Labelling AI-generated content: the European icons and the legal requirements

As of 2 August 2026, the transparency obligations of Article 50 of the European Artificial Intelligence Regulation apply. To facilitate compliance, the European Commission provides three icons that may be used to indicate that content has been generated or modified by AI.

The icons are voluntary for those who have not committed to using them via the Code of Practice on the marking and labelling of AI-generated content. The obligation to inform, when one of the scenarios provided for by the Regulation arises, is not. Nor is it sufficient to place any symbol anywhere: the information must be clear, recognisable, and accessible, and must be provided no later than the person's first exposure to the content.

The distinction is significant. The icon is a communication instrument. The obligation derives from the law.

To apply these rules to a publication, social network, video, or exhibition, I have prepared the guide How to label AI-generated or modified content, which includes a decision-making procedure, labelling formulas, resolved cases, and a final checklist.

The three icons

The set distinguishes between three situations:

• Basic icon: indicates AI intervention in a general manner. It may be used alone or accompanied by a more precise explanation.
• Content entirely generated by AI: identifies content created solely through AI, without human creative elements or human editorial control beyond the instructions provided to the system.
• Content partially modified with AI: identifies pre-existing human material that has been altered using AI until it becomes content subject to the information obligation.

The Commission provides the icons in SVG and PNG formats, in black, white, and versions with 50% transparency. They may be used freely without attribution. Design testing indicated superior results when the symbol was accompanied by a brief textual label, for example, "modified".

This result should guide its application. An icon without an explanation may become merely another technical sign that the public learns to ignore. "Image modified with AI" informs more than an isolated pictogram; "voice generated with AI" informs more than the generic expression "made with AI".

The Commission also recommends accessibility measures: visible size, plain language, alternative text or ARIA labels where possible, sufficient exposure time, and accessible navigation if a second layer with additional information exists.

Using the icons does not in itself demonstrate compliance with the Regulation. They are not a seal of approval from the European Union, nor do they certify that the content is lawful, accurate, or ethical. They serve to communicate AI intervention; legal responsibility remains with the party required to inform.

Two distinct obligations

Article 50 distributes transparency obligations among distinct actors.

Providers of AI systems, including providers of general-purpose models where applicable, must ensure that synthetic audio, image, video, or text outputs are marked in a machine-readable format and are detectable as artificially generated or manipulated. The regulation requires that solutions be effective, interoperable, robust, and reliable to the extent technically feasible, taking into account the characteristics of the content, costs, and the state of the art. This is an obligation incorporated into the system itself or its outputs.

Deployers, that is, those who professionally use a system under their authority, must inform in a visible manner when they publish or disseminate certain deepfakes or specific texts generated or manipulated with AI concerning matters of public interest. Strictly personal and non-professional uses fall outside the Regulation's definition of a deployer.

The icons are primarily directed at this second layer. They do not replace the technical marking that is the responsibility of the provider. An image may carry metadata or a machine-detectable signal and, simultaneously, require a visible warning for the public. It may also lose part of its technical information when downloaded, captured, or circulated on another platform; therefore, human communication remains necessary.

Content that must be indicated

Article 50.4 contemplates two main scenarios.

Deepfakes

The Regulation defines a deepfake as image, audio, or video content generated or manipulated through AI that resembles real persons, objects, places, entities, or events and may appear falsely authentic or truthful.

The definition does not extend to all synthetic imagery. A manifestly fantastical illustration that does not purport to be a record of reality may fall outside its scope. Conversely, an artificial photograph of a demonstration that never occurred, a voice attributing new words to a real person, or a video presenting a non-existent event as authentic may fall into this category.

Resemblance to objects, places, entities and events is as relevant as resemblance to persons. Reducing deepfakes to face-swapping ignores a substantial part of the problem: destroyed buildings that remain standing, non-existent crowds, altered heritage works or historical scenes fabricated with a documentary appearance.

Texts intended to inform on matters of public interest

The obligation also extends to text generated or manipulated by AI that is published to inform the public on matters of general interest.

There is an important exception: there is no specific obligation to inform under this section when the text has undergone human review or editorial control and a natural or legal person assumes editorial responsibility for its publication. Both conditions must be met. Clicking 'publish' or performing a superficial correction does not in itself demonstrate genuine editorial control.

The exception does not render the use of AI irrelevant. It means that Article 50.4 does not mandate the labelling of such text when review and editorial responsibility exist. An internal policy, an academic code, a professional standard or a transparency decision may require or advise a broader declaration. In the guide How to declare the use of AI in academic work I explain what information allows for a useful reconstruction of the tool's role.

Not every use of AI triggers this obligation

One should avoid the opposite error: legally labelling as AI-generated content any work in which an automatic tool has intervened.

Article 50 does not mandate the labelling of every spelling correction, assisted translation, preliminary search, or technical adjustment. Nor is every generated image a deepfake. The visibility obligation is contingent upon the type of content, its appearance, its purpose, and the manner in which it is presented to the public.

This does not preclude adopting an editorial policy that is more transparent than the legal minimum. In my declaration on the use of artificial intelligence I explain where I employ these tools and what responsibility I maintain for the result. The header images of this notebook are generated using AI and I declare this in a general manner. That decision does not mean that each one is a deepfake nor that Article 50 requires an icon to be placed on it in every case.

Distinguishing between voluntariness and obligation avoids two errors. The first is presenting a general ethical declaration as legal compliance. The second is asserting that, if the law does not require a label, no transparency responsibility exists.

Artistic works

The Regulation contemplates a specific form of compliance when a deepfake forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme. Information may be provided in an appropriate manner that does not hinder the exhibition or enjoyment of the work.

This is not a general exemption for art. The obligation is adapted in its form. A caption, gallery text, opening or closing credits, a programme, a catalogue entry or an explanation accessible via the icon itself may be reasonable solutions depending on the context.

The decisive word is 'evidently'. If a piece uses a documentary appearance, circulates outside the exhibition space or may be confused with an authentic record, hiding the information in a remote entry may prove insufficient. In an exhibition, moreover, one must consider subsequent circulation: gallery photographs, fragments on social media, press releases and downloadable materials may separate the content from the caption that provided its context.

The solution does not consist of covering the work with warnings. It consists of designing a transparency that travels with it when it leaves the framework that made its fictional character evident. In An image without provenance I analysed the inverse problem: a cultural institution published generated images without clearly explaining their origin. The issue was not to prohibit the tool, but to apply to institutional communication a standard of provenance consistent with the authority of the museum.

How to apply it in a cultural institution

An institution should not begin by asking which icon to place, but rather what kind of content it has before it and what responsibility it assumes by publishing it.

A minimum procedure may follow this sequence:

1. Identify the content. Determine whether it is image, audio, video or text, and whether it is fully generated or partially modified.
2. Check if it may appear authentic. In visual or audio material, assess whether it resembles a real person, object, place, entity or event and may induce a belief that it is true.
3. Specify the purpose. In texts, determine whether it is published to inform on a matter of public interest.
4. Document human intervention. Record who reviewed the content, what they verified and who assumes editorial responsibility.
5. Choose the location of the information. Place it at the latest at the first public exhibition, except for the adaptation provided for evidently artistic or analogous works.
6. Maintain it when redistributing the content. Prevent the warning from disappearing when downloading, sharing, cropping or reusing the material.
7. Ensure accessibility. Incorporate comprehensible text, alternative descriptions, and a navigable secondary layer when further detail is required.
8. Maintain an internal record. Archive tools, dates, objectives, transformations, sources, and editorial decisions.

The guide Artificial Intelligence in Exhibitions and Institutions extends this analysis to rights, procurement, data protection, conservation, and internal policy. Labelling represents only one component of institutional AI governance.

What a label does not resolve

A label confirms that content is synthetic or has been manipulated. It does not determine the legitimacy of its production.

It does not, in itself, resolve:

• whether an individual's voice, likeness, or identity was utilised with authorisation;
• whether training material or pre-existing work was eligible for use;
• whether a transformation respects copyright, performer rights, or moral rights;
• whether the processing of personal data has a valid legal basis;
• whether the content defames, discriminates against, deceives, or harms any party;
• whether an institution has verified the data it publishes;
• whether the work respects the conditions imposed by a community or an archive.

The AI Regulation itself applies without prejudice to legislation concerning data protection, consumer rights, fundamental rights, and other liabilities. A correctly identified deepfake may still infringe upon rights. Similarly, an unlabelled image does not automatically infringe Article 50: it must first fall within one of its prescribed categories.

Transparency is a necessary condition in the prescribed cases. To conflate it with authorisation would be to mistake the notice for permission.

Implications for artists and editors

For those who create or publish, the primary shift is not the addition of a mark at the end of the process. It necessitates considering from the outset how the content will circulate and how it will be interpreted by those who encounter it outside its original context.

An artist may achieve compliance through credits or gallery information without disrupting the work, yet must ensure that an isolated fragment retains the necessary notice. A media outlet publishing information of public interest may be exempt from the obligation to label the text if human oversight and editorial responsibility exist, provided they can demonstrate that such oversight was not merely nominal. A cultural institution may choose to provide information even when not obliged, yet must clearly distinguish whether this is a voluntary practice or a legal requirement.

The Commission has established a common visual language. Its utility depends upon it not becoming a mere alibi. Effective labelling entails informing at the appropriate moment, with sufficient precision, and without obscuring the human decisions that enabled the content behind the icon.

About the open conversation

This text distinguishes between the content of Article 50, the Commission's guidance, and my interpretation regarding its application in art and culture. It does not constitute legal advice. The classification of content depends upon its characteristics, the context in which it is disseminated, and the position of the individual utilising the system.

If you work with AI-generated content, communications, or cultural institutions and wish to clarify a point or contribute an applied case study, you may write to me.

The author's relationship with the cases

There is no economic, professional, or collection-based relationship with the individuals, companies, and institutions cited.

Sources

• European Parliament and Council of the European Union, Regulation (EU) 2024/1689, of 13 June 2024, in particular Articles 3.13, 3.60, 50, and 113. Official Journal of the European Union, 12 July 2024. The consolidated version of 27 July 2026 has also been consulted.
• European Commission, "EU Icons for labelling AI-generated content", updated 13 October 2026. Description of the three icons, use cases, variants, accessibility, placement, and licensing.
• European Commission, "Code of Practice on Transparency of AI-generated Content". Code of practice to support compliance with the marking and labelling obligations of Article 50.