Do AI-Generated Texts Have to Be Labelled?
Transparency obligations under Art. 50 EU AI Act for AI-generated texts
From 2 August 2026, the transparency obligations of the AI Act, Art. 50 AI Act, apply.
For images, videos and audio, the labelling obligation for AI content is by now widely known. For texts, by contrast, uncertainty prevails.
AI-generated text content must also be labelled under the AI Act, insofar as it is relevant to the public and has not undergone final editorial review, Art. 50(4) subpara. 2 AI Act.
There is therefore no general obligation to label AI-generated texts. The obligation under Art. 50(4) subpara. 2 AI Act is narrow and only applies when two hurdles are cumulatively crossed: The text must inform the public on matters of public interest and it must not be subject to genuine final editorial review.
Note on the state of the law: This article takes into account the draft guidelines of the EU Commission on the implementation of the transparency obligations under Art. 50 AI Act published on 8 May 2026.
The principle: no general labelling obligation for texts
The obligation under Art. 50(4) subpara. 2 AI Act falls – as with the other content types – on the deployer, i.e. the user of an AI system.
The deployer must label both texts generated with AI and texts manipulated with AI:
- Generating a text with AI means creating a text by prompt – "Write an article about …".
- Manipulating a text with AI means altering an existing text using AI – for instance "Correct the text" or "Summarise".
Public relevance
The labelling obligation requires that the text is published in order to inform the public on matters of public interest. What falls under this is specified by the Commission's draft guidelines of 8 May 2026.
The Commission considers publicly relevant in particular:
*- an AI-generated summary of a human-written newspaper article about a recent decision of a city council;
- AI-manipulated parts of a published scientific paper comparing the effects of different diets on a particular disease in middle-aged women;
- an AI-generated post on the social media profile of a meteorological institute warning citizens of stormy weather and corresponding precautions.*
Not publicly relevant, by contrast, in the Commission's understanding are:
*- AI-generated fictional novels or poems of any genre;
- AI-manipulated text as part of corporate advertising – with the exception, however, of statements relating to health, consumer safety or sustainability;
- a news summary by a chatbot that is only available to the user who operated the chatbot.*
For practice, this is considerable relief: Most editorial and promotional corporate texts already fall out at this level.
Caution is, however, required with advertising texts touching on health or sustainability.
Example: A beverage manufacturer publishes an advertising text pointing out that its lemonade strengthens the immune system and is produced in Germany, and which addresses the ingredients and their effect in the body. Here, the threshold of public relevance could be crossed because health claims are concerned.
Final editorial review
Even if a text is publicly relevant, the labelling obligation may nevertheless not apply. Under Art. 50(4) subpara. 2 AI Act, it does not apply where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication.
Both conditions must be met cumulatively: a genuine review process and clearly assigned editorial responsibility.
What matters in practice is demonstrability. Anyone wishing to rely on the exception should document the review process: who reviewed and approved the text substantively, which substantial checks and changes were made, and who holds editorial responsibility. A traceable version history evidencing human intervention beyond mere formatting, and a documented approval step, are the obvious means here. Without such evidence, the exception can hardly be relied upon robustly in a dispute.
In practice, this is the decisive lever: Whoever establishes a defined, documented approval process for publicly relevant texts keeps the labelling obligation within limits and at the same time secures quality and accountability. We design this process in a legally sound way so that your evidence holds up in a dispute.
How must publicly relevant AI-generated texts be labelled?
If the text is publicly relevant and does not undergo final editorial review before publication, the deployer must label the use of AI. The modalities are governed by Art. 50(5) sentence 1 AI Act: The information must be provided to the persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. The public should be able to recognise that it is not dealing with purely human-created informational content.
For texts, this means: The notice belongs at the beginning, before the text – not at the end and not hidden in the legal notice.
In the lemonade example: If the website carries an AI-generated short version of the article about strengthening the immune system, produced automatically and without final editorial review, the notice must appear before the text and before reading – for instance: "This is an AI-generated summary of the following text."
What infringements entail
The labelling obligations of Art. 50 AI Act are not a mere recommendation. Infringements are subject to the fining regime of the AI Act: Under Art. 99 AI Act, fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher, may be imposed. The designation of the competent supervisory and enforcement authorities takes place nationally; the sanctionable threshold, however, is thereby set by Union law.
Conclusion and recommended course of action
For AI-generated texts: no blanket labelling, but a narrowly limited obligation. The assessment proceeds in two steps:
- Check public relevance. Does the text inform the public on a matter of public interest? Pure advertising and fictional content fall out – but beware of health, safety and sustainability claims.
- Ensure and document final editorial review. Where public relevance looms, the obligation can be avoided through genuine human control with assigned editorial responsibility – provided the process is demonstrable.
Only when both filters are crossed must labelling take place – clearly and prominently before the content.
The most effective lever is therefore organisational: a defined, documented approval process for publicly relevant texts. It not only keeps the labelling obligation within limits but also secures quality and accountability – and protects against fines under Art. 99 AI Act as well as against attacks under competition law.
You use AI to create or edit texts and want to know which of them are subject to labelling? We review your content along the two decisive filters, design a legally sound editorial control process and ensure that your evidence holds up in a dispute.
FAQ
1. Do I have to label every text created with ChatGPT or similar tools?
No. There is no general labelling obligation for AI texts. It only applies if the text informs the public on a matter of public interest and is not subject to genuine final editorial review.
2. Are advertising texts subject to labelling?
In principle, no – advertising is generally not information on matters of public interest. It may be different for statements relating to health, consumer safety or sustainability.
3. What does "final editorial review" mean?
More than merely adopting the AI output. What is required is a substantive review for accuracy, selection, weighting and revision – with a clearly assigned person holding editorial responsibility.
4. Is a notice in the legal notice sufficient?
No. The notice must appear clearly, unambiguously and before the text – not hidden at the end or in the legal notice.
5. What are the consequences of infringements?
Under Art. 99 AI Act, fines of up to EUR 15 million or 3% of worldwide annual turnover. In addition, there is the risk of cease-and-desist letters under competition law if the missing labelling is deemed a breach of a market conduct rule (Section 3a UWG).
