PPROMPT WORKBOOK Build a prompt
Applies from 2 August 2026

Labelling AI-generated content

This is the part of the AI Act that reaches the most ordinary organisations first, and it was not delayed with the rest.

This is information, not legal advice. Every rule below is cited to its source and reflects the position on 26 September 2026. This area is moving quickly, and a decision that matters for your organisation should be checked by someone with legal responsibility for it.

Four duties, different addressees

Article 50 separates the party that supplies the system from the party that uses it. That distinction matters, because the duty to disclose deepfakes sits with the user of the system. In most cases that is you.

The chatbot must say so from the start

Whoever provides a system people converse with must disclose the AI before or at the beginning of the first interaction. The exception for cases where this is obvious is interpreted narrowly.

Output must be machine-readable

Generative systems must mark synthetic audio, image, video and text in a machine-detectable format, as far as technically feasible. Systems already on the market have until 2 December 2026.

Deepfakes must be disclosed

Anyone generating or manipulating image, audio or video that appears authentic must disclose that it is artificial. This applies even with no intent to deceive. Art and satire get lighter treatment.

Published AI text on public-interest topics

Text published to inform the public on matters of public interest must be disclosed, unless a human carried out genuine editorial review and takes responsibility.

Editorial review is a real threshold

The human-review exemption is the provision most likely to be stretched too far. The Commission is explicit that a superficial check does not count. Running a spell-checker and skimming the result is not editorial responsibility. What is required is that a person with genuine editorial authority went through the content and stands behind what is published.

For anyone publishing professionally this is a workflow question more than a legal one. Either a person reviews the text and signs off on it, or the text gets marked. We have written up what the threshold demands, and what a workflow that can prove it looks like.

This should qualify

  • A named person checks the factual claims
  • That person can change or block publication
  • The review leaves a record you could show
  • Responsibility clearly sits with the publisher

This probably will not

  • A spell-check and a quick skim
  • Sign-off with no power to change the content
  • “Someone looked at it” with no trace of who or when
  • Automatic publishing with review only after complaints

What to do this week

Map where AI actually publishes

Website copy, product descriptions, support replies, newsletters, ad imagery, social posts. Build the list before deciding anything.

Decide a marking practice per channel

A visible line at the foot of articles, a sentence in the chat interface, a caption under illustrations. Write the decision down so it is applied consistently.

Name who the editor is

The marking duty falls away for text a human genuinely stands behind. That requires an actual name and an actual routine.

Voluntary, but useful

The Commission's code of practice

The transparency code has been assessed as adequate by the Commission and the AI Board. About 190 organisations had signed by the end of July, and the live list continues to grow. Signing remains voluntary, but offers a more predictable route to demonstrating compliance.

See the code and signatories
Checklist
  • Does the chatbot say it is AI?
  • Are AI-generated images marked?
  • Does published AI text have an owner?
  • Is the practice the same everywhere?
  • Do the people publishing know the rule?

On penalties

Breaches of the transparency duties can carry administrative fines. The ceiling is reported as 15 million euro or 3 percent of total worldwide annual turnover, whichever is higher, with proportionality for smaller organisations. National market surveillance authorities enforce this. Check the current text of Article 99 before relying on any figure internally.

Sources