EU AI Act AI-generated content labeling requirements

Start with the two-layer model

Teams often collapse all AI transparency work into a single visible badge. Article 50 distinguishes provider-side machine-readable marking from deployer-side disclosure to people.

Machine-readable marking supports technical detection of generated or manipulated output. It belongs in the generation and export pipeline.

Human-visible or audible disclosure tells a person that specified content was generated or manipulated by AI. It belongs at the point of exposure or publication.

The Commission's official questions and answers state that a deployer cannot rely only on the provider's embedded machine-readable marking to satisfy a visible deepfake disclosure duty.

Provider marking workflow

For an AI system that generates synthetic audio, image, video or text, map every export route and transformation step:

1. Identify which component creates or materially manipulates the output.
2. Select marking techniques suitable for each format and distribution channel.
3. Test whether the mark is present, detectable and associated with the correct asset.
4. Measure what happens after resizing, transcoding, screenshots, copy and paste, metadata stripping or social-platform processing.
5. Record known limitations rather than claiming perfect detection.
6. Version the implementation and rerun tests after pipeline changes.

The standard-editing and technical-feasibility details require context-specific assessment against the current Act, Commission guidelines and Code of Practice.

Deepfake disclosure workflow

Where a deployer publishes image, audio or video content that meets the relevant deepfake definition, the disclosure should be clear and distinguishable, understandable without a specialist tool, and available no later than first exposure.

Examples of implementation surfaces include a visible label on an image, a persistent video overlay, a clear caption next to the asset, or an audible statement for audio-only distribution. Artistic, creative, satirical or fictional contexts have specific treatment, but do not create a blanket no-disclosure rule.

Public-interest text workflow

Article 50(4) also addresses AI-generated or manipulated text published for the purpose of informing the public on matters of public interest. The Act contains an exception connected to human review or editorial control and final editorial responsibility.

Do not translate that exception into a checkbox labelled "human reviewed." Retain evidence of who reviewed the text, what substantive changes or fact checks occurred, who could approve or reject publication, and which natural or legal person held editorial responsibility.

Grammar, spelling, formatting or automated moderation alone should not be recorded as substantive editorial review.

Label content checklist

A practical label should answer the most important question in plain language: was this content generated or materially manipulated by AI? Depending on context, add the responsible publisher, a link to further information, the date or version, and an accessible text equivalent.

Avoid labels that are:

• visible only on hover;
• hidden behind a generic information icon;
• removed when the asset is embedded;
• written in a language the audience does not understand;
• contradicted by nearby claims of authenticity;
• used as a substitute for provider-side marking.

Evidence to preserve

Keep the original asset, marked asset, visible-label rendering, validation output, transformation-test matrix, publication URL, wording and component version, timestamps, owners, exceptions analysis and official sources relied upon.

Use the Article 50 content-label generator for a controlled starting point, then validate the label in the actual distribution environment.