EU publishes AI Act transparency guidance

The EU just moved one step closer to telling AI companies, in plain operational terms, what “be transparent” actually means. That sounds bureaucratic, but the stakes are real — if your product chats with people, generates content, or edits media in ways that stop being “just editing,” you may need disclosures, labels, and detection-friendly output by August 2, 2026. The gap until now was that Article 50 of the AI Act set the obligation, but teams still had to guess where the lines really were. On May 8, the European Commission published draft guidelines and opened a consultation that runs through June 3. ### What did Brussels actually publish? It published draft guidelines for Article 50 of the AI Act — the transparency section — plus a public consultation for feedback before adoption. The guidelines are meant to help both regulators and companies apply the rules consistently, and they sit alongside a separate Code of Practice on marking and labelling AI-generated content. That code is expected in June 2026 and is voluntary, but it is supposed to help companies show they are complying. (digital-strategy.ec.europa.eu) ### Who is covered by these rules? Four buckets matter. First, AI systems that interact directly with people — chatbots, assistants, voice agents. Second, systems that generate or manipulate content and need that output marked in a machine-readable way. Third, deployers of emotion-recognition or biometric-categorisation systems. Fourth, deployers of deepfake systems and certain AI-generated or AI-manipulated text published on matters of public interest. (digital-strategy.ec.europa.eu) ### What does “transparency” mean here? Basically, three things. Tell people when they are dealing with AI rather than a human. Mark synthetic or manipulated output so machines can detect it later. And make those notices clear and accessible, not buried in legal sludge. The Commission is also pointing companies toward technical methods like watermarks, metadata, cryptographic provenance, logging, and fingerprinting — not one magic standard, but a toolbox. (digital-strategy.ec.europa.eu) ### Why does machine-readable marking matter so much? Because the law is not just asking for a visible “made with AI” sticker. It wants content to be detectable in ways platforms, tools, and investigators can work with at scale. That is a much harder engineering problem. A chatbot disclaimer is copywriting. Reliable marking for audio, images, video, and text is infrastructure — especially if the output gets remixed, compressed, screenshotted, or lightly edited downstream. The draft says the solution should be effective, interoperable, robust, and reliable as far as technically feasible. (digital-strategy.ec.europa.eu) ### Are there carve-outs? Yes — and they matter. If it is already obvious to a reasonably informed person that they are interacting with AI, the interaction notice may not be required. There are also exceptions tied to law-enforcement uses in the legal text. And for synthetic-content marking, the Act carves out assistive functions for standard editing or cases where the system does not substantially alter the input or its meaning. That last bit is the line many product teams will obsess over. (digital-strategy.ec.europa.eu) ### So what products should worry first? Customer-service bots, meeting assistants, writing tools, image and video generators, voice cloning features, and anything that can create realistic public-facing content. The tricky part is that the rule often hits both providers and deployers. So a model company, an app builder, and the enterprise using the app may each carry different pieces of the compliance load. (digital-strategy.ec.europa.eu) ### Why is this landing now? Because Article 50 takes effect on August 2, 2026. The Commission is trying to narrow ambiguity before that date instead of leaving companies to reverse-engineer enforcement from first cases. Consultation feedback closes on June 3, 2026, so this is still draft guidance, not the final word. But the direction is clear — the EU wants AI transparency to be product behavior, not a policy PDF. (digital-strategy.ec.europa.eu) ### Bottom line? If you ship AI in Europe, “users should probably know” is turning into a checklist with dates, scopes, and technical expectations. The hard part is not adding one label. It is proving, across the whole product, when AI is present, when output is synthetic, and when your exception really is an exception. (digital-strategy.ec.europa.eu)