---
title: "The EU icons for AI-generated content: what they are, when to use them and where to place them"
url: https://zerofive.ai/en/blog/compliance/eu-ai-icons-labelling-ai-generated-content-article-50
canonical: https://zerofive.ai/en/blog/compliance/eu-ai-icons-labelling-ai-generated-content-article-50
language: en
published: 2026-10-25
updated: 2026-10-21
author: "ZeroFive.AI"
tags: AI Act, Article 50, deepfake, AI labelling, transparency
abstract: "AI Act Article 50 and the EU Code of Practice: which content to label, the three EU icons and where to place them on video, images, audio and text."
---

# The EU icons for AI-generated content: what they are, when to use them and where to place them

**In brief.** Since 2 August 2026, deepfakes and AI-generated text on matters of public interest must be disclosed to the people who see them, at the latest at first exposure. The European Commission has released three free icons (AI, AI GENERATED, AI MODIFIED) with different placement rules for video, images, audio and text. The icon must be embedded in the content and survive downloads and reshares, while text reviewed by a qualified person who takes editorial responsibility is exempt, provided you can prove it.

Article 50 of Regulation (EU) 2024/1689 has applied since 2 August 2026, and anyone who publishes a deepfake or an AI-generated text on a matter of public interest must tell the people who watch, listen or read, at the latest when the content first reaches them. On 10 June the European Commission had published the final version of the [Code of Practice on Transparency of AI-Generated Content](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content) and, inside the Code, three free icons designed to sit on a video, an image or an article. The final Guidelines on Article 50 followed on 20 July, and with them the framework is complete.

If you look at how many companies flag their use of AI today, two solutions dominate: the sparkle symbol next to "smart" features, or a line in the website terms of use. Neither meets the European rules, and understanding why is the fastest way to understand how the icons work and how to use them.

## The scope of Article 50

Article 50 contains four separate transparency obligations, split between those who develop systems (providers) and those who use them professionally (deployers). Providers must inform people when they interact with an AI system, such as a chatbot, and must mark everything their generative systems produce in a machine-readable format, with signed metadata and invisible watermarks. Deployers must inform people exposed to emotion recognition or biometric categorisation systems and, above all, visibly label deepfakes and AI-generated text on matters of public interest. The EU icons serve this last obligation, the one in paragraph 4.

The distinction has a practical consequence that is often missed: the invisible marking added by the model provider is not enough for the deployer. The Commission says so explicitly in its [FAQ on Article 50](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act), because the label must be perceivable by a person without technical tools or dedicated actions, and metadata hidden in a file is not.

A deployer is any natural or legal person using an AI system under its own authority for a professional activity. A company remains the deployer even when the content is produced by freelancers or suppliers working under its control, and employees (designers, content creators, journalists) are not separate deployers. Personal use is excluded, but a creator who earns a regular economic benefit from their activity falls within scope.

## The content that must be labelled

Not everything that passes through a generative model needs a label. The obligation covers two categories, and they deserve a careful reading because they are broader than they look.

The first is deepfakes, meaning generated or manipulated images, audio or video that resemble existing (or plausible) persons, objects, places, entities or events and could appear authentic to the viewer. The Guidelines ask you to assess three criteria together: the degree of resemblance, the reference to something that exists or could exist, and the capacity of the content to look real. Context matters, and so does the composition of the audience likely to see it. A deepfake does not require a politician or a famous face: among the Commission's official examples is an authentic photo of an empty flat furnished with AI, a case that directly concerns estate agents, e-commerce and anyone producing product images. Background scenes and special effects in an ordinary film production, by contrast, are unlikely to make something look authentic when it is not.

The second category is text generated or modified with AI and published to inform the public on matters of public interest: politics, health, justice, security, the environment, consumer protection, economic, financial and scientific developments. An automated news summary on a newspaper website is clearly in scope, and so is an investor report reworked with AI and published on a listed company's website.

For text there is an exemption, and it is the subject of most of the questions we receive. No label is needed if the text has gone through human review or editorial control, and a natural or legal person has taken editorial responsibility for it. The Commission defines human review as a deliberate examination of the substance by someone with expertise in the subject, and editorial control as the actual power to approve, change or reject a text on its merits, including source checking. Spell-checking or a formal check does not count.

There are two exceptions. If a deepfake is part of an evidently artistic, creative, satirical or fictional work, the obligation is reduced to a disclosure that does not hamper enjoyment of the work. Advertising benefits from this regime only in specific cases, though, and when content is both informative and creative the Guidelines let the informative character prevail. The second exception covers uses authorised by law to prevent or prosecute criminal offences.

## The EU icons and when to use them

The icons are in Annex 1 of the Code and were user-tested before publication. There are three, all built on the acronym "AI" written in English and in capitals.

![The three EU icons for AI-generated content: basic AI icon, AI GENERATED for fully generated content, AI MODIFIED for partially modified content, with use cases and examples](/eu-ai-icons-which-to-use-en.png)

The basic icon shows only "AI" and is used when AI was involved in creating the content and is paired with a custom text label or an interactive second layer, for example a tooltip explaining what was generated. The Commission's example is a video with the text "voices generated with" followed by the icon.

The "AI GENERATED" icon is used when the content is entirely produced by AI, with no human-created elements and no human editorial control beyond the prompt: a synthetic video of a politician, music or images composed by the model, an automatically generated news summary.

The "AI MODIFIED" icon is used when original, human-made content has been partly modified with AI to the point of becoming a deepfake or a text of public interest: a face swapped in a real photograph, the digitally furnished flat from the earlier example.

Each icon comes in four variants (black, white and the two 50% transparent versions), in SVG and PNG, can be downloaded from the [Commission's page](https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content) and can be used freely, without attribution. User testing showed that the basic icon performs better when paired with a text label. Two clarifications prevent the most common misunderstandings. Using the EU icons is optional while the labelling obligation is not, and an equivalent proprietary label must still carry the acronym "AI" in English, in capitals, and follow the same design and placement specifications. Using the icon without having signed the Code also does not signal adherence to the Code.

At this point it is clear why the sparkle falls short: it signals that a product has an AI feature, not that a specific piece of content was generated or modified with AI.

## Placement by format

The Code applies the same principle to every format, namely a label that is recognisable immediately, without interaction, present at first exposure and visible long enough, and translates it into different rules for video, audio, images and text.

![Where the AI label goes, format by format: video at the start and after every interruption, image with the icon embedded in the file, audio with a spoken notice at the start, text with the icon next to the headline](/ai-label-placement-formats-en.png)

| Format | Where the label goes | Notes |
|---|---|---|
| Video | At the start and after every interruption, at regular intervals where possible | Continuous display is encouraged, and for live streams it is the safest choice |
| Image | Embedded in the image, in a fixed and clearly visible position | No graphic elements overlaid on the icon |
| Audio only | A short spoken notice in plain language at the start | In long formats, such as podcasts, it must be repeated |
| Audio with a screen | Spoken notice plus a visual icon | Additional sound cues are possible |
| Text | Above or next to the headline, or in the opening colophon | Short texts must be labelled too, and you may label only the generated part |
| Creative and satirical works | Adjacent or contextual placement | The disclosure must not hamper enjoyment of the work |

For a company website or blog, two practical choices follow. Product images or illustrations that meet the deepfake definition carry the icon inside the file, not only in the page layout. Articles that do not go through editorial review carry the icon next to the headline, in the same position for every piece, so readers learn to recognise it.

## Visibility, accessibility and resharing

A label that only appears when hovering over an image, or that sits on the terms and conditions page, is not a label for the purposes of Article 50: the disclosure must be clear and distinguishable at the latest at first exposure, and the Guidelines expressly rule out a note in the terms of use.

The rule that causes teams the most trouble is the one on resharing. The icon must be embedded directly in the deepfake or the text, unless equivalent alternatives such as an interface overlay are available, and it must remain visible when the content is downloaded or reposted elsewhere. An icon added via CSS on top of an image disappears at the first download, and the same goes for a badge applied only in the site's video player: the file circulating on another channel arrives without a label.

On accessibility, the Code asks for an icon of clearly readable size, any text label in plain language, without jargon and without abbreviations other than "AI", alt text or an ARIA attribute readable by assistive technologies, a display time long enough for people with cognitive difficulties, and an interactive second layer, if there is one, that can be navigated by keyboard and screen reader.

## The Italian layer

In Italy the AI Act sits alongside Law 132/2025, which designated AgID and ACN as the national AI authorities and also amends the Consolidated Act on Audiovisual Media Services. For audiovisual and radio media service providers, the Consolidated Act provides for an identifying mark on content produced wholly or partly with AI, to be shown at the start and end of the broadcast and of the content and at every resumption after an advertising break, excluding manifestly creative, satirical, artistic or fictional works and with AGCOM as the reference authority. Broadcasters therefore have to coordinate two regimes, the European "AI" icon and the national one.

Penalties for breaching Article 50 reach 15 million euros or 3% of annual worldwide turnover, with proportionality criteria for SMEs and small mid-caps. Enforcement is largely in the hands of national market surveillance authorities, while the AI Office steps in only for systems built on general-purpose models from the same provider or integrated into the very large platforms designated under the Digital Services Act.

## Timeline and grace period

Article 50 has applied since 2 August 2026 to everyone, providers and deployers (what became binding on that date is covered in our article on [2 August 2026](/en/blog/compliance/ai-act-august-2-2026-what-is-binding)). The only grace period, introduced by the AI Act revision under the Digital Omnibus, concerns the invisible marking required of providers for generative systems already on the market before that date, which have until 2 December 2026. There are no extensions for deployers.

Content generated before 2 August does not need to be labelled retroactively, although the Commission recommends it. For images, audio and video the generation date counts, for text of public interest the publication date counts: a text written with AI in July and published in September must be labelled.

The Code is voluntary. The Commission and the AI Board assessed it as adequate to demonstrate compliance with paragraphs 2, 4 and 5 of Article 50, and those who do not sign it must demonstrate compliance by other means, expecting more requests for information from the authorities.

## Checks before publishing

Before publishing AI-produced content, it is worth running through a short sequence of questions.

- Is the content a deepfake under the three criteria of the Guidelines, assessed against the real audience that will see it, or an informative text on a matter of public interest?
- If it is text, has a competent person reviewed the substance, and has someone taken editorial responsibility for it? Is it documented who, when and on which version?
- Which of the three icons describes the case: generated, modified, or the basic icon with a text label or second layer?
- Is the icon embedded in the file or the text, and does it survive downloads and resharing?
- Does the placement follow the rules for the format (start and interruptions for video, spoken notice for audio, headline or colophon for text)?
- Are there alt text or ARIA attributes, and a plain-language label?
- If the content is creative or satirical, is the disclosure present without hampering enjoyment, and does the informative component not prevail?

The human review exemption is only worth the evidence you can show, and in an inspection "we always reread it" proves nothing. A minimal editorial log (content, tool used, reviewer, date, outcome) turns a practice into evidence, and costs far less than any penalty.

## Before the label

The icon is the last step in a chain of decisions made earlier: which content the company produces with AI, who reviews it, who is accountable for it, where the trail is kept. These are governance decisions: those who have already made them apply the icons as a production step, those who have not must first reconstruct which generated content they have published since 2 August. It is worth asking how many of the images and texts online right now, on European company websites, would pass the sequence of questions above.

## Where to start

In the [AI Rating](/en/services/ai-rating) model, transparency obligations fall under the AI Risk dimension, where baseline compliance and governance are non-compensable gates: the assessment shows which generated content the company already publishes, who reviews it and where the review trail is missing. For the full picture of the obligations there is our [guide to the EU AI Act](/en/blog/compliance/eu-ai-act-complete-guide-obligations-liability-data).

To check where you stand on content transparency you can [start an AI Rating](/en/start-ai-rating) or [book an assessment call](https://calendly.com/fabiolalli/zerofive).