Chatbot and deepfake transparency obligations
Article 50 requires that users be informed when they are interacting with an AI system or consuming AI-generated content. The goal is to ensure transparency so that individuals can distinguish between human-authored/human-led interactions and those produced by artificial intelligence.
What it means
The obligation applies primarily to three scenarios: chatbots, deepfakes (synthetic images, audio, or video), and AI-generated text intended for public information on matters of public interest. The core intent is to prevent deception and ensure the user's right to know they are not interacting with a human being.
For chatbots, the disclosure must be explicit unless it is already obvious from the context that the interaction is synthetic. For media—such as images or audio—the requirement extends beyond simple visual labels; content must be marked in a machine-readable format to ensure transparency persists even if a visible label is removed.
Regarding AI-generated text, the obligation focuses on content meant to inform the public on matters of general interest. This ensures that synthetic information masquerading as human journalism or official communication is identifiable.
How to meet it
- Implement clear and prominent disclosure notices at the start of any chatbot interaction (e.g., "You are now interacting with an AI assistant").
- Integrate machine-readable markers, such as digital watermarks or metadata tags, into all synthetic images, audio, and video files generated by the system.
- Apply visible labels to AI-generated media that clearly state the content is artificially generated or manipulated.
- Ensure that any text produced for public interest purposes contains a clear disclaimer indicating its AI origin.
- Review User Interface (UI) designs to ensure transparency notices are placed where users will see them before the interaction begins, rather than buried in Terms and Conditions.
- Establish a technical pipeline that automatically attaches required metadata to outputs without requiring manual intervention by the end-user.
Evidence an auditor asks for
- Screenshots or screen recordings of the user interface showing the disclosure messages provided to chatbot users.
- Technical specifications describing the machine-readable marking method used (e.g., implementation of C2PA standards or specific metadata fields).
- A copy of the internal policy governing how and when transparency labels are applied to generated content.
- Validation logs or test reports demonstrating that AI-generated files contain the required markers upon export.
Common pitfalls
- Placing disclosures in a "Terms of Service" link rather than presenting them prominently at the point of interaction.
- Using ambiguous language (e.g., "Enhanced by technology") instead of explicitly stating the content is AI-generated or that the entity is an AI system.
- Relying solely on visual watermarks that can be easily cropped out, while failing to implement the required machine-readable metadata.