Skip to content
AI Act Checker

Article 50(4) and (5) · Code of Practice, Section 2

How to label AI-generated content.

Article 50 requires you to say when AI made it. Free icons, the wording that works, and where to put it.

The EU icons

Which EU AI icon should I use?

Published by the AI Office as Annex 1 of the Code. No attribution needed. Pick the icon that matches what the AI did and the tone that stays readable on your background.

AI-generated content

Icon

Fully created by AI, with no human elements beyond the prompt.

Tone

Background

Markup for this choice

<img src="/eu-ai-icons/ai-generated-white.svg" alt="AI-generated content" width="126" height="40">
Download: ai-generated-white.svg
Add an icon to your own image with the free labelling tool

The Commission makes the icons publicly available for everyone to use freely, without attribution. Signatories of the Code must follow its placement rules; anyone else may use the icons too.

Official packs:SVG pack (ZIP)PNG pack (ZIP)

Scope

Do I have to label my content?

Answer up to four questions. The verdict names the article, commitment or exception it rests on.

  1. 01

    What are you publishing?

Measure 1.1

What should an AI label look like?

  1. 01

    "AI" is the main element

    The capitalised acronym in English, unless national language law requires otherwise. Both letters share the same height; keep the proportions when you resize.

  2. 02

    A second layer says what happened

    "Generated" or "modified", and where possible what was changed: a face, a voice, the background. Interactive layers that expand on tap are encouraged.

  3. 03

    Style is free, clarity is not

    Size, colour, contrast and typography may vary as long as the label stays readable and recognisable to natural persons.

  4. 04

    Audio gets words, not pixels

    Where no visual is possible, open with a short spoken disclaimer in plain language, in the language of the content or in English. Earcons may support it, not replace it.

  5. 05

    Accessible by design

    Alt text or ARIA that names the AI origin, audio description for visual labels, haptic cues for audio, high contrast for colour-vision deficiencies, detectable by assistive technology. WCAG 2.1 AA and EN 301 549 are the reference standards.

  6. 06

    Think about who is watching

    Children, older people, low AI literacy, sensitive contexts such as health or finance. The label has to work for the least prepared person in the audience.

Measure 1.2

Where do I put the label?

Four principles apply everywhere: immediate recognition without any user action, visible long enough to be noticed, embedded in the content unless an equivalent overlay exists, and clear at first exposure with enough distance from other overlays.

  • AI-generated content

    Image

    Top right corner, no overlay in the way. The alt text names the AI origin.

    Sub-measure 1.2.2(a)

  • AI-modified content
    Label reappears:StartIntervalAfter break

    Video

    At the start, at regular intervals, and again after every ad break. Ideally throughout the deep fake segment.

    Sub-measure 1.2.2(b)

  • AI-generated or manipulated content

    00:00 · "This recording was generated with AI."

    05:00 · Reminder after five minutes

    Audio

    Spoken disclaimer first, reminders at intervals. On any screen, show the icon too.

    Sub-measures 1.2.2(e) and 1.2.3

  • AI-generated content

    Council approves new flood defences for the lower Danube

    The plan covers 140 km of embankments and is due to start next spring.

    Newsroom automation · 24 August 2026

    Published text

    Above or at the top, near the headline, or in the colophon. Label only the AI part when that is clearer.

    Sub-measure 1.2.2(f)

  • Replies in this chat are generated by AI.

    Your parcel left the depot at 07:40 and should arrive today.

    Short outputs

    Where a label would wreck a one-line output, disclose once in the interface, at the start of the session or at first contact.

    Sub-measure 1.2.2(f)

  • "Harbour, 2031", video installation. Faces generated with AI.

    Artistic works

    Adjacent to the frame, in the credits or the description, or a small icon that expands on click. Never in a way that spoils the work.

    Commitment 3

Checks

What are the dos and don'ts of AI labels?

Do

  • Show the label before or with the content, never after a click.
  • Keep it visible when the content is downloaded, screenshotted or reshared.
  • Use plain words. "AI" is the only abbreviation you may rely on.
  • Pair every visual label with alt text or an ARIA label.
  • Publish what your labels mean, on a page like this one.
  • Keep a way for people to report missing or wrong labels.

Don't

  • Hide it under a cookie banner, a player control or another overlay.
  • Rely on audio alone for a video, or on a visual alone for audio in a car.
  • Show it for one frame and drop it.
  • Strip the provider's watermark or metadata when you edit or re-upload.
  • Use grey-on-grey, tiny type or jargon such as "synth" or "genAI".
  • Assume the exception applies because a human glanced at the feed.

For developers

How do I add the label in HTML?

Three patterns that satisfy the design and accessibility rules. When adjacent text already states the AI origin, the icon gets an empty alt so screen readers do not announce it twice.

Image with embedded icon

<figure class="ai-media">
  <img src="/campaign/hero.jpg" width="1600" height="900"
       alt="AI-generated image of a flooded street in Linz">
  <img class="ai-label" src="/eu-ai-icons/ai-generated-white.svg"
       width="126" height="40" alt="AI-generated content">
</figure>

<style>
  .ai-media { position: relative; display: inline-block; }
  .ai-label { position: absolute; top: 1rem; right: 1rem;
              height: 2.5rem; width: auto; }
</style>

Article header

<article>
  <header>
    <p class="ai-disclosure">
      <img src="/eu-ai-icons/ai-generated-black.svg"
           width="126" height="40" alt="">
      This article was generated by an AI system
      and published without human review.
    </p>
    <h1>Council approves new flood defences</h1>
  </header>
  ...
</article>

Audio with spoken and visual disclosure

<!-- the recording itself opens with the same sentence,
     spoken in the language of the content -->
<audio controls src="/briefing.mp3" aria-describedby="ai-note"></audio>
<p id="ai-note" class="ai-disclosure">
  <img src="/eu-ai-icons/ai-white.svg" width="40" height="40" alt="">
  This recording was generated with AI.
</p>

Machine-readable marking (metadata, watermarks) is a separate duty for providers under Article 50(2). Do not remove markings you receive; the deployer label sits on top of them.

Commitments 2 and 4

What should I document?

Proportionate to your size, but written down.

  1. 01

    Compliance process

    Describe how you decide what gets labelled, who does it, and how you check that placement follows the rules. Media service providers may rely on their existing editorial procedures.

  2. 02

    Public explanation

    Publish which icon or label you use and what each variant means. If your label differs by context, explain the variations.

  3. 03

    Awareness and training

    Everyone who creates, edits or publishes such content knows when a label is required, how to apply it, and how to fix a missing one.

  4. 04

    Feedback channel

    Let readers, fact-checkers and researchers flag missing or wrong labels. Review substantiated reports and remedy them without undue delay.

  5. 05

    Editorial responsibility policy

    If you rely on the human-review exception for text: name the person or entity with editorial responsibility (name, role, contact), describe the review resources, and publish the contact details.

  6. 06

    Cooperation with authorities

    Keep the documentation ready for market surveillance authorities and cooperate with them under national law.

Legal basis

What is the legal basis for labelling?

Article 50(4)
Deployers must disclose deep fakes and AI text on matters of public interest, with exceptions for law enforcement, artistic works and human editorial control.
Article 50(5)
Disclosure must be clear and distinguishable at the latest at first exposure, and meet accessibility requirements.
Article 3(60)
Defines a deep fake: 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.
Recitals 133 and 135
Why marking and labelling matter, and why the Commission encourages codes of practice.
Article 99(4)(g)
Breaches of Article 50 can be fined up to EUR 15 million or 3 % of worldwide annual turnover, whichever is higher.
Code of Practice, Section 2
Commitments 1 to 4 and Measures 1.1 to 1.3 for deployers; the EU icons in Annex 1. Section 1 covers the marking and detection duties of providers.

Sources: EU icons for labelling AI-generated content · Code of Practice on Transparency of AI-Generated Content · Guidelines on transparency obligations · SVG pack (ZIP) · PNG pack (ZIP)

See whether your site already says it.

The scan checks the four Article 50 rules, quotes what it finds and tells you what is missing.

Scan a URL