Texts written with AI: when do I have to say so?
Updated: 13 October 2026
When you publish text generated or rewritten by AI that informs people about matters of public interest, such as health, law, finance or safety. The duty falls away if a competent human has checked the content and the site names who is responsible for publishing it. Product descriptions and ordinary advertising copy are generally outside the duty.
What the rules say and what status they have
The EU Artificial Intelligence Act has a transparency provision that has applied since 2 August 2026. Anyone who, in a professional capacity, publishes text generated or altered by AI in order to inform the public on matters of public interest must disclose it.
The detail comes from the European Commission guidelines and the Code of Practice. The guidelines are not binding, and their formal adoption in all languages could not be confirmed by 13 October 2026. The Code is voluntary. Even so, they are the best description available today of how to read the provision.
The date of publication is what counts. Text published before 2 August 2026 needs no notice. Text generated earlier but published on or after that day does.
Three conditions and the list of topics
The duty arises when three conditions are met together. The text is published, meaning it is open to a wide circle of unrelated readers, including behind a paywall. It is meant to inform, meaning it conveys knowledge, opinions or facts. It concerns a matter of public interest.
The list of topics is broad and is not limited to news sites. The guidelines name politics, public administration and services, justice, fundamental rights, public security, public health, environmental protection and consumer safety. Economic, financial, scientific and cultural developments that are a subject of public debate count as well.
An example from the guidelines: an article on a lifestyle website comparing the effects of diets on a disease. A guide on a company blog about tax, data protection rules, medical conditions or energy prices is covered in the same way, even if its main purpose is search visibility.
What is outside the duty
Product descriptions and advertising copy are expressly outside the duty. The exception: when they make claims about health, consumer safety or sustainability. So an ordinary shampoo description is outside the provision. A description that promises a health effect may fall within it.
A post on how to use a product, company news, recipes and style tips are outside the duty too. Novels, private correspondence and advice sent to a single client do not count.
Minor help from AI is not altering the text. Fixing typos and grammar and lightly polishing the style is standard editing.
When no notice is needed
No notice is required when two conditions are met together. First, a person who knows the subject has checked the substance before publication and was able to change or reject it. Fact-checking is the minimum. Second, the site names, in a place that is easy to find, the person or company responsible for the publication, with contact details.
These are not enough: spell-checking and grammar correction alone, having another AI review the text, a cursory sign-off, or simply having an editorial policy. If AI makes a substantive change to the text after sign-off, the exception is lost.
It is worth writing down briefly who checks texts and what they check. That is your evidence if anyone asks.
Where and how to place the notice
The notice has to be visible from the reader's first contact with the text. The Code of Practice puts it above the text or at the top, near the headline. A note at the end of the article is a weak solution, because it is easy to miss.
A general line in the footer or the terms, such as "we use AI on this site", is not a notice. Each text is labelled separately. You may label only the part that AI wrote or altered.
A plain statement such as "Text generated by AI" will do, or the optional EU icon in its variant for generated or for modified content. The notice must also be readable by screen readers.
Text detectors and enforcement
Programs that claim to detect AI in text are unreliable. In a 2023 study, seven detectors flagged as AI-written an average of 61 per cent of essays written by people whose first language was not English. A 2025 study covering Polish, Czech and Hungarian, among others, found detectors built around English almost unusable for other languages. The maker of one such tool itself says the result is not proof.
A detector score also says nothing about what matters in the provision: the topic, and whether a human checked the content. The decision is yours.
In Slovenia and Hungary an authority can already fine a missing notice today. In Poland the authority will be able to do so from 28 October 2026. The duty itself applies everywhere now.
This describes the rules. It is not advice on your particular case. Legal basis: Article 50(4) and (5) of Regulation (EU) 2024/1689.
Key points
- The notice covers AI texts on matters of public interest, including health, law, finance, safety, the environment and science.
- Product descriptions and adverts are outside the duty unless they make health, safety or sustainability claims.
- No notice is needed when a competent human has checked the facts and the site names who is responsible for publication.
- The notice goes at the top of the text, near the headline. A general line about AI in the footer is not enough.
- A text detector score is not proof. The decision is yours.
Sources
- Regulation (EU) 2024/1689, the Artificial Intelligence Act
- European Commission: guidelines on Article 50, annex to document C(2026) 5054
- Code of Practice on Transparency of AI-Generated Content
- Liang et al. (2023): AI text detectors and non-native English writers (arXiv 2304.02819)
- Polish Act of 3 July 2026 on artificial intelligence systems (Journal of Laws 2026, item 1003)
This guide is for information only and is not legal advice.