EU AI Act Article 50: what you must disclose.
Since 2 August 2026: if you let people talk to an AI or publish AI content, you must disclose it. This guide explains the four duties in Article 50, who carries them, which notices fall short and which deadlines apply, with examples.
As of 13 October 2026 · checked monthly · not legal advice
Key takeaways
- Article 50 has applied since 2 August 2026: regardless of risk class and also for open-source systems.
- Four triggers: AI talking to people, AI-generated content, emotion recognition, deepfakes and AI texts on public matters.
- The notice must be clearly visible at the first contact at the latest, not in the terms or the footer.
- Fines up to €15m or 3% of worldwide annual turnover; in Germany the Bundesnetzagentur is responsible.
What does Article 50 of the AI Act regulate?
Article 50 is the transparency chapter of the EU AI Act. It requires that people can tell when they are talking to an AI and when content is artificially generated. Unlike the duties for high-risk systems, Article 50 was not postponed by the Digital Omnibus.
The European Commission adopted guidelines on Article 50 on 20 July 2026. They are not legally binding, but they are the reading supervisory authorities work with. Only the Court of Justice of the EU can interpret the regulation bindingly.
The four transparency duties in Article 50.
1 · AI talks to people (para. 1)
Chatbots, voice assistants and AI agents: people must be told they are interacting with an AI, unless it is obvious.
2 · AI generates content (para. 2)
Text, images, audio and video from generative AI must be marked as artificially generated in a machine-readable way.
3 · AI recognises emotions (para. 3)
Anyone using emotion recognition or biometric categorisation must inform the people affected.
4 · Deepfakes and AI texts (para. 4)
Deepfakes and AI texts that inform the public on matters of public interest must be disclosed.
Provider or deployer: who carries which duty.
| Duty | Carried by |
|---|---|
| Notice “you are talking to an AI” (para. 1) | Provider |
| Machine-readable marking of generated content (para. 2) | Provider |
| Notice for emotion recognition and biometric categorisation (para. 3) | Deployer |
| Disclosure of deepfakes (para. 4) | Deployer |
| Labelling AI texts on public matters (para. 4) | Deployer |
The provider develops an AI system and places it on the market under its own name; the deployer uses it under its own responsibility. If you run a system yourself, offer it under your own name or adapt it heavily, you may become a provider too: clarify your role with your legal counsel.
Three notices the guidelines say are not enough.
Hidden in the terms
A notice nobody sees during the conversation is not a notice.
Only a watermark
A machine-readable mark does not replace telling the person.
Only “assistant”
The word does not tell anyone it is an AI.
Deadlines for Article 50 and the EU AI Act.
| Date | What applies |
|---|---|
| 10 June 2026 | Commission publishes the voluntary Code of Practice on marking and labelling AI-generated content |
| 20 July 2026 | Commission adopts guidelines on Article 50 |
| 2 August 2026 | Article 50 applies to new systems and content |
| 2 December 2026 | End of the grace period for marking (para. 2) for systems already on the market |
| 2 December 2027 | Duties for high-risk systems under Annex III |
| 2 August 2028 | Duties for high-risk systems in products under Annex I |
Sources: AI Act (EU) 2024/1689, Digital Omnibus (EU) 2026/1744, Commission guidelines of 20 July 2026. We check the deadlines monthly.
Article 50 in practice: six examples.
Chatbot on the website
A visible notice in the chat window before the first message: “You are chatting with an AI.”
Voice bot on the hotline
An announcement at the start of the call, not only when asked.
Internal knowledge assistant
The notice applies internally too; for tools clearly named as AI it may be obvious.
AI images in marketing
The tool marks them machine-readably; you visibly label realistic people or events.
AI text in a press release
Labelling needed unless a person reviewed it editorially and takes responsibility.
Translation and proofreading
Pure assistive functions such as spell-checking are exempt from the marking duty.
Implementing Article 50: checklist for AI assistants.
Tick off what you’ve already clarified. Nothing is stored.
Code of Practice, supervision and fines.
The Commission’s voluntary Code of Practice covers marking content (para. 2) and labelling deepfakes and texts (para. 4). It does not cover the chatbot notice under para. 1. Signing is voluntary; those who do not sign must show in another way that they meet the duties.
Breaches of Article 50 can be fined up to €15m or 3% of worldwide annual turnover, whichever is higher. In Germany, under the AI market surveillance act (KI-MIG), the Bundesnetzagentur is the central authority.
Questions about EU AI Act Article 50.
What does Article 50 of the AI Act regulate?
Article 50 sets transparency duties for certain AI systems. People must be told when they interact with an AI. AI-generated content must be marked in a machine-readable way. Emotion recognition, deepfakes and AI texts on public matters must be disclosed. The duties have applied since 2 August 2026, regardless of risk class.
Does a website chatbot need an AI notice?
As a rule, yes. A chatbot is a system that interacts directly with people. The notice must be clearly visible, at the first message at the latest, and must not be hidden in the terms or the footer. The only exception is when it is obvious to a reasonable person that they are talking to an AI.
Does Article 50 also apply to open-source AI?
Yes. The AI Act exempts free and open-source AI from duties in several places, but not from the transparency duties in Article 50. Anyone who runs or offers a chatbot based on open-source software must show the AI notice in the same way. The software licence changes nothing about this duty.
Must older AI content be labelled retroactively?
No. Content generated before 2 August 2026 does not need retroactive marking, though the Commission recommends it where feasible. For generative systems already on the market before that date, a grace period until 2 December 2026 applies to machine-readable marking.
What fines apply for breaches of Article 50?
Breaches of the transparency duties can be fined up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. For small and medium-sized companies the lower amount applies. In Germany the Bundesnetzagentur is the central supervisory authority for the AI Act.
Who is the provider and who the deployer?
The provider develops an AI system and places it on the market under its own name. The deployer uses it under its own responsibility, such as a company with a chatbot on its website. Anyone who adapts a system heavily or offers it under their own name can become a provider. Clarify your role with your legal counsel.
Do you know which risk class your AI falls into?
The free risk class check places your AI use in a few minutes. Or try RAGSuite free for 7 days.