AI Act Article 50: How to Label Your Chatbot and AI Content
Article 50 of the EU AI Act requires that people are told when they talk to an AI and that deepfakes and certain AI-written text are labelled. The chatbot and deepfake duties have been enforceable since 2 August 2026, with no grace period. The Digital Omnibus pushed the high-risk rules to 2 December 2027 and left the rest of Article 50 alone. It granted one narrow transition: machine-readable marking for generative systems already on the market before 2 August 2026 must be in place by 2 December 2026.
Many product teams read the Omnibus headlines and relaxed. The delay covers the expensive part of the Act, not the cheap part. If your chatbot does not say it is an AI before the first message, you are probably out of compliance today, and the fix takes an afternoon.
What Article 50 Actually Requires
- Art. 50(1) (Providers): Systems interacting directly with humans must be designed so users know they are dealing with AI, unless it is obvious from the context. See the Article 50 text for the full requirement. This is a core part of EU AI Act compliance for builders.
- Art. 50(2) (Providers): Systems generating synthetic audio, image, video, or text must mark outputs in a machine-readable, detectable format that is as robust and reliable as technically feasible.
- Art. 50(3) (Deployers): Users of emotion recognition or biometric categorisation systems must inform the people exposed to those systems and process data under GDPR.
- Art. 50(4) (Deployers): Deepfakes must be disclosed as AI-generated. AI-generated text published to inform the public on matters of public interest must be disclosed, unless it went through human review and someone holds editorial responsibility.
- Art. 50(5) (Providers & Deployers): All transparency information must be clear, distinguishable, and provided at the latest at the first interaction or exposure, meeting accessibility requirements.
Are You the Provider or the Deployer?
The law distinguishes between the provider (who develops the AI) and the deployer (who uses it). In practice, if you build a chatbot on top of a model API and put it on your site under your own brand, you are likely the provider of that system. You cannot delegate the Article 50(1) disclosure duty to the model vendor.
If you buy a finished SaaS chatbot and use it as is, you are the deployer. If you wrap an LLM in your own code and UI, you are most likely the provider. Your role determines which specific duty you carry.
How to Label a Chatbot
Disclosure must happen before or during the first interaction. Placing a notice in the footer or hidden within the Terms of Service is insufficient. The notice should appear above the input box or as the first message sent by the bot. Put it in the first sprint when you are building LLM features for production.
If the bot hands over the conversation to a human agent, and later back to the AI, the status must remain clear. Voice bots must state they are AI audibly at the start of the call. For accessibility, use screen-reader readable text rather than relying solely on an icon.
Do not rely on the “obvious from context” exception if your bot uses a human name and a realistic avatar. The exception is narrow and rarely applies to professional customer-facing interfaces.
How to Label Images, Video and Voice
Deepfakes, meaning image, audio or video that would falsely appear authentic, must be clearly disclosed as AI-generated or manipulated. For content that is evidently artistic, satirical, or fictional, you only need to disclose that it was generated or manipulated, in a way that does not spoil the work.
The Commission published its final, voluntary Code of Practice on 10 June 2026 (see this summary). For deployers it recommends the EU icon or an equivalent label on deepfakes and public-interest text. For audio content, an audio disclaimer is a valid alternative to a visual label.
AI-Written Text: When You Need a Label and When You Do Not
The requirement to label AI-generated text applies specifically to content published to inform the public on matters of public interest. A product description or a support answer is not that; a news-style article or a public statement is.
There is an exemption for text that has undergone human review or editorial control. If a named human editor reviews the content and the company takes editorial responsibility, the disclosure duty is removed. An auto-published news feed that pumps LLM output directly to the web without a human in the loop does not qualify for this exemption.
Machine-Readable Marking: the 2 December 2026 Deadline
This requirement applies only to providers who generate synthetic content. Systems launched after 2 August 2026 had no grace period. Systems already on the market before that date must implement machine-readable marking by 2 December 2026.
The Code of Practice asks providers for at least two layers of machine-readable marking, for example metadata plus a watermark. The European AI Board’s plenary on 17 September 2026 set no new deadline and gave no relief, according to the CSA research note.
Test your markings: they have to survive your own pipeline, including image resizing, re-encoding, or the stripping of metadata during copy-paste operations.
Breaches fall under the Act’s general penalty tier in Article 99: up to EUR 15 million or 3 percent of worldwide annual turnover, whichever is higher. This article is a builder’s guide, not legal advice. Edge cases, such as emotion recognition in the workplace or law-enforcement exemptions, should be referred to legal counsel.
A 30-Minute Article 50 Check
- Does the chatbot disclose its AI nature before the user sends the first message?
- Is the disclosure visible and accessible (not hidden in a footer or Terms of Service)?
- If you publish public-interest text, is there a named human editor responsible for the content?
- Are synthetic images or videos marked with the EU icon or an equivalent label?
- If you are a provider of generative content, is machine-readable marking implemented?
- Do the machine-readable marks survive your content delivery pipeline?