Right to AI transparency

What it is

The right to AI transparency is the right to know when you are interacting with an artificial intelligence system and to understand:

• that you are dealing with an AI and not with a person;
• how the system works;
• what data it was trained on;
• what decisions or outputs it produces;
• who is responsible for it.

The two levels of transparency in the AI Act

Minimum transparency (all AI systems)

Art. 52 requires users to be informed when:

• they interact with an AI chatbot or virtual assistant;
• content is generated by AI (text, images, video, audio);
• an AI system categorises or profiles people;
• an AI system recognises emotions.

Enhanced transparency (high-risk AI systems)

Art. 50 requires detailed documentation on:

• the purpose of the system;
• performance, limits and risks;
• training data;
• human oversight measures;
• accuracy, robustness and security.

What transparency is for

Awareness

• You know who, or what, you are dealing with.
• You are not misled into believing you are talking to a person.

Understanding

• You understand how the AI makes its decisions.
• You can assess how reliable its outputs are.

Accountability

• You can identify who is responsible for the system.
• You know whom to turn to with a complaint.

When it applies: minimum transparency (Art. 52)

Chatbots and virtual assistants

When you interact with:

• conversational AI assistants (ChatGPT, Claude, Gemini);
• customer-service chatbots on websites;
• voice assistants (Alexa, Siri, Google Assistant).

The obligation:

• a clear notice: “You are interacting with an AI”;
• before or at the start of the interaction;
• in plain, visible language.

Exceptions. The notice is not required if:

• it is obvious from the context (for example on an AI assistant’s own website, where it is already clear);
• the system clearly assists a person (a declared “co-pilot”).

AI-generated content (deepfakes, images, text)

When content is created or manipulated by AI.

Images and video:

• photos generated with tools such as Midjourney, DALL-E, Stable Diffusion;
• deepfake videos (face swaps);
• images manipulated with AI.

The obligation: a digital watermark or a visible label, such as “AI-generated image” or “Video manipulated with AI”.

Audio:

• voices generated or cloned with AI;
• music created by AI;
• synthetic podcasts.

The obligation: an audio notice or accompanying text, such as “AI-generated voice”.

Text:

• articles written by AI;
• personalised automated emails;
• reports generated by algorithms.

The obligation: the label “AI-generated content”, unless it is already obvious.

Exceptions:

• AI used as creative assistance (editing, suggestions), when a person keeps final control;
• AI content that is obvious by its nature (for example machine translation).

When it applies: enhanced transparency (Art. 50)

It concerns AI systems considered “high-risk” because of their impact on fundamental rights:

• employment: HR systems for recruiting and assessing staff;
• education: AI systems for assessing students;
• finance: AI systems for granting credit.

Mandatory technical documentation

For every high-risk system, the provider must prepare documentation containing:

• a general description of the system;
• its intended purposes and context of use;
• the algorithmic architecture;
• the training data (origin, characteristics);
• performance (accuracy, precision, recall);
• known limitations;
• residual risks;
• bias-mitigation measures;
• human oversight procedures;
• cybersecurity measures.

Accessible documentation

• Available on request to users and authorities.
• In understandable language, not just technical.
• Updated whenever the system changes.