Labelling AI-written articles and synthetic media under Article 50

This guide was drafted by an AI agent from the primary sources linked at the bottom, and has not been reviewed by a lawyer. It is general information, not legal advice. The original is at https://mdodeja.github.io/eu-ai-act-article-50-guide/guides/labeling-ai-written-articles.html.

Short answer. AI-written text on matters of public interest needs a disclosure unless a human reviewed it and a named person or organisation holds editorial responsibility. Realistic synthetic images, audio and video need a visible label either way.

Two separate duties

Article 50(4) puts two duties on deployers of generative AI, and agencies often mix them up.

Deepfakes. If you use AI to generate or manipulate image, audio or video content that realistically resembles real people, places or events, you have to disclose that it is artificially generated or manipulated.
Text on matters of public interest. If you use AI to generate or manipulate text published to inform the public on matters of public interest, you have to disclose that it is AI-generated.

The marking duty in Article 50(2) is separate again, and it belongs to the provider of the generation tool. The marking guide covers it.

The editorial-responsibility exception

For the text duty, Article 50(4) has an exception. You do not have to disclose AI generation if the content has been through human review or editorial control, and a natural or legal person holds editorial responsibility for its publication.

That gives an agency a practical route for AI-drafted articles:

Every AI-drafted article goes to a named human editor before it is published.
The review is substantive. The editor checks facts and changes the piece, because they are the person answerable for mistakes.
The site says who holds editorial responsibility, for example on an about page, a masthead, or a byline convention.
You keep a record per article or per pipeline: the editor's name, the date, and what the review involved.

If no human is in the loop, label the piece. A plain line works: "This article was generated by AI." It should appear at or before first publication, where readers can see it.

The Act does not spell out how deep a review has to be. The Commission's guidelines on Article 50 address the scope of these duties, so read them for your use case, and treat a skim as too little.

What counts as public-interest text

Ordinary marketing copy, such as product pages and category descriptions, is generally not text published to inform the public on matters of public interest. Content that reads like news, commentary, or civic information is. When a piece could be taken for either, handle it as public-interest text.

Images, audio and video

For realistic synthetic media, a metadata mark alone is not enough. The disclosure has to be visible to the person looking at it.

Put a label on or next to the asset, such as "AI-generated image". Make it readable and not fine print.
For video the label should persist. For audio, state it audibly or show it.
AI avatars and synthetic spokespeople count.
Evidently artistic, creative, satirical or fictional works get a lighter disclosure that does not get in the way of enjoying the work, for example a credit line. Keep a note of why you used that exception.
Using synthetic media of a real, identifiable person without their consent is a legal risk beyond Article 50. Take that to a lawyer before you proceed.

The Commission's Code of Practice includes a set of icons that deployers may use for labelling.

A record to keep for each pipeline

Pipeline:        ____________   Client: ____________   Date: ________
Classification:  fully AI-generated / AI-assisted / assistive only
50(2) marking:   tool-native metadata kept / C2PA added / not applicable (why)
50(4) media:     visible label applied / not applicable
50(4) text:      disclosed / editorial responsibility held by NAME / not applicable
Evidence:        links or screenshots