In short: The AI Act transparency obligations are in Article 50 and apply from 2 August 2026, whatever the risk level of your AI. People must know when they’re talking to an AI, unless it’s obvious. Deepfakes must be disclosed. AI-generated content must carry a machine-readable mark. For most businesses, the chatbot is where this starts.
You use a chatbot or an AI assistant to help customers at any hour, and AI tools to write and design faster. That’s what they’re for. The AI Act now adds one simple condition: be open about it. The AI Act transparency obligations are among the lightest duties in the law, and also among the easiest to get wrong without noticing. Here’s what they ask and how to check that your AI keeps to them.
What are the AI Act transparency obligations?
The AI Act transparency obligations are four duties in Article 50 of the EU AI Act. They apply to every AI system that falls under them, high-risk or not. First, people must be told they’re dealing with an AI when it talks with them directly, unless that’s obvious. Second, providers of AI that creates text, images, audio or video must mark that output. The mark must be machine-readable, so the content can be detected as AI-made. Third, anyone using AI to recognise emotions or sort people by biometric traits must tell the people involved. Fourth, anyone publishing deepfakes must say the content is AI-made. The same goes for AI-written text that informs the public on matters of public interest. The information must be clear and come at the latest at the first interaction. These duties have applied since 2 August 2026.
Do you have to tell customers they’re talking to a chatbot?
Yes, unless it’s obvious to a reasonably well-informed person. In the law this is a duty for the provider. That’s the party that develops the AI system, or has it developed, and offers it under its own name. If you bought a ready-made chat widget, the vendor carries this duty. If you built your own assistant on a vendor’s model and put your brand on it, you may well be the provider yourself. Our post on provider or deployer helps you work out which role you have.
Whatever your role, the safe choice is the same. Say it in the first message: “I’m the AI assistant of…”. Don’t give the bot a human name and photo without saying it’s an AI. A small icon in the corner is easy to miss, so don’t rely on it alone.
Do you have to label AI-generated text and images?
It depends on what you make and why. The rules split into two parts.
- Deepfakes. Images, audio or video that look like real people, places or events must be disclosed as AI-made. For clearly artistic or satirical work, a lighter disclosure is enough.
- Text for the public. AI-written text published to inform people about matters of public interest must be disclosed. That duty falls away when a person has reviewed or edited it and someone holds editorial responsibility.
So a product description or a blog that a colleague checks and signs off isn’t caught by the label duty. The machine-readable mark is mostly a job for the makers of AI tools, not for you as a user. Our advice is still to be open where readers would care. Trust is easier to keep than to win back.
The European Commission has been working on guidelines and a code of practice for these labels. Proposals to adjust some AI Act deadlines have also been discussed. Check the current state on the Commission’s AI Act pages before you settle your approach.
How do you check that your chatbot is transparent?
You check it by testing, the same way you’d test any other promise your AI makes. A line in the system prompt saying “tell users you are an AI” is a good start. It’s not proof. Models drift, prompts change, and users push. Even well-designed AI systems drift and misbehave over time. So test these points:
- Does the first message say clearly that this is an AI?
- What happens when someone asks “Am I talking to a real person?” in ten different ways?
- Does the bot stay honest when a user asks it to pretend to be human, or to play a role?
- Does it give the same answer in the other languages your customers use?
- Does it still pass after a new model version or a prompt change?
Keep the questions and the exact answers with a date. That record is what you show when someone asks how you know your AI is open about itself.
What happens if you don’t comply?
Breaking Article 50 can lead to fines of up to 15 million euro or 3% of worldwide annual turnover, whichever is higher. For small and medium-sized companies the lower of the two applies. Our post on EU AI Act fines explains how the levels work.
To keep it in proportion: a fine is not the most likely outcome of a missing disclosure. A customer who finds out later and feels misled is more likely. The fix is cheap, though. A clear first message and a regular test cost far less than either risk. There’s no good reason to wait.
What this means for your team
Make a short list of every place your AI meets people: website chat, email replies, phone bots, generated images. For each one, write down who the provider is and how people are told. Fix the wording where it’s vague. Then add the five test questions above to your regular AI tests and rerun them after every change. Our EU AI Act compliance page shows how this fits with the rest of the law.
Would you rather not run those tests by hand? The VeritasAI PRO plan runs tests mapped to the EU AI Act, ISO 42001, OWASP and NIST every week. You review the exact responses your model gave, so you can see for yourself whether it stays open about being an AI.
Show every week that your AI keeps to the AI Act.
Frequently asked questions
Does Article 50 apply to internal AI tools?
The duty to say people are talking to an AI covers systems that interact directly with people, and staff are people too. For an internal assistant it’s often obvious to users, but saying it costs nothing.
Is a small “AI” label in the chat window enough?
The law asks for clear and distinguishable information at the first interaction. A label that most users overlook may not meet that. A plain sentence in the first message is safer.
Does Article 50 apply if our AI is not high-risk?
Yes. The transparency duties stand on their own. A simple customer service chatbot with minimal risk still has to make clear that it’s an AI.
Do I have to label a blog written with AI?
Not if a person reviewed or edited it and someone takes editorial responsibility. The label duty is for unreviewed AI text that informs the public on matters of public interest.