The EU AI Act's Disclosure Rules Start August 2026
Correct as of 2 August 2026. Article 50 of the EU AI Act — the transparency-obligations article covering AI-generated content and chatbot interactions — becomes applicable today. Guidance and an in-development Code of Practice on marking and labelling accompany it, which tells you something on its own: the rule reads simply, and the practical detail of complying with it is still being worked out even as the deadline lands. Content was rephrased for compliance with licensing restrictions.
What the obligation actually covers
Article 50 has two distinct halves, and conflating them is the most common misreading. The first: a person interacting with an AI system must be told they are talking to one, unless it is obvious from context. The second: AI-generated or AI-manipulated content — image, audio, video, or text published to inform the public on a matter of public interest — must be marked as such. These are different obligations with different triggers, aimed at different harms, and a compliance plan that treats them as one item on a checklist will miss cases.
Provenance tagging — attaching a durable, checkable marker to generated content at the point it's produced — is the technical answer to the second half. Article 50 is the regulatory reason that technique now has a deadline rather than being a nice-to-have, and the data governance discipline of knowing which downstream artifact came from which source is the same lineage problem this obligation raises, one level up.
Where the boundary cases actually are
A support chatbot clearly identified as a bot at the top of the conversation is the easy case. Harder: a chatbot embedded in a product flow with no explicit self-identification, where "obvious from context" is doing real legal work and reasonable people disagree about what counts as obvious. Harder still: AI-assisted content where a human materially edited machine output — the disclosure obligation is about the output, not about crediting an authorship split, and "materially edited" has no bright line in the text.
The public-interest qualifier on the content-marking half is its own trap. A synthetic product photo is not obviously "informing the public on a matter of public interest." A synthetic image illustrating a political or health claim plausibly is. Teams building content pipelines that touch anything adjacent to news, politics, health, or public safety should assume the stricter reading applies rather than the narrower one, because the cost of over-disclosure is low and the cost of under-disclosure, once flagged, is not.
What to build now rather than later
Two engineering artifacts fall directly out of this. A chatbot identification pattern — a persistent, unambiguous signal in the interface that the user is talking to a system, not a person, checked at the point of first interaction rather than buried in terms of service. And a content-provenance tag attached at generation time to any output that could plausibly be published — not retrofitted later, because retrofitting requires knowing which historical outputs need it, and that knowledge degrades fast once a pipeline has been running unlabeled.
Neither of these is a compliance checkbox you bolt on afterward. Both are cheapest when they're part of the generation pipeline from the start, which is the argument for building them now rather than waiting for an enforcement action to clarify the boundary cases for you.