EU AI Act labelling requirements: when to label AI-generated content and how to do it in Jahia

The AI Act, Regulation (EU) 2024/1689, is EU law. Since 2 August 2026, its Article 50 transparency obligations have applied to providers that place AI systems on the market in the EU and to deployers established or located in the EU. They have also applied to providers and deployers outside the EU whose AI output is used in the EU. From the same date, Article 50 has required organisations to disclose that a text published to inform the public on matters of public interest was AI-generated or manipulated, unless the text went through human review or editorial control and someone holds editorial responsibility for it. In Jahia, you put this into practice with four tools: the publication workflow, roles, content history and a disclosure field rendered by the template.

This article is for marketing and editorial teams that write with the help of AI. It summarises the official text and shows how to apply it in Jahia. For a specific case, your legal counsel has the final word.

What Article 50 says

Article 50 of the AI Act sets transparency obligations. Two of them concern the content you publish.

Technical marking, on the AI tool side. Paragraph 2 is addressed to providers of generative AI systems, that is the companies behind the tools that produce the text. Since 2 August 2026, their outputs must be "marked in a machine-readable format and detectable as artificially generated or manipulated", with a deadline of 2 December 2026 for systems already on the market. This marking is invisible to the reader. It does not always survive the journey: a plain-text copy and paste carries no metadata.

Visible disclosure, on the publisher side. Paragraph 4 is addressed to deployers, that is the organisation that uses the tool to produce what it publishes. Since the same date, for text "published with the purpose of informing the public on matters of public interest", deployers must "disclose that the text has been artificially generated or manipulated". This is the disclosure that editorial teams own.

The form of the disclosure. Since 2 August 2026 as well, paragraph 5 requires the information to be provided "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure" and it "shall conform to the applicable accessibility requirements". For text, the Commission guidelines of 20 July 2026 ask for a clear and perceivable notice that cannot easily be overlooked (points 132 and 142) and give as an example a notice "at the start of an AI-generated or manipulated text publication on matters of public interest" (box following point 143).

Who is concerned and since when

• Since when. Article 50 has applied since 2 August 2026. Generative AI systems placed on the market before that date have until 2 December 2026 to comply with the marking of paragraph 2: this is Article 111(4), added by Regulation (EU) 2026/1744.
• Who. A deployer is any person or organisation using an AI system under its authority, except in the course of a personal non-professional activity (Article 3, point 4). A company or a public body that publishes text written or edited with AI is therefore a deployer.
• Where. The AI Act is EU law. Under Article 2, the Article 50 obligations, applicable since 2 August 2026, cover in particular deployers established or located in the EU and those established outside the EU when the output produced by the AI system is used in the EU. It is not a worldwide rule.
• Which texts. Text published to inform the public on matters of public interest. Whether your news, blog posts or corporate pages fall into that category is a question to settle with your legal counsel.

The editorial control exception: what it takes

Paragraph 4, applicable since 2 August 2026, lifts the obligation "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". It also exempts uses authorised by law to detect, prevent, investigate or prosecute criminal offences, which does not concern ordinary publishing.

The two conditions are cumulative: a real review and editorial responsibility that someone holds. The regulation does not say what the review must involve. The Commission guidelines of 20 July 2026 do:

• "Fact-checking the accuracy of the content is a minimum requirement that should be part of that review" (point 134);
• neither spell-checking, grammatical correction nor a cursory editorial approval is enough (point 135);
• "Any substantive AI intervention occurring after the human review or editorial control process has taken place will therefore cause the exception to become void" (point 136);
• the identity and contact details of the person, or of the function, holding editorial responsibility should be made public in an easily findable location, for example in the legal information of the website (point 138).

In practice, the exception holds if you can show, for each text, who reviewed it, on what date and that AI did not change the substance after that review. Without that record, a text that falls within paragraph 4 must carry the disclosure, as it has since 2 August 2026.

In Jahia: how to set it up

Jahia provides features that help you apply these requirements and keep a record of them.

Human review before publication

In Jahia, a writer with the Editor role creates content and requests its publication: they cannot publish it themselves. A reviewer then validates or rejects the request. Depending on your organisation, the publication workflow can include additional steps.

The reviewer does not approve blindly. The review screen compares the live version with the version to be published and highlights what was added, changed or deleted. That is the place to check the facts.

Jahia's AI features go through the same workflows. Every AI operation triggers the validation workflows set up for that content type. If your workflows require human sign-off before publishing, AI cannot skip that step. Some low-risk operations, such as updating a meta description, can be configured to run automatically. Your team controls that configuration, as the page on Jahia's built-in AI explains.

Identified editorial responsibility

In Jahia, the right to publish is carried by roles. The documentation describes three of them. The Editor creates content and requests its publication. The Editor in chief can also publish directly and validate requests. The Reviewer validates or publishes but cannot create or edit content. A role is granted to users or groups, on a site or on part of a site.

The validation role sets who reviews and approves in Jahia; it does not confer editorial responsibility. That lies with whoever holds the ultimate legal responsibility over the publication, review included, whether an individual, an editorial board or the publishing company (point 138). What remains is to publish that person's identity or function, for example in the legal information of the site.

A clear notice, when it is due

A notice that the writer has to remember to type will eventually be missing. It is better to make it a field of the content.

In Jahia, a content type is defined by its properties. A mixin adds the same property to several content types and the view that renders the content displays that property where you choose. Your technical team can therefore add a field such as "Text generated or edited with AI" to your articles, which the view displays at the start of the text when it is ticked. The documentation describes this mechanism, not a ready-made field: it is a development to plan.

A field can be checked across every page of a site. A habit cannot.

For the form of the notice, the Commission provides EU icons for labelling AI-generated content, which can be used for text on matters of public interest. They are part of the Code of Practice on Transparency of AI-generated Content, which organisations sign voluntarily and which the Commission and the AI Board have confirmed as an adequate tool to demonstrate compliance with the AI Act transparency obligations. The icons are optional and do not establish compliance by themselves. Where a disclosure is due, it has been due since 2 August 2026, with or without an icon. If you adopt them, plan them in the same development as the disclosure field, with a text label and alt text. According to the Commission, the basic icon performed better in user testing when accompanied by a text label and it encourages making the icons readable by assistive technologies where possible.

A record of who did what

Jahia keeps three useful records:

• content history, in the advanced options of the editor: the date and type of each action, the user who performed it and the property concerned;
• published versions, which you can compare with the current version and restore;
• workflow comments, which cannot be edited or removed once added.

For a given text, these records help you find out who reviewed it, on what date and what changed afterwards. That is what you need to be able to show to rely on the exception.

An accessible notice

The guidelines state that "Article 50 AI Act does not impose distinct or additional accessibility requirements" (point 144). They give two directives as examples: Directive (EU) 2016/2102 on the websites and mobile applications of public sector bodies and Directive (EU) 2019/882 on products and services. Each organisation has to check whether its content is subject to them.

On the tool side, a notice rendered as text by the template can be read on screen and by a screen reader. A notice carried by an image alone, by colour alone or by an attribute that screen readers ignore will not hold up against any standard. Among the uses of its built-in AI, Jahia's product page also lists a WCAG accessibility check before publishing.

What to decide internally

1. Which texts fall within Article 50(4): to settle with your legal counsel.
2. Who holds editorial responsibility and where their identity is published.
3. What the review must include at a minimum, starting with fact-checking.
4. Which AI operations may run without human approval.
5. The wording and position of the notice, when it is due.
6. What to do with content already published. Under the Commission guidelines of 20 July 2026 (point 154), AI-generated texts on matters of public interest "published before 2 August 2026" do not need to be labelled retroactively, but "if texts that have been generated or manipulated before 2 August 2026 are published on or after that date, they need to be labelled." For the rest of the backlog, information that was not recorded at the time of writing cannot be rebuilt: that is a decision to make with your legal counsel.

What we do not claim

No platform makes an organisation compliant with the AI Act. Compliance depends on what you publish, who reviews it and what you can show. A tool can make proof possible, it does not produce it for you.