EUAICheck is an independent self-assessment tool. We are not a law firm and have no link with any official authority.
EUAICheck is an independent self-assessment tool published by Vival Solutions (EI). This content is informational, is not legal advice and may change: always check the official text on EUR-Lex.
EU AI Act and chatbots: Article 50 applies since 2 August 2026
Updated 30 September 2026 · Article 50 of Regulation (EU) 2024/1689 · Commission guidelines of 20 July 2026
Since 2 August 2026, a person chatting with an AI chatbot or assistant must be told they are interacting with AI, unless it is obvious. This obligation was not postponed by the Digital Omnibus.
1. The four Article 50 obligations
- Interacting with AI (§1): systems intended to interact directly with people are designed so that people know they are dealing with AI, unless obvious to a reasonably well-informed person.
- Marking generated content (§2): audio, image, video or text outputs generated by an AI system are marked in a machine-readable format.
- Emotion recognition and biometric categorisation (§3): people exposed are informed.
- Deepfakes and public-interest text (§4): deepfake content, and generated text published to inform the public on matters of public interest, is disclosed as generated or manipulated, unless the text underwent human review with editorial responsibility.
2. Who carries which obligation?
| Obligation | Responsible party |
|---|
| Disclosing interaction with AI (§1) | Provider of the system |
| Machine-readable marking (§2) | Provider of the system |
| Emotion recognition, biometric categorisation (§3) | Deployer |
| Deepfakes, public-interest text (§4) | Deployer |
Mind the role. If you build your own chatbot on a model API (OpenAI, Anthropic, Mistral…) and put it into service under your name, you are in principle the provider of that system: the §1 disclosure is yours. If you use an off-the-shelf chatbot, its vendor designs the disclosure first; in practice, check that it actually appears on your site.
3. Your chatbot: practical cases
- Customer service chatbot on an e-commerce site: disclosure required, unless the context makes the AI obvious.
- AI assistant in an internal tool used by your teams: disclosure is in principle also required, though the context may make it obvious.
- AI agent acting on behalf of a company: per the guidelines, it states that it is AI and on whose behalf it acts.
- Marketing text (blog, product pages): §4 only covers text published to inform the public on matters of public interest; promotional articles are generally not covered. Realistic generated images or videos may, however, be deepfakes.
According to the Commission guidelines published on 20 July 2026, the information must be clear, noticeable and given at the latest at the first interaction. In practice:
- An explicit greeting: "I am an AI assistant".
- A visible "AI" label in the chatbot interface, in the user's language.
- No instruction that would make the assistant pass as human.
- For generated content: enable the marking offered by the tool (metadata, watermark, C2PA or equivalent). A code of practice on transparency of AI-generated content has been recognised by the Commission.
5. Penalties and enforcement
- Breaching Article 50 falls under the Article 99(4) ceiling: up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher.
- For SMEs and start-ups, the lower of the two amounts applies (Article 99(6)).
- Enforcement is carried out by market surveillance authorities designated by each Member State. In France, the Government's announced scheme gives coordination to the DGCCRF, with other authorities by use case (CNIL, Arcom…); check the status of the designation.