How to label content generated or modified with AI

Verified on 14 October 2026.

Since 2 August 2026, the transparency obligations of Article 50 of the European Artificial Intelligence Act have applied. They do not require the labelling of everything in which an AI has intervened. They affect, as far as this guide is concerned, two specific groups: certain image, audio or video content that constitutes deepfakes, and certain texts generated or manipulated with AI that are published to inform about matters of public interest.

This guide assists in determining whether there is an obligation to disclose, which European icon may be utilised, and how to document the decision. The complete legal and cultural analysis is available in Labelling AI-generated content: European icons and legal obligations.

The European Commission icons are voluntary, except for those who have committed to using them by adhering to the Code of Practice. The legal obligation to disclose, where applicable, is not voluntary. Furthermore, using an icon does not in itself demonstrate compliance with the Regulation.

Before you begin

This guide is intended for artists, media organisations, museums, archives, universities, foundations, publishers, production companies, and communications professionals who use AI systems in a professional capacity.

It does not replace the analysis of a specific case. The Regulation defines a 'deployer' as the natural or legal person who uses an AI system under their authority, except when doing so in a personal, non-professional activity. An institution remains the deployer when its employees use the system following its instructions and under its control.

Before labelling, retain a copy of the content, identify the tool used, and note which parts were generated or modified. Without such a record, it will be more difficult to justify the chosen approach at a later date.

Step 1. Identify the type of content

Classify the output before assessing the obligation.

• Image, audio, or video. Verify whether it represents or reproduces an existing person, object, place, entity, or event with a realistic appearance.
• Text. Verify whether it is published to inform the public about a matter of public interest.
• Artistic, creative, satirical, or fictional work. Determine whether its nature is evident to the recipient within the specific context.
• Ancillary use. Distinguish the generation or manipulation of content from simple spell-checking, searching, assisted translation, classification, or organisational tasks.

Do not label every AI-generated image as a deepfake. The Regulation reserves this term for image, audio, or video generated or manipulated via AI that resembles existing persons, objects, places, entities, or events and may appear falsely authentic or truthful.

An overtly fantastical illustration may not meet this definition. A synthetic photograph of a protest that never occurred, a recording that attributes new words to a real person, or a visual reconstruction that appears to be a documentary photograph may indeed meet it.

Step 2. Determine if there is an obligation to inform

Image, audio, or video

Disclose when these three conditions are met:

1. The content has been generated or manipulated via an AI system.
2. It resembles an existing person, object, place, entity, or event.
3. It may appear falsely authentic or truthful to a person.

If the appearance of authenticity is absent, the specific obligation under Article 50.4 may not apply. Nevertheless, you may disclose voluntarily for editorial transparency, due to institutional policy, or because the context may be misleading.

Text concerning matters of public interest

Disclose when the text:

1. Has been generated or manipulated via AI.
2. Is published to inform the public about matters of public interest.
3. Has not undergone sufficient human review or editorial control, or no natural or legal person assumes editorial responsibility.

The specific obligation does not apply when the text has been subjected to human review or editorial control and a natural or legal person assumes editorial responsibility. Both conditions must be met.

A superficial reading is insufficient. Document who reviewed the content, which sources were verified, what changes were made, and who authorised its publication. If an individual is accountable for the text, but no one verified the generated claims, it is inadvisable to treat the exception as an automatic formula.

Content authorised by law

The disclosure obligation under Article 50.4 does not apply when the use is authorised by law for the detection, prevention, investigation, or prosecution of criminal offences. This exception does not render any police or judicial use free: there must be an applicable legal authorisation.

Step 3. Decide if the artistic regime modifies the form

When a deepfake forms part of a work or programme that is evidently artistic, creative, satirical, fictional, or analogous, the obligation does not disappear. It may be fulfilled in an appropriate manner that does not impede the exhibition or enjoyment of the work.

Depending on the context, the information may appear in:

• the caption;
• the gallery text;
• the work's documentation;
• the catalogue;
• the programme;
• the opening or closing credits;
• the exhibition webpage;
• a second layer accessible via the icon.

Do not conceal the information on a remote page if the content may circulate independently of it. A gallery photograph, a video clip, or an image used in the press may lose the caption that rendered its fictional nature evident. Provide a version of the statement that accompanies every reuse.

Artistic adaptation affects the manner of disclosure; it does not legitimise what might otherwise infringe upon rights of identity, authorship, performance, personal data, or contract.

Step 4. Choose the icon

The European Commission offers three functional icons for deployers:

• Basic icon. Use this when AI has intervened in the creation of the disclosed content or when you add your own textual label or an interactive second layer.
• Fully AI-generated content. Use this when all disclosed content has been generated via AI and contains no elements created by a person or human editorial control, beyond the instructions provided to the system.
• Partially AI-modified content. Use this when pre-existing human content has been altered via AI to become a deepfake or a text subject to disclosure.

Each icon is offered in black, white, black with 50% transparency and white with 50% transparency. It is available in SVG and PNG. Its use is free and does not require attribution to the Commission or the AI Office.

Do not present the icon as a European seal of approval. It does not certify that the content is lawful, accurate, secure, or compliant with the Regulation. It merely communicates the intervention of AI.

Step 5. Add an understandable label

The Commission indicates that user testing improved when the basic icon was accompanied by a brief text. Use a specific description.

Recommended phrasing:

• Image generated with AI.
• Image modified with AI.
• Voice generated with AI.
• Video with image and voice generated with AI.
• Fictional recreation generated with AI.
• Original photograph modified with AI.
• Text generated with AI and not editorially reviewed.

Phrasing to avoid:

• Digitally created. Does not identify AI intervention.
• Enhanced with technology. Conceals the nature of the modification.
• Innovative content. Is promotional and does not inform.
• May contain AI. Does not specify whether the tool actually intervened.
• Responsible AI. Makes an assessment that the icon does not verify.

If a real person appears, add the necessary information to avoid false attributions. For example: "Fictional recreation generated with AI. The person depicted did not speak these words."

Step 6. Place it at the appropriate time

The information must be clear, distinguishable, and accessible, and provided no later than the person's first exposure to the content.

On a webpage

• Place the icon and the label next to the image, the player, or the beginning of the text.
• Do not relegate them solely to the general footer of the site.
• Add alternative text or an ARIA label indicating that the content was generated or modified with AI.
• If an expanded explanation exists, ensure it can be opened and navigated using assistive technologies.

On social media

• Include the declaration in the image or video itself when it is necessary for it to survive redistribution.
• Repeat it in the text of the post.
• Do not rely solely on an external link or metadata that the platform may remove.

In video

• Display the information before or at the start of the affected segment.
• Keep the notice visible long enough to be read.
• If only a portion is generated or modified, identify the corresponding section.
• Repeat the information in the description and credits when it helps to preserve the context.

In audio

• Include a comprehensible announcement before the first appearance of the synthetic voice or sound.
• Repeat the information in the description, file details, or transcript.
• Do not rely on a visual icon as the sole means of information.

In an exhibition

• Place the information on the label, at the entrance to the room, or on the mediation device before the encounter with the content.
• If a warning about the work interferes unjustifiably with its enjoyment, utilise the artistic regime and offer appropriate disclosure within the exhibition context.
• Include the declaration in press images and materials that may circulate outside the room.

In printed publications

• Place the icon and the text in the image caption, credits, or legal page, depending on the type of content.
• Do not refer solely to a URL that may cease to exist.
• Maintain a sufficient explanation within the publication itself.

Step 7. Ensure accessibility

The Commission recommends that the icon be of a clearly visible size, that the text use simple language, and that abbreviations other than "AI" be avoided.

Also check:

• that there is sufficient contrast with the background;
• that the icon is not covered by controls, subtitles, or interface elements;
• that the temporary notice remains visible for the necessary time;
• that there is a text alternative for those who cannot see the icon;
• that there is an audio alternative or transcript where appropriate;
• that the second layer is navigable via keyboard and assistive technologies;
• that the meaning does not depend on colour alone.

Transparency that cannot be perceived by the public does not fulfil its function.

Step 8. Make the statement travel with the content

The Code of Practice insists that the icon be visible when content is shared or downloaded, except for the adaptations provided for creative works. Design the publication with captures, crops, insertions, and republications in mind.

Useful measures:

• embedding the icon in the image or video;
• keeping a version with a label for the press and social media;
• including the declaration in photo captions and credits;
• adding machine-readable metadata when the tool allows;
• preventing a thumbnail from suppressing the only notice;
• maintaining a stable link with expanded information;
• providing third parties with reuse instructions.

The provider's technical marking and the visible label serve different functions. Maintain both when available.

Step 9. Keep an internal record

For each piece of content, note:

• title or identifier;
• date of creation and publication;
• system and version used, if known;
• relevant instructions;
• initial human elements;
• generated parts and modified parts;
• real persons, places, objects, or events depicted;
• rights and authorisations verified;
• human review conducted;
• editorial responsibility;
• decision on the application of Article 50;
• icon and phrasing used;
• places where it was published;
• versions replaced or withdrawn.

This record does not have to be entirely public. It serves to justify the decision, correct errors, and maintain the labelling when the content is reused.

Resolved cases

Illustrated article header

A completely synthetic image represents a clearly illustrative composition and does not intend to pass as a photograph of a real person, place, or event.

Result: it may not be a deepfake and may not trigger the specific obligation of Article 50.4. It is advisable to declare the use of AI through editorial policy or the image caption. If it realistically depicts an existing place or event and could be mistaken for a photograph, it must be re-evaluated.

Voluntary phrasing: "Header image generated with AI."

Photograph of an exhibition expanded with AI

A real photograph incorporates works, visitors or architectural elements that were not present via AI, while maintaining a documentary appearance.

Result: it is likely to constitute a deepfake if it may appear authentic.

Icon: content partially modified with AI.

Formula: "Original photograph modified with AI. Elements not present in the exhibition were added."

Historical recreation

A museum generates a realistic scene of an event for which no images exist.

Result: it may constitute a deepfake because it represents an existing event with an authentic appearance. The educational context does not remove the obligation.

Icon: content entirely generated by AI.

Formula: "Fictional historical recreation generated with AI. This is not a period document."

Synthetic voice of a real person

A video uses a vocal replica to attribute new phrases to an identifiable person.

Result: it may constitute a deepfake. The label does not replace consent nor does it resolve rights regarding the voice and performance.

Icon: basic or entirely generated, depending on the content as a whole.

Formula: "Voice generated with AI. The person represented did not make this recording."

Informative text generated and published without review

An institution automatically publishes a summary of a new regulation without anyone verifying the sources or assuming editorial responsibility.

Result: disclosure is required if the text intends to inform the public about a matter of public interest.

Icon: content entirely generated by AI.

Formula: "Text generated with AI and published without human editorial review."

Text prepared with AI and substantively reviewed

A person uses AI to prepare a draft, checks the sources, corrects the content and the institution assumes editorial responsibility.

Result: the specific obligation of Article 50.4 does not apply to this text. An academic, professional or institutional policy may exist that requires declaring the use.

Voluntary formula: "AI was used as support in the preparation of the draft. The text was reviewed, verified and editorially assumed by [responsible person]."

Satirical work exhibited in a museum

A piece uses a synthetic video of a public figure and its satirical nature is evident within the exhibition.

Result: the obligation may be fulfilled in an appropriate manner without disrupting the exhibition. An accessible label or catalogue entry may suffice in the gallery, but fragments disseminated outside of it must retain the context.

Formula: "Satirical work. Image and voice generated with AI."

What the label does not authorise

Before publishing, verify separately:

• copyright on the source works and materials;
• rights of performing artists;
• consent to use voice, face, image or identity;
• personal data protection;
• licences from archives, museums and image banks;
• conditions of communities and holders of cultural heritage;
• accuracy of the claims;
• advertising, competition and consumer protection;
• contracts with suppliers and collaborators.

A correct label does not make an unauthorised use lawful. The guide Artificial intelligence in exhibitions and institutions expands upon these checks. For academic work, consult How to declare the use of AI in an academic paper.

Final checklist before publishing

• ☐ I have identified which parts were generated or modified with AI.
• ☐ I have checked whether the content may appear authentic or truthful.
• ☐ If it is text, I have determined whether it informs about a matter of public interest.
• ☐ I have documented the human review and editorial responsibility.
• ☐ I have decided whether the regime for artistic or analogous works applies.
• ☐ I have chosen the icon and a precise label.
• ☐ The information appears, at the latest, upon the first exposure to the content.
• ☐ The notice remains when sharing or downloading the material.
• ☐ The statement is accessible and does not rely solely on a visual icon.
• ☐ I have kept a record of the tool, the changes, the sources and the decision.
• ☐ I have checked rights, consent, data protection and accuracy separately.

Corrections and contributions

If you detect an incorrect interpretation, a regulatory update or a case that should be incorporated, you can write to me. I will correct confirmed errors and update the verification date.

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", update of 14 October 2026. Icons, variants, uses, placement, accessibility and licence.
• European Commission, "Code of Practice on Transparency of AI-generated Content", update of 31 July 2026. Voluntary code for providers and deployers.
• European Commission, "Transparency obligations under Article 50 of the AI Act", 24 July 2026. Questions and answers regarding providers, deployers and personal and professional uses.