European Commission issues AI transparency guidance, spells out labeling requirements

The EU just moved AI transparency from principle to product spec. On 8 May, the European Commission published draft guidelines for Article 50 of the AI Act and opened a consultation that runs until 3 June. The point is simple — if people are talking to AI, or looking at certain AI-made or AI-manipulated content, they need to be told clearly. And for companies building these systems, the deadline is no longer abstract: these duties apply from 2 August 2026. ### What is Article 50 actually about? Article 50 is the transparency part of the EU AI Act. It covers four buckets: AI systems that interact with people, systems that generate or manipulate content, emotion-recognition or biometric-categorisation systems, and deepfakes or certain AI-generated public-interest text. The basic idea is that users should know when AI is involved, and platforms should make synthetic content detectable. (digital-strategy.ec.europa.eu) ### What changed this week? The law itself was already on the books, but the Commission has now published draft implementation guidelines. That matters because companies were staring at broad legal duties and trying to guess what “inform people” or “mark content” would mean in practice. The draft is meant to make compliance more consistent across the EU, and the Commission says it was prepared alongside a separate Code of Practice on marking and labelling AI-generated content. (digital-strategy.ec.europa.eu) ### So what do chatbot makers have to do? If a system interacts directly with a person, the provider has to make sure the person is informed that they are dealing with AI rather than a human — unless that is already obvious from the context. That sounds narrow, but it hits a lot of real products: customer-service bots, AI assistants, booking flows, and voice agents. The practical question is not whether a model is “advanced.” It is whether the user could reasonably think a human is on the other end. (digital-strategy.ec.europa.eu) ### What about AI-generated images, audio, and text? Here the EU is pushing for two layers. One is user-facing disclosure in some cases — especially for deepfakes and AI-generated or AI-manipulated public-interest text. The other is machine-readable marking, so downstream systems can detect that content was artificially generated or altered. The Commission’s materials point to tools like watermarks, metadata, cryptographic provenance methods, fingerprints, and logging. (digital-strategy.ec.europa.eu) Basically, not just a visible sticker, but a technical trail. ### Why is “machine-readable” such a big deal? Because a visible label helps a person, but machine-readable marking helps an ecosystem. Platforms, archives, moderation tools, and verification systems can use those signals automatically. That is the difference between “we told one user” and “we made this content traceable at scale.” The catch is that this becomes an engineering problem fast — how marks survive editing, recompression, screenshots, or reposting. (digital-strategy.ec.europa.eu) ### Are these rules final? Not quite. The guidelines published on 8 May are draft guidance, and the consultation stays open until 3 June 2026. The Commission also says the final Code of Practice is expected in June 2026 and will be voluntary, but useful as a way to demonstrate compliance. So the legal obligations are coming either way on 2 August — the remaining question is how much extra implementation detail companies get before then. (digital-strategy.ec.europa.eu) ### Who else is covered? It is not just chatbot vendors. Deployers of emotion-recognition and biometric-categorisation systems have notice duties too. And deployers of systems that generate or manipulate deepfakes — plus certain AI-generated text meant to inform the public on matters of public interest — also have to disclose the artificial origin of that content, subject to defined exceptions. (digital-strategy.ec.europa.eu) ### What is the real takeaway? The EU is telling AI companies to stop treating transparency as a policy slogan. From August, it becomes interface copy, metadata, provenance tooling, and compliance records. The draft guidance does not change the direction of travel — but it makes the work much harder to postpone. (digital-strategy.ec.europa.eu 1) (digital-strategy.ec.europa.eu 2)