What exceptions and limits apply in Article 50?

How should exceptions be approached as a whole?

Exceptions in Article 50 should be approached paragraph by paragraph, because an exception for one paragraph does not create an exception for another. A system can be exempt from interaction disclosure because disclosure is obvious, while still facing a marking duty for its synthetic outputs, and a deployer can be exempt from a text disclosure because of human review, while still facing a deepfake disclosure duty for video content. Teams should therefore create an exception map that lists each Article 50 paragraph, whether it applies, whether an exception is relied on, the conditions for that exception, and the evidence. The map should be reviewed when the system, the content, or the deployment changes, because a change in context can remove the basis for an exception that previously applied. A single blanket claim that Article 50 does not apply should never be used.

What is the obviousness exception in Article 50(1)?

Article 50(1) does not require disclosure where it is obvious from the circumstances and context of use that a person is interacting with an AI system. Obviousness is judged from the perspective of the interacting person in the actual context, and the provider should test it with representative users. The exception is narrow in practice for systems that mimic human interaction, use human identities, or operate through channels where a human would be expected. The assessment should record the interface, the name and presentation, the channel, and the user test relied on. Where the context changes, for example where a system moves from a clearly labelled AI channel to a telephone or messaging channel where a human would be expected, the obviousness assessment must be redone. A provider should not carry an old obviousness conclusion into a new context without a fresh assessment.

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

What is the assistive editing exception in Article 50(2)?

Article 50(2) provides that the marking duty does not apply where the system performs an assistive function for standard editing or does not substantially alter the input data or its semantics. The exception turns on two questions: whether the function is assistive and standard, and whether the input is substantially altered. Standard editing includes, in general terms, routine corrections, formatting, and minor enhancements that do not change the meaning or create new content. Generation, substantial manipulation, or creation of realistic content that did not exist before falls outside the exception. The provider should assess each function separately, record the reasoning, and seek qualified review where a function sits near the boundary. A system that offers both assistive editing and generation should apply the exception only to the assistive functions, and should mark the generated outputs.

What law enforcement exceptions exist?

Article 50 provides exceptions 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, in paragraphs (1) and (3), and with a distinct scope for each paragraph. The exception requires legal authorisation, not merely a law enforcement purpose, and the safeguards must be appropriate to the system and the deployment. The deployer or provider relying on the exception should record the legal basis, the authorising law, the safeguards, the scope of the system and deployment covered, and the review date. Where a system is used for both authorised law enforcement purposes and other purposes, the exception should be limited to the authorised use, and the disclosure duties should be met for the other uses. The exception should be reviewed when the authorising law or the system changes.

How do artistic works and human review fit together?

Article 50(4) provides distinct treatments for evidently artistic, creative, satirical, fictional or analogous works, and for text that has undergone human review with editorial responsibility. The artistic treatment concerns the manner of disclosure appropriate to the work, while the human review exception concerns whether the text disclosure duty applies at all. Both treatments require a factual assessment and a record: for artistic works, the nature of the work and the manner of disclosure; for human review, the review process, the reviewer, and the editorial responsibility holder. Neither treatment should be applied as a blanket exemption for a publisher or a platform. Each work and each publication should be assessed on its facts, and the assessment should be kept with the publication record, so that a later question about a specific item can be answered from the file for that item.

Continue with Article 50 interaction disclosure.

What does this guide not decide?

This guide does not decide whether any exception applies to a particular system, function, content item or deployment. It does not assess obviousness, assistive editing, legal authorisation, artistic character, or human review. Those questions depend on the facts, the context, the legal basis, and the consolidated legal text for the paragraph in question. A provider or deployer should map each paragraph separately, record the exception relied on with its conditions and evidence, and confirm the position with a competent authority or legal adviser where the exception is borderline or where reliance on it would remove a disclosure that an audience would otherwise expect.

Frequently asked questions

Continue through this AI Act cluster

• Transparency duties for AI
• Article 50 interaction disclosure
• Article 50 marking and detection for synthetic content
• Deepfakes and public-interest text
• Provider responsibilities
• What must deployers disclose for emotion recognition and biometric categorisation?