Key takeaways
- Providers must design chatbots, voice agents and avatars so that people are informed, at the start of the first interaction, that they are dealing with an AI system, unless that is obvious.
- Generative systems must mark their outputs in a machine-readable way, and deployers must disclose deepfakes and AI-written text published to inform the public on matters of public interest.
- Article 50 applies from 2 August 2026, with a grace period to 2 December 2026 for the marking duty on systems already on the market.
- A voluntary Code of Practice on marking and labelling AI-generated content has been confirmed as an adequate way to show compliance.
When do you need an AI notice?
If your chatbot, voice agent or avatar speaks directly to people, design it to identify itself as AI at the first interaction unless that is obvious. Article 50 has applied since 2 August 2026. Check your role separately for generated content: provider marking and deployer disclosure are different duties.
What Article 50 requires
| Paragraph | Who | What | Example |
|---|---|---|---|
| 50(1) | Providers of systems that interact directly with people | Design the system so people are informed that they are dealing with AI, unless obvious from the context | A support chatbot, a voice agent answering the phone, a sales avatar |
| 50(2) | Providers of generative AI systems | Mark audio, image, video and text outputs in a machine-readable format so they can be detected as AI-generated | An image generator, a text-to-speech tool |
| 50(3) | Deployers of emotion-recognition or biometric-categorisation systems | Inform the people exposed to the system that it is operating | A system that infers emotions from faces or voices |
| 50(4) | Deployers of deepfake generators and of AI text published to inform the public | Disclose that the content is artificially generated or manipulated. Text is exempt where a human has reviewed it and someone holds editorial responsibility | A synthetic video of a real person. An AI-written news article |
When is it obvious that it is AI?
Assess whether the intended audience would recognise the interaction as AI in its actual context. The Commission says to interpret the exception narrowly. A robot icon or synthetic voice alone is not a dependable basis for skipping a clear notice.
The duty applies only where the system is an AI system, allows a genuine two-way exchange, interacts directly with natural persons and does not work through a human intermediary. An AI that drafts a reply which a person reads, edits and sends is a different situation from one that answers the customer itself. When in doubt, disclose. It costs one sentence and removes the argument.
What good disclosure looks like
| Channel | Example wording | Where to put it |
|---|---|---|
| Web chat | You are chatting with an automated assistant. You can ask for a person at any time. | First message, visible before the user types |
| Voice call | Hello, you have reached our automated assistant. I can help with appointments and opening hours, or pass you to a colleague. | First sentence of the call |
| Messaging apps | This is our AI assistant. Reply HUMAN to reach a person. | First reply in the thread |
| Generated images and audio | A visible or audible label, plus machine-readable metadata in the file | On the content and in the file |
The notice has to be clear, distinguishable and accessible, and it has to come at the first interaction at the latest. A line buried in the terms and conditions does not do the job.
Who is responsible when you buy a chatbot?
The provider has to design a system that makes the disclosure possible, so a bought chatbot should ship with it. As a deployer, you should check that the notice is switched on, never remove or hide it, and put the duty in the contract. If you build a system yourself, or commission one and put it into service under your own name, you are the provider and the duty is yours.
On marking, the 50(2) duty falls on providers of generative systems, but deployers carry the 50(4) disclosure for deepfakes and published text. A company using a generative tool for marketing is usually the deployer.
Dates, grace period and the Code of Practice
- 2 August 2026: Article 50 began to apply.
- 2 December 2026: end of the grace period for providers of generative systems that were already on the market before 2 August 2026, for the 50(2) marking duty only. This date was set by the Digital Omnibus on AI, Regulation (EU) 2026/1744.
- Content generated before 2 August 2026 does not have to be labelled retroactively.
- A Code of Practice on the transparency of AI-generated content, covering marking and detection for providers and labelling for deployers, has been published. It is voluntary, and the Commission and the AI Board have confirmed it as an adequate tool, but following it does not replace the legal duty.
Breaching Article 50 falls in the middle tier of Article 99: fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher, and for SMEs the lower of the two.
Six steps to put it in place this week
- List every chatbot, voice agent, avatar and generative tool your company uses or sells.
- For each one, write the disclosure sentence in every language you serve.
- Put the sentence at the start of the interaction and test it on a phone, a laptop and a screen reader.
- Add a way to reach a person, even where the law does not require it.
- Check your generative tools for built-in marking, and ask vendors how they meet 50(2).
- Keep a dated record of the wording and where it appears.
Tell us where your chatbot or voice agent meets users. We can help you build clear notices, accessible interactions and a route to a person.
Discuss AI disclosureFAQ
Do voice agents need to say that they are AI?
Yes. Article 50(1) is not limited to text. It covers any AI system intended to interact directly with people, so a voice agent should say so in its first sentence unless it is obvious from the context.
Do internal AI assistants for employees need a disclosure?
Article 50(1) also covers direct interactions with employees. Assess the actual interface and audience before relying on the obviousness exception. A clear notice at first use avoids leaving employees to infer how the system works.
Does our blog need an AI label if AI helped write it?
Article 50(4) applies to AI-generated text published to inform the public on matters of public interest. It does not apply where a human has reviewed or edited the text and a person or organisation holds editorial responsibility for it.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 50 and Article 99, Official Journal of the EU
- European Commission, FAQ on transparency obligations under Article 50 of the AI Act
- European Commission, Code of Practice on transparency of AI-generated content
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of the EU, 24 July 2026
About the author
Artem Synytskyi · Co-founder · AI
Artem runs AI at ESADEX, building LLM, RAG and speech-to-text systems for regulated environments.