What marking methods exist in general terms?

What kinds of marking exist in general terms?

In general terms, marking methods fall into several families. Metadata marking attaches information to the file or stream, such as provenance, generator, and creation details, in a structured form that a reader can parse. Watermarking embeds a signal in the content itself, visibly or invisibly, that can be detected by a reader even after the file is copied. Cryptographic provenance uses signed assertions about the origin and history of the content, allowing a reader to verify the chain of custody. Detection approaches use classifiers or forensic signals to identify synthetic content without a mark, and can complement marking but do not replace the marking duty. Each family has different strengths, different failure modes, and different compatibility with delivery paths, and a provider should understand those differences before choosing a method for a content type.

How should a method be chosen for a content type?

Choice should start from the content type and the delivery path. An image delivered through social platforms, a video delivered through streaming, an audio file delivered through a podcast feed, and a text output delivered through an application interface each undergo different handling, compression, and transformation, and a method that survives one path may not survive another. The provider should list the output types, the handling each undergoes, the readers or detectors that should be able to detect the mark, and the interoperability requirements for those readers. The choice should be recorded with the reasons, the alternatives considered, and the limits found. A single method chosen for all content types, without considering delivery paths, is unlikely to be effective for every type, and the file should show that the choice was made type by type.

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

How should robustness be tested?

Robustness testing should subject marked content to the handling it will foreseeably undergo, including compression, resizing, cropping, format conversion, re-encoding, and sharing through the channels the provider supports, and should check whether the mark remains detectable afterwards. Testing should also consider removal attempts that are reasonably foreseeable, without requiring the provider to defeat a determined adversary in every case. The test should record the method, the handling applied, the detection result, and the limits found, for each content type and delivery path. Where a mark does not survive a foreseeable handling step, the provider should consider a complementary method, a change to the delivery path, or a documented limitation, rather than leaving the gap silent. Testing should be repeated when the method, the content type, or the delivery path changes.

How do standards and codes of practice help?

Standards, specifications, and codes of practice can help a provider choose and implement a marking method, by setting out interoperable formats, provenance models, and good practice for marking and labelling. A code of practice on marking and labelling of AI-generated content, facilitated at Union level, can provide a practical framework for providers and deployers, including common icons and disclosure approaches for certain content. Following a standard or code does not by itself prove that the Article 50(2) standard is met for a particular system, and the provider should still test its method and document its limits. The provider should record which standard or code was followed, which version, and how it maps to the marking method used, so that a reviewer can see the basis for the choice and the testing that supports it.

How should marking be combined with visible disclosure?

Machine-readable marking and visible disclosure serve different purposes and should be combined where both duties apply. Marking makes content detectable by machines and by readers that check provenance, while visible disclosure informs the human audience that the content is artificially generated or manipulated. A provider or deployer should design both together, so that the visible label and the machine-readable mark are consistent, refer to the same content, and travel together as far as possible when content is shared. The design should be tested in the publication channel, checking that the visible label remains visible and the mark remains detectable after sharing. A design where the label is present but the mark is stripped, or the mark is present but no human can tell the content is synthetic, leaves one of the two purposes unmet.

Continue with Article 50 marking and detection for synthetic content.

What does this guide not decide?

This guide does not endorse any vendor, product, standard, or method, and it does not decide whether a particular marking method meets Article 50(2) for a specific content type or delivery path. It does not assess a watermark, a metadata scheme, or a provenance system. Those questions depend on the method, the testing, the content type, the delivery path, and the state of the art at the time, together with the consolidated legal text. A provider should test its chosen method, document its limits, keep the choice under review as standards and practice develop, and confirm the position with a competent authority or legal adviser where the method is novel or the limits are significant.

Frequently asked questions

Continue through this AI Act cluster

• Transparency duties for AI
• Article 50 marking and detection for synthetic content
• Synthetic content workflows for transparency
• Codes, standards and sandboxes overview
• Provider responsibilities
• When must people be told they are interacting with AI?