Contents8
Should you state that a text was written by AI, at the bottom of an article, a product page or a customer reply? Some see a duty of transparency, others an irrelevant detail of how it was made. European law does not fully side with either: it requires the disclosure in certain specific cases, and waives it when a human has reviewed the text and takes responsibility for it.
For a company that publishes with the help of AI, the question is very practical: which content to flag, and how to prove that a person has reviewed the rest.
In short
The transparency obligations of Article 50 have applied since 2 August 2026; technical marking is the responsibility of tool providers, not of the companies that use them.
A product page or a commercial email does not generally fall under mandatory disclosure, but remains subject to the rules against misleading practices.
For Google, appropriate use of AI is not against its guidelines, and a disclosure is useful when readers might wonder how the content was made.
For a company, what matters is being able to show who reviewed the text, when, and who signs it off.
Position 1: the reader has a right to know
The transparency argument rests on trust. A reader does not read a first-hand account, a signed analysis and a text mass-produced by a machine in the same way. If they discover afterwards that an article they thought was written by an expert was generated, they feel deceived, and from then on they doubt everything the same source publishes.
Supporters of this position add a substantive argument: a language model can state a mistake with complete confidence. Knowing that a text is generated invites readers to read it with more caution, just as we read an anonymous source differently. We come back to this risk in our analysis of AI and disinformation.
Position 2: a text is worth what it says, not how it was made
The other camp points out that nobody declares having used a spell checker, translation software or the help of a copywriter. Between “AI wrote the whole text” and “AI rephrased two sentences”, the line is blurred, and a systematic disclosure would end up appearing everywhere, and so stop meaning anything.
This camp also stresses a perverse effect: a “generated by AI” label can get a sound, accurate text rejected, while a false text written by hand will pass without a label. Accuracy does not depend on the tool, but on the work of checking.
What the European regulation says
Article 50 of the EU AI Act is precisely about these transparency obligations. Three points to remember, starting with deepfakes.
Definition
Deepfake
An image, audio or video generated or manipulated by AI that resembles existing people, places or events and could wrongly pass for authentic.
Deepfakes must always be flagged. Deployers, meaning the professionals who use the system that produces them, must disclose that the content has been generated or manipulated by AI. For a clearly artistic, satirical or fictional work, the obligation is limited to flagging the existence of such content, in a way that does not ruin the work.
Texts of public interest, unless a human has reviewed them. Deployers of a system that generates or manipulates text “published with the purpose of informing the public on matters of public interest” must disclose that the text has been generated or manipulated by AI. But the obligation falls away “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 of the content”.
Technical marking, on the providers’ side. Providers of systems that generate text, images, audio or video must mark their outputs in a machine-readable format, so that they can be identified as generated. This obligation does not fall on the company that uses the tool, but on the vendor of the model. Nor does it apply when the tool merely provides standard editing assistance, or does not substantially alter the text it is given or its meaning: here the regulation partly agrees with the argument of position 2.
The timetable comes down to three dates:
- July 2026
Regulation (EU) 2026/1744 amends the AI Act.
- 2 August 2026
The transparency obligations apply (Article 113).
- 2 December 2026
Providers whose system was already on the market before 2 August 2026 must have put the technical marking in place.
Two remarks on scope. The mandatory disclosure for texts only covers those that inform the public on matters of public interest, not every product page or commercial email. And the European rule does not replace other obligations, such as those of consumer law on misleading practices, or the rules specific to customer reviews.
What Google says
On the SEO side, Google gave its position in a post of 8 February 2023 on its Search Central blog. Appropriate use of AI is not against its guidelines; what is against them is generating content primarily to manipulate rankings in search results. Google considers a disclosure useful for content where a reader might wonder how it was made, and advises adding one when it would be reasonably expected. It also recommends an accurate byline where a reader would wonder who wrote the text, but considers that presenting AI as the author is probably not the best way to say that it took part in the writing.
What we think
We find the European criterion smarter than the two starting positions. It does not ask “who typed the words?” It asks “who answers for it?”
A text that is generated, then reviewed, and for which an identified person takes responsibility, is no longer a machine’s text: it is an author’s text, written with a tool.
A text published without anyone having read it, on the other hand, deserves to be flagged, and probably should not be published at all.
On this blog, a generated article starts as a draft, and it is the human review that takes it out of draft to publish it. Because our site is static, publishing amounts to merging a change: we explain how this works in our article on the reasons for moving to a static site with Astro.
One nuance all the same: when AI did most of the work, an honest disclosure costs very little, even where there is no legal obligation, and it spares a reader from finding out later and feeling deceived. That is also the thrust of Google’s recommendation, which advises a disclosure where it would be reasonably expected.
What would change our mind: solid studies showing that readers trust a reviewed, signed text less when it mentions the help of AI, to the point of making the disclosure counterproductive.
What this changes in practice for a company
It comes down to four steps:
- Classify your content.
Distinguish what informs the public on matters of public interest (news, health, finance, analysis of social issues), what is commercial (product pages, service pages) and what is private (emails, customer replies). The legal obligation only covers the first group, and only without human review.
- Formalise the review.
If you are relying on the review exception, you must be able to prove it: who reviewed, when, and who takes editorial responsibility. A publishing chain where nothing goes out without approval, like that of a static site managed in a repository, keeps a record of it with no extra effort.
- Flag deepfakes.
A generated visual that has a real person say or do what they never said or did, and that could pass for authentic, is a deepfake. Unlike texts, it must be flagged whatever its subject, even after review.
- Write a short rule for the team.
A one-page rule is enough: which tools, for which content, who reviews, when you add a disclosure. It can also be part of the measures that Article 4 of the same regulation asks you to take to support AI literacy among staff, which we detail in our guide to what an SME that uses AI must do under the AI Act and the GDPR.
Where to start
The law does not ask you to label everything AI has touched. It asks you to flag deepfakes, and for a person to answer for the texts that inform the public. So start by classifying your content, then make the review visible: who reviewed, on what date, and who signs off. If your current site keeps no record of what was published or when, a migration to a static site, where each publication goes through a preview and each version can be restored gives you that record, on top of a faster site.
Frequently asked questions
Do I have to flag a product page written by AI?
Generally not under the AI regulation: its Article 50 only covers texts published to inform the public on matters of public interest, and a product page is not in principle one of them. It remains subject to the rules against misleading commercial practices: a feature invented by the tool binds the company as if it had written it itself.
Is a simple review enough to avoid the disclosure?
The regulation sets two conditions: a human review or editorial control, and a natural or legal person who takes editorial responsibility for the publication. A quick glance with no identified person in charge does not meet the second. To be able to prove it, keep a record of who reviewed the text, on what date, and who signs off the publication.
Who has to mark generated content technically?
The provider of the system, meaning the vendor of the tool that generates the text, image, audio or video, in a machine-readable format. The obligation has applied since 2 August 2026, with a grace period until 2 December 2026 for systems already on the market. The company that uses the tool, for its part, flags deepfakes and texts of public interest that have not been reviewed.
Can you present AI as the author of an article?
Google does not forbid it, but it considers that it is probably not the best way to tell readers that AI took part in the writing. It recommends instead an accurate byline and a disclosure where one is reasonably expected. The European regulation, for its part, says nothing about bylines: for texts of public interest, it asks that a person take editorial responsibility.
Sources
- EU AI Act (new tab)eur-lex.europa.eu
- Regulation (EU) 2026/1744 (new tab)eur-lex.europa.eu
- Search Central blog (new tab)developers.google.com
Jim · founder of Brain Plus AI
Also in websites, migration and astro
Stepping back7 min
AI and disinformation: lying machine or remedy?
One study shows that AI writes more convincing false information than humans do, another that it reduces conspiracy beliefs, but its journal has expressed a reservation. What we can take from them, and what the law requires of those who publish.
Practical guide10 min
AI Act and GDPR: what an SME using AI must do
An SME that uses ChatGPT or a CRM with built-in AI already has obligations. This guide takes stock in October 2026: your role under the AI Act, the timetable revised in July 2026, what the GDPR requires and the steps to get compliant.
Stepping back7 min
AI and copyright: should authors be paid?
Plundering works, or learning like a reader? European law and a first German ruling are starting to separate learning from memorising, and a content publisher can already choose to opt out of training.