# EU AI Act transparency obligations When you must tell people they are dealing with AI, mark generated content, or label deep fakes under Article 50, and from which date. Source: https://sted.digital/docs/standards/eu-ai-act/transparency-obligations · Last reviewed 17 October 2026 · Strand Standards and regulations EU AI Act Article 50 sets four transparency duties from 2 August 2026: providers tell people they are interacting with AI (50(1)) and mark synthetic content as machine-readable (50(2)); deployers inform people exposed to emotion recognition or biometric categorisation (50(3)) and disclose deep fakes and some AI-written public-interest text (50(4)). ## The requirement in plain English The duties fall on whoever controls the relevant step: the provider builds the disclosure or the marking in, and the deployer tells people about how it uses the system. **Interacting with people (EU AI Act Article 50(1)).** A provider designs a system that interacts directly with people so they are told they are dealing with AI, unless that is clear to a reasonably well-informed person from the circumstances. Systems authorised by law to detect or investigate crime are exempt unless the public can use them to report a crime. **Marking synthetic content (EU AI Act Article 50(2)).** A provider of a system that generates synthetic audio, images, video or text marks the output in a machine-readable format, so it can be detected as artificially generated or manipulated. The technical solution must be effective, interoperable and reliable as far as is technically feasible. Systems that only assist standard editing, or do not substantially change the input or its meaning, are exempt. **Emotion recognition and biometric categorisation (EU AI Act Article 50(3)).** A deployer of either informs the people exposed to it and processes their personal data under data protection law. The workplace and education bans on the [prohibited practices page](/docs/standards/eu-ai-act/prohibited-practices) apply first. **Deep fakes and public-interest text (EU AI Act Article 50(4)).** A deployer that generates or manipulates images, audio or video which resemble real people, objects, places or events, and would falsely appear authentic, discloses that the content is artificial. For evidently artistic, creative, satirical or fictional work, the disclosure only has to avoid spoiling the work. A deployer that publishes AI-generated or manipulated text to inform the public on matters of public interest discloses it too, unless a person has reviewed it and someone holds editorial responsibility for it. **How and when to tell people (EU AI Act Article 50(5)).** The information is clear and distinguishable, given at the latest at the first interaction or exposure, and accessible. **Dates.** These duties apply from 2 August 2026. A generative system placed on the market before that date has until 2 December 2026 to meet the marking duty (EU AI Act Article 111(4)). **Codes of practice (EU AI Act Article 50(7)).** The Commission published a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026. Following it is one way to show how you meet these duties; it is not the law itself. This is guidance to help you understand the requirement, not legal advice. ## What good looks like A 35-person marketing software company runs three AI features, and its transparency policy says which paragraph applies to each. Its customer support assistant opens every conversation with a one-line statement that the customer is chatting with an AI system, and offers a route to a person. The statement is in the interface, not in the terms. Its image generator for customers' social posts is a generative system the company provides, so every image carries machine-readable marks: embedded metadata and a watermark that survives resizing. The engineering lead tests detection each release against compressed and cropped copies, and the results go in the release record. Its own marketing team uses the image tool too. For any image that shows a realistic person or event, the team adds a visible "AI-generated" label; for its blog posts, a named editor reviews every AI-drafted article and holds editorial responsibility, so the text duty does not apply. The policy owner reviews the notices every six months and whenever a feature changes what it generates. Each notice version is kept with the date it went live. ## Evidence an auditor expects - **A map of systems to Article 50 paragraphs**: the market surveillance authority asks which duty applies to which system and who carries it. - **The disclosure as users see it**: screenshots or interface copy, dated, showing it appears at first interaction. - **The marking method and detection tests** (provider): what is embedded, how it is detected, and test results on altered copies. - **Deep-fake labels** (deployer): examples of labelled content and the rule that decides when a label is needed. - **Editorial review records** (deployer): who reviewed AI-drafted public-interest text and who holds responsibility for it. - **Notice versions**: which text was live when, so a complaint can be answered for the right date. ## Common mistakes ### Disclosure buried in the terms A clause in the terms of service is not information given "at the latest at the time of the first interaction or exposure" (EU AI Act Article 50(5), in the [consolidated text on EUR-Lex](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02024R1689-20260727)). Put it in the interface. ### Marks your own pipeline strips Image compression, format conversion and social platforms often remove metadata. Test marking on the files your customers actually publish, and combine methods where one is fragile. ### Each side assuming the other labels Marking is the provider's duty; labelling deep fakes is the deployer's. A provider cannot rely on customers to label, and a deployer cannot rely on the provider's marks to meet its own disclosure duty. ### Treating human review as a general exemption The editorial-review exemption covers public-interest text only. A reviewed deep-fake video still needs its disclosure. ### Forgetting audio Voice cloning and synthetic voice-overs are synthetic content. A text-to-speech feature needs marking, and a cloned voice of a real person can be a deep fake. ## How Sted helps Sted's knowledge base includes an AI Transparency and Notices Policy (POL-AIT) with an AI Transparency Notice Record (REC-ATN), tagged to the EU AI Act and to ISO 42001 controls A.3.3, A.8.3 and A.8.5. Records are part of the Full Management System plan. Sted treats the EU AI Act as applying when at least one of your AI systems is in EU scope and carries obligations: it is prohibited, high-risk through the product route (it embeds a medical device under EU MDR classed above Class I) or through an Annex III use you recorded, or it is a generative system under the Article 50 transparency duties; a general-purpose model you trained and that is in EU scope counts only when it carries systemic risk; in-scope systems that are all minimal-risk give applies in part; and a stored high-risk assessment overrides the derived tier. Sted holds your transparency policy and notice records. It does not detect or test the content your AI systems produce. ## Related guides and terms - [EU AI Act](/docs/standards/eu-ai-act): the hub, with the risk tiers and the roles. - [Risk classification](/docs/standards/eu-ai-act/risk-classification): a system can carry these duties and the high-risk ones at once. - [GDPR transparency and privacy notices](/docs/standards/gdpr/transparency-and-privacy-notices): the privacy notice that sits beside your AI disclosure. - [ISO 42001 Annex A controls](/docs/standards/iso-42001/annex-a-controls): the controls on information for users and interested parties. - [General-purpose AI models](/docs/standards/eu-ai-act/general-purpose-ai-models): the duties of the provider whose model your generative system is built on. - [Interested parties](/docs/glossary/interested-parties): the people your notices are written for. ## Frequently asked questions Q: Does a chatbot have to say it is AI? A: Yes, from 2 August 2026, unless it is already clear to a reasonably well-informed person from the context. The provider designs the system so people are told, at the latest at their first interaction. A line in the terms of service does not meet that timing. Q: Do we have to label marketing copy written with AI? A: Not as such. The labelling duty for text covers AI-generated or manipulated text published to inform the public on matters of public interest, and it falls away when a person has reviewed it and someone holds editorial responsibility. Images, audio or video that look like real people or events are different: those deep fakes must be disclosed.