When must people be told they are interacting with AI?

Which systems face Article 50(1)?

Article 50(1) applies to providers of AI systems that are intended to interact directly with natural persons. That includes, in general terms, chatbots, voice assistants, virtual agents and other conversational or interactive systems where a person engages directly with the system. The test is intended direct interaction, not incidental contact with a system that operates in the background. A system that processes data without direct interaction, or a system whose AI nature is part of a different transparency duty, should be assessed under the paragraph that fits its function. Teams should record, for each interactive feature, whether it is intended for direct interaction, who the interacting person is, and which Article 50 paragraph was considered, so that the disclosure decision is traceable to the feature rather than to a general assumption about the product.

When is disclosure obvious and not required?

Disclosure is not required where it is obvious from the circumstances and the context of use that the person is interacting with an AI system. Obviousness is judged from the perspective of a reasonably well-informed, observant and circumspect person in the actual context, not from the provider's knowledge that the system is AI. A system presented in a way that mimics a human, uses a human name and image without any AI indication, or operates through a channel where a person would reasonably expect a human, is unlikely to meet the obviousness test. A system whose interface, name, and context make its AI nature clear from the first contact may meet it. The assessment should be recorded with the context relied on, and should be tested with representative users, because a provider's view of what is obvious is not a reliable substitute for user testing.

For the broader regulatory framework, see Transparency duties for AI.

How and when must the information be given?

The information must be provided in a clear and distinguishable manner, at the time of the first interaction or, where the article provides, before the interaction begins. It should be prominent enough to be noticed, in language the interacting person can understand, and in a form that works in the channel used, whether text, voice, or visual interface. A disclosure buried in terms and conditions, shown only after a long interaction, or written in technical language, does not meet the requirement. The provider should design the disclosure into the interaction flow, test that it is noticed and understood, and record the design and the test. Where an interaction continues across sessions or channels, the provider should consider whether the disclosure needs to be repeated or reminded, so that a person who returns later is not left without the information.

What exceptions and limits apply?

Article 50(1) provides exceptions, including for AI systems authorised by law to detect, prevent, investigate or prosecute criminal offences, subject to appropriate safeguards for the rights and freedoms of third parties, unless the system is available to the public to report a criminal offence. Other limits in Article 50, including those concerning artistic, creative and similar works, apply to the relevant paragraphs and should be checked in the consolidated text for the system in question. An exception should not be assumed from a general law enforcement or creative purpose; the specific conditions, safeguards, and scope in the article must be checked and recorded. Where an exception is relied on, the file should state the legal basis, the safeguards, and the review date, so that the reliance can be checked if the system or the law changes.

How should disclosure be tested with users?

Disclosure should be tested with representative users in the actual interaction context, checking whether they notice the information, understand that they are interacting with AI, and know what that means for the interaction. Testing should cover different devices, languages, accessibility needs, and levels of familiarity with AI, and should include users who might reasonably expect a human in that context. Findings should be recorded, with the design changes made in response, and the test should be repeated when the interface, the channel, or the context changes. A disclosure that passes a test with internal staff who know the system is AI, but fails with external users who do not, has not been shown to work for its intended audience. User testing is the evidence that the disclosure design achieves its purpose.

Continue with Article 50 marking and detection for synthetic content.

What does this guide not decide?

This guide does not decide whether a particular system must disclose under Article 50(1), whether disclosure is obvious in a specific context, or whether a disclosure design is clear and distinguishable. It does not assess an interface, a script, or a user test. Those questions depend on the system's intended interaction, the context of use, the perspective of the interacting person, and the consolidated legal text. A provider should test its disclosure with representative users, record the assessment and the test, and confirm the position with a competent authority or legal adviser where the interaction is borderline or the exception is relied on.

Frequently asked questions

Continue through this AI Act cluster

• Transparency duties for AI
• Article 50 marking and detection for synthetic content
• Testing disclosures with users
• Provider responsibilities
• EU AI Act overview
• What must deployers disclose for emotion recognition and biometric categorisation?