EU AI Act Article 50 Transparency Obligations: Full Compliance Checklist
Article 50 of Regulation (EU) 2024/1689 is the provision that applies to the widest range of ordinary businesses: not just high-risk AI, but any chatbot, content generator, or emotion/biometric tool. Here is what it actually requires, paragraph by paragraph.
- In force since 2 August 2026, not delayed by the 2026 Digital Omnibus (apart from one marking extension, see below).
- Applies regardless of whether the AI system is classified as high-risk.
- Four paragraphs: 50(1) AI interaction disclosure, 50(2) machine-readable marking of AI-generated content, 50(3) emotion recognition and biometric categorisation notices, 50(4) deepfake and public-interest text disclosure.
- Marking under 50(2), for generative systems already on the market before 2 August 2026, has an extended deadline of 2 December 2026.
- Enforced by national market surveillance authorities; fines fall under the general AI Act penalty regime (see the fines guide).
Who Article 50 actually applies to
Article 50 is not limited to high-risk AI systems. It applies to providers and deployers of four kinds of system: systems that interact directly with people (chatbots, voice assistants), systems that generate synthetic audio, image, video or text, emotion recognition and biometric categorisation systems, and systems used to produce deepfakes or public-interest text. If your company uses a customer-facing chatbot, an AI writing tool, an AI image generator, or emotion/biometric recognition, Article 50 already applies to you, whatever your company size.
This is the obligation most SMEs actually need to act on right now. High-risk obligations (Annex III/Annex I) were pushed back by the 2026 Digital Omnibus; Article 50 was not.
Article 50(1): AI interaction disclosure
Providers must design AI systems that interact directly with people so that those people are informed they are interacting with an AI system, unless this is obvious to a reasonably well-informed, observant and circumspect person from the circumstances and context. The exception for systems authorised by law to detect, prevent, investigate or prosecute criminal offences is narrow. In practice: a support chatbot, a voice assistant, an AI sales agent all need a clear, unambiguous disclosure at or before the point of interaction. If you deploy a third-party chatbot, check that the disclosure is switched on and visible in your setup.
Article 50(2): Marking AI-generated content
Providers of AI systems that generate synthetic audio, image, video, or text content must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. This is the "watermarking" obligation, and it’s a provider duty (the company that builds or offers the AI system), not a deployer duty. For generative systems already on the market before 2 August 2026, the deadline was extended to 2 December 2026 by the Digital Omnibus. See the Article 50(2) guide for the detail.
Article 50(3): Emotion recognition and biometric categorisation
Deployers of an emotion recognition system or a biometric categorisation system must inform the people exposed to it that the system is operating, and process their personal data in line with the GDPR. Remember that emotion recognition in workplaces and educational institutions is prohibited outright under Article 5, apart from medical or safety reasons, so a transparency notice is only relevant where the system is lawful in the first place.
Article 50(4): Deepfakes and AI-generated public-interest text
Deployers of an AI system that generates or manipulates image, audio, or video content constituting a deepfake must disclose that the content has been artificially generated or manipulated. For evidently artistic, creative, satirical or fictional work, the disclosure can be limited to not hamper the display or enjoyment of the work. This deployer duty sits alongside the provider’s marking obligation under 50(2): the two aren’t interchangeable, and both can apply to the same piece of content.
If you deploy an AI system to generate or manipulate text published to inform the public on matters of public interest, you must disclose that the text was artificially generated or manipulated, unless the content has gone through human review or editorial control and a natural or legal person holds editorial responsibility for it. This exemption is why most professionally edited AI-assisted journalism doesn’t trigger a disclosure requirement, but an unedited auto-published AI news feed likely does.
Practical checklist
Before you can claim Article 50 compliance, you need to be able to answer yes to each of these for every relevant AI system:
- Every AI system in the company is logged somewhere, not just the ones IT approved.
- Each logged system has an Article 50 relevance tag: 50(1), 50(2), 50(3), 50(4), or none.
- Every chatbot or voice AI discloses it’s AI, before or during the interaction, not buried in a footer.
- Every generative feature you provide marks its output in a machine-readable way where 50(2) applies.
- Everyone exposed to an emotion recognition or biometric categorisation system is informed that it is operating.
- Every deepfake your company publishes is disclosed as AI-generated or manipulated.
- Any AI-generated public-interest text is either disclosed as AI-generated or has a named human editor with editorial responsibility.
- You have a timestamped record of when each disclosure was implemented, in case a regulator or a client asks for evidence.
Questions
Does Article 50 apply to internal-only AI tools?
The disclosure duties are triggered by interaction with natural persons, by exposure to emotion or biometric systems, or by content that reaches an audience. A purely internal tool with no external-facing output or interaction generally doesn’t trigger 50(1)-(4), but staff are natural persons too, and the tool should still be logged in your AI asset inventory, since its classification can change if its use expands.
Is a small "AI-generated" badge on a chatbot enough?
The Act requires disclosure to be clear to a reasonably well-informed person at the relevant point of interaction. A generic badge buried where users won’t see it before or during the conversation is unlikely to meet that bar. This is an operational judgment call, not a legal one, and if you need certainty, confirm the specific wording and placement with qualified counsel.
Not sure which risk tier your AI system is in?
Take the free 5-question Risk Tier Quiz. No email needed, result in under a minute.
Want this tracked automatically, not read about?
The EU AI Act Governance Cockpit is a Notion + n8n workspace that logs your AI systems, runs this exact risk classification, tracks the required actions, and keeps the evidence. One-time payment from €49.
Related guides
Founder of AI Compliance Lab. Specialist in EU AI Act governance workflows and automation. LinkedIn →