How must synthetic content be marked and made detectable?

Which outputs must be marked?

Article 50(2) applies to providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text content. The duty concerns the output of the system, not the system as a whole, and it applies to content that is artificially generated or manipulated by the system. Teams should inventory the output types a system produces, the channels through which they are delivered, and the points at which marking can be applied, before choosing a marking method. An inventory that lists only the main output type, while the system also produces thumbnails, transcripts, or compressed versions that lose the marking, will miss the outputs where detection matters most. Each output type and each delivery path should be assessed separately, with the marking approach recorded for each.

What does machine-readable marking mean?

Machine-readable marking means a mark embedded in or attached to the output in a form that a machine can read and use to detect that the content was artificially generated or manipulated. In general terms, that can include metadata, watermarks, cryptographic provenance, or other technical methods that meet the effectiveness, interoperability, robustness and reliability standard in the article. The method should survive the normal handling the content will undergo, as far as technically feasible, including compression, resizing, or format conversion that the provider can reasonably foresee. A mark that is present only in a form that is stripped by the first normal handling step does not meet the purpose of the duty. The provider should document the method chosen, the standard or technique relied on, and the testing carried out to show that detection works after foreseeable handling.

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

How are effectiveness, robustness and reliability judged?

Effectiveness, interoperability, robustness and reliability are judged as far as technically feasible, taking into account the type of content, the costs of implementation, and the generally acknowledged state of the art. That standard is not a promise that a mark can never be removed, and it recognises that no method is perfect against a determined adversary. It does require a provider to choose a method that is appropriate to the content, to test it against foreseeable removal and handling, and to document the limits found. A provider should record the threats considered, the tests performed, the results, and the residual limits, and should revisit the method as the state of the art changes. A method chosen without testing, or a method whose limits are known but not documented, cannot be shown to meet the standard.

What exceptions apply to marking?

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 provided by the deployer or the semantics thereof, subject to the wording in the consolidated text. That exception is narrow and fact-specific: it turns on whether the function is assistive and standard, and whether the input is substantially altered in content or meaning. A system that generates new content, changes the meaning of an image or text, or creates a realistic depiction that did not exist before is unlikely to fall within the exception. The provider should record, for each function, whether the exception is relied on, the reasoning, and the evidence, and should seek qualified review where the function sits near the boundary between assistive editing and generation.

How does the December 2026 transition work?

Article 50 transparency duties apply from 2 August 2026. For certain providers of AI systems generating synthetic content that were already placed on the market before 2 August 2026, Regulation (EU) 2026/1744 provides a transitional deadline of 2 December 2026 to comply with the Article 50(2) marking and detection obligation. That transition is limited to that obligation and to those systems; it does not postpone the other Article 50 duties, which apply from 2 August 2026. A provider relying on the transition should record the placing date, the system version, and the plan to meet the marking duty by 2 December 2026, with dated milestones. A provider that treats the transition as a general postponement of transparency duties will have missed the duties that already apply.

Continue with December 2026 transition for Article 50 marking.

What does this guide not decide?

This guide does not decide whether a particular output must be marked, whether a marking method is effective, robust or reliable, whether the assistive editing exception applies, or whether a system qualifies for the December 2026 transition. It does not choose a marking technique or assess a detector. Those questions depend on the output type, the system function, the marking method and testing, the placing date, and the consolidated legal text. A provider should inventory its outputs, test its marking method, record the transition position where relied on, and confirm the position with a competent authority or legal adviser where the function or method is borderline.

Frequently asked questions

Continue through this AI Act cluster

• Transparency duties for AI
• December 2026 transition for Article 50 marking
• Marking methods in general terms
• Exceptions in Article 50
• Provider responsibilities
• When must people be told they are interacting with AI?