Starting today, any company operating in the European Union that deploys a chatbot, publishes AI-generated video, or uses AI to categorize people’s emotions faces a new legal obligation: tell people when artificial intelligence is involved. Article 50 of the EU’s landmark AI Act becomes enforceable on August 2, 2026, marking one of the broadest AI transparency mandates anywhere in the world.
What to know:
- Article 50 of the EU AI Act applies from August 2, 2026, requiring AI systems that talk to people, generate synthetic content, or create deepfakes to disclose their artificial nature.
- Chatbots and virtual assistants must be designed so users are told upfront they’re interacting with AI — a disclaimer buried in the terms and conditions doesn’t count.
- Deepfake creators must label manipulated images, audio, or video that could pass as authentic, with disclosure required at the point the content is first seen or heard.
- Generative AI providers must add machine-readable watermarks to synthetic text, images, audio, and video, though systems already on the market get until December 2, 2026 to comply with that specific marking requirement.
- The rules aren’t limited to “high-risk” AI — the European Commission’s own compliance data shows transparency obligations are the second most common compliance trigger businesses face under the Act, affecting an estimated one-third of organizations it surveyed.
- Publishers using AI-generated text to inform the public on matters of public interest must disclose it, unless a human editor has reviewed the content and taken responsibility for it.
The obligations fall into four categories, and together they touch a surprisingly wide swath of the tech industry. Under Article 50(1), any AI system designed to interact directly with people — chatbots, automated phone systems, AI agents — must make its artificial nature obvious at the very start of the interaction. The European Commission’s draft guidelines make clear that a vague reference to an “assistant,” a watermark buried in metadata, or a disclosure tucked into a website’s fine print will not satisfy the requirement. The disclosure has to be something an ordinary, reasonably attentive person would actually notice.
Article 50(2) targets the output side of generative AI: any system that produces synthetic audio, images, video, or text must mark that output in a machine-readable format so that detection tools can verify its origin. This is arguably the most technically demanding piece of the law, and regulators have built in some breathing room — under a provisional agreement reached in May 2026, generative AI systems already on the market before today have until December 2, 2026, to meet the machine-readable marking requirement specifically, even though the broader transparency duties apply now.
The other two provisions are narrower but carry real consequences for specific industries. Article 50(3) requires companies deploying emotion-recognition or biometric-categorization systems — tools that infer stress, sentiment, or demographic traits from a person’s face or voice — to inform anyone exposed to them, layering onto existing obligations under the EU’s General Data Protection Regulation. Note this is separate from the outright ban on emotion recognition in workplaces and schools, which has been in force since February 2025; outside those settings, the technology is generally permitted, but only with disclosure.
Article 50(4), meanwhile, covers deepfakes and AI-generated news content. Anyone deploying AI to create image, audio, or video content realistic enough to be mistaken for authentic footage of a real person, place, or event must disclose that it was artificially generated or manipulated — through a persistent visual label, an opening disclaimer, or an audible warning, depending on the medium. Satirical or clearly fictional content gets lighter-touch treatment. Separately, publishers who use AI-generated text specifically to inform the public on matters of public interest must flag it as such, unless the material has gone through substantive human editorial review.
Enforcement infrastructure is still catching up to the deadline. The European Commission has published draft guidelines interpreting the scope of Article 50 and is finalizing a voluntary Code of Practice on AI-generated content — expected to settle on a standardized EU label, tentatively a simple “AI” tag, localized across member-state languages — that regulators are likely to treat as the practical benchmark for compliance even though adherence to it isn’t mandatory.
For businesses, the reach of the rule is the headline. Unlike the Act’s high-risk system requirements, which apply only to a defined subset of AI uses such as hiring tools or credit scoring, Article 50 applies to any qualifying AI system regardless of risk classification. A retailer’s customer-service chatbot, a marketing team’s AI-generated product video, or a local newsroom’s AI-assisted article can all trigger obligations. The Commission’s own compliance-checker data puts transparency requirements just behind AI literacy training as the most common reason organizations find themselves in scope of the law at all.
The rollout adds to a year of incremental AI Act milestones — prohibitions on the most dangerous uses of AI took hold back in February 2025, obligations for general-purpose AI model providers arrived in mid-2025, and now the transparency layer aimed squarely at the tools ordinary consumers encounter every day is arriving. What’s still unsettled is exactly how strictly national regulators will enforce the disclosure standard in practice, and how quickly companies outside the EU — many of whom serve EU customers regardless of where they’re headquartered — adjust their products to comply.
Leave a Reply