On this page
If people talk to a machine, they must know
The EU AI Act transparency obligations apply since 2 August 2026. Article 50 asks three things: a chatbot or other AI that talks to people says it is AI; providers of generative AI mark their output so a machine can detect it; and anyone publishing a deepfake, or unreviewed AI text informing the public on matters of public interest, discloses that it is AI.
The Digital Omnibus on AI delayed the high-risk rules but left these on schedule, and the EU AI Office and national authorities enforce them since August. They reach every chatbot used by people in the EU, including the support bot of a US software company or a UK online shop.
Company with a website chatbot
Act now
Visitors must know they talk to AI, unless that is already obvious.
Put a line at the top of the chat, in every language you serve.
Online shop with AI images
Check it
Photorealistic people wearing your product who do not exist are deepfakes.
Label those images visibly, or use real photos.
Agency producing content
Act now
You publish for clients: deepfakes and unreviewed AI text on public matters need disclosure.
Agree per client who reviews and who labels.
Non-EU company with EU customers
Act now
The rules apply to any system used in the EU, wherever the provider sits.
Show the notice in German, French or Dutch too, not only in English.
Internal AI tool, no public
Be aware
Summarising or translating internally triggers no outward disclosure.
Do take measures to support your team's AI literacy.
The machine-readable mark on AI output sits with the provider of the system, not with you as a user.
Provider or deployer: which role is yours
The provider develops an AI system, or has it developed, and places it on the market under its own name. The deployer is the business that uses it. A Dutch SaaS company that ships an in-app assistant to customers across Europe is the provider of that assistant, even when the model underneath comes from OpenAI or Anthropic. A German online shop that embeds a vendor's chat widget is a deployer; the vendor builds the disclosure in, the shop makes sure it shows. Building your own bot on your documents, for example with retrieval-augmented generation, makes you the provider; see how to build a chatbot and AI chatbot development.
Five duties, and when each one lapses
| Duty | Who | Lapses or eases when |
|---|---|---|
| Tell people they interact with AI | Provider | A reasonably well-informed person can already tell from the context |
| Mark output machine-readably | Provider of generative AI | The AI only does standard editing or does not substantially change your input |
| Disclose a deepfake | Deployer who publishes | The work is evidently artistic, creative or satirical: a disclosure that does not spoil it suffices |
| Disclose AI text on matters of public interest | Deployer who publishes | A person reviewed it and someone holds editorial responsibility |
| Inform people about emotion recognition or biometric categorisation | Deployer | The law permits it to detect or prosecute crime |
The information has to be clear, given at the latest at the first interaction, and accessible to people with disabilities.
Questions about chatbots and labels
Our company is in the US. Does Article 50 really apply to us?
Yes, when your AI system is placed on the EU market or its output is used in the EU. Where you are incorporated does not matter.
Must I label every text I wrote with AI?
No. The text duty covers AI text informing the public on matters of public interest, and it lapses once a person reviews it under editorial responsibility. Product copy or a post about your service is usually outside it.
What counts as a deepfake?
Image, audio or video generated or manipulated with AI that resembles existing persons, objects, places, entities or events and would falsely appear authentic.
Does an assistant that drafts replies for my support team count?
When an agent reads and sends the reply, the AI does not interact directly with the customer, so no notice to the customer is needed. Your team still needs to know what the tool does.
The line at the top of your chat window
Do not rely on "obvious". A chat window with a first name and a friendly tone reads as an employee to many visitors. One line at the start removes the doubt, in every language you serve:
- "You are talking to our AI assistant. Ask for a colleague and we take over."
- "Sie sprechen mit unserem KI-Assistenten. Fragen Sie nach einem Kollegen, dann übernehmen wir."
- "Vous parlez avec notre assistant IA. Demandez un collègue et nous prenons le relais."
A voice bot says it at the start of the call. Translating the notice is the easy part of serving several markets; how we build sites in more languages is in multilingual website development, and how a bot hands over to a person in AI customer service.
A bot that says what it is should also say what it does not know. A model that hallucinates your prices is not a transparency breach, but it is a problem with your customer. Let it answer from your own sources and pass the conversation on when unsure. What running one costs is in AI chatbot cost; if you want a model hosted in Europe, see is there a European chatbot.
Who enforces, and what a breach costs
For most chatbots the market surveillance authority of the member state is in charge. The EU AI Office takes systems offered by the maker of the underlying general-purpose model, and systems inside very large online platforms. Anyone can file a complaint through the AI Office's complaint tool. Breaching Article 50 costs up to EUR 15 million or 3 percent of worldwide turnover; for SMEs the lower of the two applies.
What the omnibus left in place
High-risk AI moved to December 2027 and August 2028; see the EU AI Act high-risk delay. Article 50 kept 2 August 2026, with extra time only for marking by older generative tools; see AI-generated content labelling. AI literacy has applied since February 2025, and since the omnibus it means taking measures that support your team's knowledge. A written policy is not mandatory but helps show it; read do I need an AI policy. Our EU AI Act checker shows the duties per application.