and effective implementation" of the rules. In practice, guidance like this is what regulators and companies reach for when a dispute arises, so its interpretations carry real weight even without the force of a statute.

Part of a hard fortnight for compliance

August 2 is a date we have flagged before. It is the same deadline that brings the AI Act's high-risk obligations into force, and it arrives while the Commission is also leaning on Google over Android interoperability. For any company shipping a chatbot or a generative tool into the European market, the message from Brussels this month is consistent. The grace period is ending, and the details are now on paper.

For users, the practical effect should be gradual but noticeable. More chatbots opening with a plain statement that they are automated. More AI-made images and clips carrying an embedded label your software can read even when your eye cannot. Whether that marking survives a screenshot or a re-upload is the harder question, and one the guidelines cannot answer on their own.