Labelling AI-generated content: what the EU AI Act requires
AI and Search

Labelling AI-generated content: what the EU AI Act requires

2026-03-30 13 min read

Article 50 of the AI Act has applied since 2 August 2026. Who must label what, what is exempt, the penalties, and what a company publishing AI-assisted content should do.

Since 2 August 2026, Article 50 of the EU AI Act has applied, setting out when AI-generated content must be marked or disclosed. It covers four situations: AI systems that interact with people, machine-readable marking of synthetic content, deep fakes, and AI-generated text published to inform the public on matters of public interest. Ordinary AI-assisted web copy such as product descriptions and advertising is normally outside the rules, and text that has been reviewed by a human and has a party holding editorial responsibility is expressly exempt.

In brief

  • Article 50 applies from 2 August 2026. The date was not moved by the EU's Digital Omnibus on AI, which entered into force on 27 July 2026.
  • Providers of AI systems must build in disclosure and machine-readable marking. Those who use the systems professionally must label deep fakes and certain text.
  • There is no general duty to label every text written with AI assistance.
  • Infringements of Article 50 can lead to fines of up to EUR 15 million or 3 percent of worldwide annual turnover.
  • Google's guidance on AI content is a different thing from the law. It is covered separately at the end.

Note: this article is general information on the position as of November 2026 and is not legal advice. The assessment in an individual case depends on the circumstances. Consult a lawyer if you are unsure.

The position in November 2026: what has been adopted

Everything below is based on adopted legal acts and documents published by the European Commission, not on proposals.

DateWhat happened or applies
1 August 2024The AI Act, Regulation (EU) 2024/1689, entered into force.
10 June 2026The final Code of Practice on Transparency of AI-generated Content was published.
27 July 2026Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force.
2 August 2026Article 50 started to apply.
5 November 2026Date of the Commission's guidelines on Article 50, C(2026) 7702 final.
2 December 2026Deadline for providers whose systems were placed on the market before 2 August 2026 to comply with the machine-readable marking duty in Article 50(2).

The base text is the consolidated version of the AI Act on EUR-Lex. The application dates are in Article 113: the Regulation applies from 2 August 2026, with certain exceptions that do not concern Article 50.

The Omnibus Regulation postponed the rules for high-risk AI systems, but not the transparency rules. For Article 50 it changed two things. Paragraph 7 on codes of practice was rewritten. And a new Article 111(4) gives providers of systems already on the market until 2 December 2026 to comply with Article 50(2). That transitional period covers only the providers' marking duty. The duties of those who publish content have applied in full since 2 August.

The Omnibus also added new prohibitions to Article 5 that apply from 2 December 2026, including one on AI systems that generate realistic sexually explicit images, video or audio of identifiable people without their consent. That is a ban, not a labelling rule.

Provider or deployer: which role do you have?

The Regulation splits responsibility between two roles.

  • A provider, under Article 3(3), develops an AI system or has one developed and places it on the market or puts it into service under its own name or trademark. Example: the company behind a chatbot or an image generator.
  • A deployer, under Article 3(4), is a person or body "using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity". Example: a business that writes web copy or creates images with an AI tool.

A company that publishes AI-assisted content is almost always a deployer. According to the Commission's guidelines, individual employees are not treated as separate deployers, and the company remains the deployer even when freelancers or contractors operate the tools on its behalf and under its control. A company that builds its own AI system and puts it into service under its own name can be provider and deployer at once.

The four obligations in Article 50

ParagraphWhoObligation
50(1)ProviderAI systems that interact directly with people must be designed so that the person is informed they are interacting with an AI system, unless this is obvious.
50(2)ProviderSynthetic audio, image, video and text output must be marked in a machine-readable format and be detectable as artificially generated or manipulated.
50(3)DeployerAnyone using an emotion recognition or biometric categorisation system must inform the people exposed to it.
50(4)DeployerDeep fakes must be disclosed. So must AI-generated or manipulated text published to inform the public on matters of public interest, with exceptions.

Under Article 50(5) the information must be given "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure" and must conform to applicable accessibility requirements.

Chatbots and other interactive systems

The duty sits with the provider, who must build the disclosure in. The Commission says the exception for what is "obvious" should be interpreted restrictively. The guidelines also state that some techniques are insufficient when used alone, such as a disclosure only in the terms and conditions, or a general statement that services on the website use AI. If you run a chatbot on your site, check that it identifies itself as AI at the start of the conversation, whether you bought it or built it.

Machine-readable marking

This is the tool vendor's responsibility, not yours as a user. The marking duty does not apply where the system performs "an assistive function for standard editing" or does not substantially alter the input the user provided. Recital 133 mentions watermarks, metadata identifications and cryptographic methods as examples of techniques. The Commission's guidelines encourage those who distribute content to preserve the marking, so do not strip metadata without a reason.

Deep fakes

Article 3(60) defines a deep fake as "AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful". So the definition is not limited to faces and voices. The guidelines give examples that matter for marketing:

  • Deep fake: an AI-generated video in which a celebrity influencer appears in an advertisement, and an AI-generated product image that can mislead about what the product actually looks like or does.
  • Not a deep fake: a real product shown against an AI-generated background, as long as the image does not mislead about the product, and evidently unreal scenes such as talking mice in a commercial.

Colour correction, noise reduction and similar standard editing do not normally turn content into a deep fake, according to the guidelines. For evidently artistic, satirical or fictional works it is enough to disclose the AI element in a way that does not hamper enjoyment of the work. The disclosure has to be perceivable by a person without special tools. The provider's machine-readable marking does not replace it.

The text rule in Article 50(4), word for word

This is the provision that settles the question for most people publishing on the web. The second subparagraph of Article 50(4) reads:

Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated. This obligation shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences or 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.

When is a text covered?

The Commission reads three conditions into the rule, all of which must be met, according to its questions and answers on Article 50 and the guidelines:

  1. The text is published. It is accessible to an indeterminate and fairly large number of readers, including behind a paywall. Internal texts and private correspondence do not count.
  2. It aims to inform the public. It communicates knowledge, opinions or facts.
  3. It concerns matters of public interest. The guidelines mention politics and democratic processes, public administration, the administration of justice, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, political, scientific or cultural developments that may be a subject of public debate.

As examples of text within the rule, the guidelines mention an AI-generated summary of a newspaper article about a town council decision, AI-manipulated parts of a lifestyle article comparing the effects of diets on a disease, and AI-manipulated corporate reports containing investor information. As an example of text outside the rule they mention AI-manipulated text that is part of a company's advertisement or product descriptions, with the reservation that this does not include claims relating to, for example, health, consumer safety or sustainability.

The exception: human review and editorial responsibility

Even a text that is covered does not need a label if two conditions are met together.

  • Human review or editorial control. The guidelines describe this as a deliberate examination of the substance of the content by one or more people with relevant knowledge, with fact-checking as a minimum requirement. Spell-checking, purely formal or automated review and cursory sign-off are not enough. Nor is the mere existence of an editorial policy.
  • Editorial responsibility. A natural or legal person must hold ultimate responsibility for the publication. The guidelines say the identity and contact details of that person or function should be publicly available in an easily findable place, for example in the website's legal information.

One detail worth knowing: if AI is used to modify, supplement or reformulate the text after editorial sign-off, the exception falls away according to the guidelines, and the text must be treated as AI-generated.

The guidelines are not binding. Only the Court of Justice of the European Union can give an authoritative interpretation of the Regulation.

What does not need a label

Based on the wording of Article 50 and the Commission's examples, the following can be said.

  • Article 50 contains no general duty for publishers to label all AI-assisted text. The duties of deployers are in paragraphs 3 and 4.
  • Advertising copy and product descriptions are not text published to inform the public on matters of public interest, as long as they do not contain claims about, for example, health, consumer safety or sustainability.
  • AI-assisted text that a knowledgeable person has reviewed in substance, and for which an identified person or organisation is responsible, is exempt even when the subject is of public interest.
  • Internal documents and private correspondence are not published.
  • Use by private individuals outside professional activity is not covered by the deployer obligations.

Two caveats. A company blog is not advertising simply because it sits on a company website: if you write about health, personal finance, law or the environment for a broad audience, the text may concern matters of public interest, and then the review and the responsibility are what decide the matter. And Article 50(6) makes clear that the rules are without prejudice to other transparency obligations in Union or national law. Misleading advertising is unlawful however the text was produced.

What the disclosure should look like

The Regulation prescribes no set wording. The requirement is that the disclosure is clear, distinguishable and present no later than the reader's first exposure to the content. The EU has produced a set of icons that anyone may use freely. They are optional, and the Commission points out that using an icon does not by itself establish compliance. A simple text label at the top of the content could look like this:

<p><strong>AI-generated text.</strong> This summary was produced with an
AI system and has not been reviewed by our editors.</p>

The voluntary Code of Practice on Transparency of AI-generated Content sets out more detailed measures for both providers and deployers. On 8 July 2026 the Commission concluded that the code adequately covers the obligations in Article 50(2), (4) and (5). Those who do not sign it have to demonstrate compliance by other means.

According to the guidelines, content that was generated and published before 2 August 2026 does not need to be labelled retroactively. Text generated earlier but published on or after that date is covered.

Penalties and supervision

Under Article 99(4)(g), non-compliance with the transparency obligations in Article 50 is subject to administrative fines of up to EUR 15,000,000 or, for an undertaking, up to 3 percent of its total worldwide annual turnover for the preceding financial year, whichever is higher. For small and medium-sized enterprises, Article 99(6) makes the lower of the two the cap. Member States lay down the detailed rules on penalties under Article 99(1), and enforcement is mainly carried out by national market surveillance authorities.

In Sweden, the government in a decision of 4 June 2026 assigned the Swedish Post and Telecom Authority (PTS) as market surveillance authority for the transparency requirements in Article 50(1), (2) and (4). The Swedish Authority for Privacy Protection (IMY) is responsible for Article 50(3). PTS is also the single point of contact for the AI Act. According to the decision the assignment runs until 31 December 2026. A government inquiry, SOU 2025:101, has proposed a new act and ordinance with supplementary provisions. Check with the Swedish government or PTS for the arrangement that applies from 2027. Companies in other Member States should look up their own national authority.

What a company publishing AI-assisted content should do

  1. Map your AI use. List where AI is used for text, images, video, audio and customer contact, and note whether you are deployer or provider in each case.
  2. Sort your texts. Separate advertising and product descriptions from content that informs the public on matters of public interest. Take particular care with health, safety, finance and sustainability.
  3. Make the review real and traceable. A person with subject knowledge checks facts and sources and approves the text. Record who and when. Do not let AI rewrite the text after approval.
  4. Publish who is responsible. State the person or function holding editorial responsibility, with contact details, on a page that is easy to find. Named authors and a clear publisher also strengthen credibility, see the guide to E-E-A-T in the AI era.
  5. Label what has to be labelled. AI-generated or manipulated images, video and audio that realistically depict real people, places, products or events, and AI text on matters of public interest that is published without human review.
  6. Review product images. An AI image that makes the product look different from or better than it really is can be a deep fake. This matters most for online retail, see GEO for e-commerce.
  7. Check your chatbot. It should say it is AI when the conversation starts.
  8. Write down an AI policy. Which tools are used, who reviews and how labelling is done. It is also your evidence if an authority asks.

Google's guidance is not a legal requirement

Do not confuse the AI Act with what Google says about AI content in search results. They are separate matters. Google's guidance on generative AI content says generative AI can be useful, but that generating many pages without adding value for users may violate its spam policy on scaled content abuse. Google calls for manual fact-checking and review of all AI-generated content before publishing, and suggests that you consider adding information on how the content was created. That is a recommendation, not a labelling requirement for ordinary web pages.

For online retailers, Google Merchant Center does have its own rules, according to the same page: AI-generated images must carry IPTC metadata indicating this, and AI-generated titles and descriptions must be specified separately and labelled.

The conclusion is the same from both directions. What counts is that a knowledgeable person has reviewed the content and stands behind it. How to write so that the text is also useful as a source is covered in writing content AI wants to cite, and the usual pitfalls of mass-produced AI content are in common GEO mistakes. If you want to see how a single page holds up technically and in terms of content, Visicite's free scan checks, among other things, whether there is an identifiable sender and named authors. It does not assess legal compliance.

Frequently asked questions

Do I have to label AI-generated content on my website in 2026?

Only in certain cases. Since 2 August 2026 you must label deep fakes, and AI-generated text that informs the public on matters of public interest unless the text has been reviewed by a human and someone holds editorial responsibility. Advertising and product descriptions are normally not covered.

What counts as a deep fake under the AI Act?

AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful. Text cannot be a deep fake within the meaning of the Regulation.

Is it enough for someone to read through the AI text before publishing?

Not if the read-through only concerns language and form. The Commission's guidelines require a review of substance including fact-checking, by someone with relevant knowledge, and a person or organisation that holds editorial responsibility.

How high are the fines for failing to label AI content?

Up to EUR 15 million or 3 percent of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises the lower of the two is the cap.

Does older content have to be labelled retroactively?

No. According to the Commission's guidelines, content generated and published before 2 August 2026 does not need to be labelled retroactively. The Commission nevertheless encourages it where possible.

Kristoffer Pihlfelt

Developer and founder, Pihlfelt & Vänner

Developer and founder of the Swedish web agency Pihlfelt & Vänner since 2014. Builds Visicite and writes the GEO guide.

About the author

Share:

Related Articles