EU AI Labels Since 2 August 2026: What Businesses Need to Know Now

Which AI content must be disclosed?

The European Commission makes it explicit: not all AI-generated or manipulated content must be visibly labelled.

The three icons are primarily meant to support the disclosure obligations of Article 50(4) of the AI Act. This provision covers two main types of content.

1. Deceptively realistic deepfakes

This refers to AI-generated or AI-altered images, audio recordings, and videos that resemble real people, objects, places, entities, or events.

Disclosure is particularly relevant where people could mistakenly believe the content to be authentic or true.

This can include, for example, a video in which a person appears to say something they never actually said. A swapped face in a real photo or an artificially generated voice can also fall under this category.

2. Certain texts on matters of public interest

AI-generated or AI-modified texts can also be covered by the disclosure obligation. This applies to texts published to inform the public about matters of public interest.

An important exception applies where the content:

• has been reviewed or subject to editorial control by a human, and
• a natural or legal person holds editorial responsibility for the publication.

Businesses therefore do not need to automatically label every text drafted with ChatGPT or another AI tool. What matters is the purpose of the text, editorial review, and who takes on responsibility.

The three EU icons explained simply

Basic icon: AI

The basic icon generally indicates that AI was involved in creating or altering a given piece of content.

It is particularly suited to cases where a plain-language note additionally explains exactly what was generated or changed using AI. One example given by the European Commission is a deepfake video with a note such as "Voices generated with AI", to which the basic AI icon can be appended.

The icon can also be used where an additional, interactive layer provides further information on how the content was created.

In short: the basic icon means AI was involved. An accompanying note explains more precisely how it was used.

AI GENERATED: fully AI-generated

The AI GENERATED icon is intended for content that was created entirely by AI. According to the European Commission's description, that means:

• the entire relevant content was created by AI,
• there are no human-created elements,
• no human editorial control took place, and
• human involvement was essentially limited to the prompt.

The Commission cites fully AI-generated deepfake videos, AI-generated news summaries, and music or art fully composed by AI as examples.

For obviously artistic, creative, satirical, or fictional works, however, more limited requirements regarding the form of disclosure can apply.

In short: AI GENERATED means the entire content was created using AI.

AI MODIFIED: partially altered with AI

The AI MODIFIED icon indicates that pre-existing, human-created content was subsequently altered using AI. The original content already existed; AI was then used to edit, add to, or replace certain parts of it.

The European Commission cites, among others, these examples:

• a person's face is replaced with another face in a real photo,
• photos of an empty apartment are virtually furnished using AI.

In short: AI MODIFIED means the original content was human-made and was later altered using AI.

Are the EU AI icons mandatory?

No. Using precisely these three icons is voluntary.

That does not mean, however, that businesses are free to decide whether to carry out a legally required disclosure. In the cases covered by Article 50, disclosure is mandatory. Businesses can use the EU icons for this or choose another suitable, clear, and understandable form.

The European Commission also points out that merely using an icon does not by itself establish full legal compliance. Responsibility remains with the respective AI system operator or the publishing entity.

The icons can be used freely. No reference to the European Commission or the European AI Office is required for this. However, non-signatories of the associated code of conduct must not, by using the icons, create the impression that they have signed up to the code.

How should the icons be displayed?

The labelling must be clearly perceivable and understandable to the people it is aimed at. According to the European Commission's guidance, the icon should:

• be recognisable at the latest upon first exposure to the content,
• not be obscured by other elements,
• be embedded as directly as possible into the content,
• remain visible even when the content is downloaded or reshared,
• be displayed sufficiently large and legible.

Accompanying notes should be phrased in plain language. Unnecessary technical terms, unclear wording, and abbreviations should be avoided.

The Commission also recommends considering accessibility — for example, suitable alt text or ARIA labels that let assistive technologies recognise that content is AI-generated or manipulated.

What exceptions exist?

Editorially reviewed texts

For texts on matters of public interest, the disclosure obligation described above does not apply where two conditions are met:

1. the content has been reviewed or subject to editorial control by a human, and
2. a natural or legal person takes on editorial responsibility.

For businesses, this means a clearly organised review and approval process for AI-assisted texts can be legally relevant.

Artistic, creative, or satirical works

Where a deepfake is part of an obviously artistic, creative, satirical, fictional, or comparable work, an appropriate disclosure is still required. It must not, however, unduly impair the display or enjoyment of the work.

Legally permitted use

The disclosure obligation also does not apply where the use of such content is legally permitted for detecting, preventing, investigating, or prosecuting criminal offences.

What should businesses do now?

For businesses, whether AI is used is not the only decisive factor. What matters more is what it is used for and how the results are published.

An initial internal review process can be structured around five questions:

1. Is it an image, a video, an audio file, or a published text?
2. Was the content fully created with AI, or only partially altered?
3. Could outsiders mistakenly believe the content to be authentic?
4. Does a text inform the public about a matter of public interest?
5. Was the text professionally reviewed, and does someone clearly hold editorial responsibility?

In addition, businesses should define:

• who reviews AI-generated content before publication,
• who decides on a required disclosure,
• which labelling is used,
• how approvals are documented, and
• who holds editorial responsibility.

Especially for small and medium-sized businesses, this does not need to result in an extensive set of rules. A clear, documented process with defined responsibilities is often the most sensible first step.

Conclusion

The new EU icons give businesses a uniform way to transparently label certain AI-generated and manipulated content.

The most important distinction is this: using the EU icons is voluntary. A legally required disclosure is not.

Businesses should therefore not label every use of AI as a blanket rule. Instead, they need clear criteria for assessing deepfakes, published texts on matters of public interest, and content that was fully or partially created with AI.

A traceable editorial review process, clearly defined responsibilities, and understandable communication form the most important basis for this.