How do you make an AI label accessible?

An AI label is the notice that tells people content was made or changed with AI: an "AI" icon on an image, a caption under it, a line at the top of a chat window. Most advice about labels covers wording and placement. This guide covers the part the EU rule makes explicit and many implementations skip: whether everyone can perceive the label, including people who use screen readers, magnification or high-contrast settings.

It covers the EU AI Act, the Commission's own guidance, the WCAG criteria a label touches, and the problem that a label drawn into an image is invisible to assistive technology. It is general information, not legal advice, and we checked every source on October 21, 2026. For which content needs a label at all, start from our AI disclosure laws hub and the page on Article 50 of the EU AI Act.

What Article 50(5) says about accessibility

Article 50 sets transparency duties for AI systems and the content they produce. Paragraph 5 says how the information must be delivered: "The information referred to in paragraphs 1 to 4 shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. The information shall conform to the applicable accessibility requirements." That is the text of Regulation (EU) 2024/1689. Article 50 has applied since August 2, 2026.

Does that create a new accessibility standard? The Commission says it does not. Its Guidelines on Article 50 (C(2026) 5054 final, July 20, 2026, non-binding according to paragraph 5) say in paragraph 144: "Article 50 AI Act does not impose distinct or additional accessibility requirements." The same paragraph tells providers and deployers to be "aware of the applicable accessibility requirements (e.g. under Directive 2016/2102 and Directive 2019/882)" and to "assess whether their product, service or content needs to comply with those rules."

So the test for an AI label is the accessibility law that already applies to you and your service. Paragraph 142 adds a plain-language test for the word "clear": information is clear where it is "noticeable, easy to understand by and accessible for the natural person concerned (including persons with disabilities)". It is not clear if it "can be easily overlooked or missed by natural persons under normal exposure or interaction conditions."

What the Code of Practice and the icon page add

The Commission has also published a Code of Practice on transparency of AI-generated content (final text June 10, 2026, found adequate by the Commission on July 8, 2026) and a set of free EU icons. Both say more about accessibility than the Act does, with one limit. The Code binds only the organizations that signed it; the Commission's page of July 31, 2026 lists 95 signatories to Section 1 and 192 to Section 2. For everyone else, paragraph 148 of the Guidelines says to show how you complied by other adequate means, for instance with a gap analysis against the Code.

Section 2 of the Code, which covers organizations that publish deep fakes, asks in Measure 1.1 for accessible disclosure in line with the European Accessibility Act and the Web Accessibility Directive. It lists four applications:

• "audio descriptions or alternative cues for visual disclosure elements";
• "tactile and haptic cues for audio-only content";
• "high contrast icons or equivalent labels and screen-reader compatibility", including for end-users with color vision deficiencies;
• "detectability of the icon or equivalent label by assistive technologies".

The Code encourages EN 301 549 and WCAG 2.1. Its placement measure asks that the label be recognized "without requiring user (inter)action or sustained attention" and "remain visible against any background".

The Commission's EU icons page (last updated October 13, 2026) lists accessibility measures that users of the icons are "encouraged to implement":

• The icon should be in a clearly visible size, and any accompanying label should use plain language.
• "If possible, the icon should be readable by assistive technologies using alt text or ARIA labels indicating that the content is AI-generated or manipulated."
• A disclosure that appears for a limited time should stay visible long enough for people with cognitive or processing difficulties to read it.
• A second, interactive layer should be signaled by the icon and navigable with assistive technologies.

The same page says "The use of these EU icons does not establish legal compliance by itself." In our reading, neither the Code, the Guidelines nor the icons page names a minimum contrast ratio, a size in pixels or a duration in seconds. The Commission leaves those measurements to you.

Which accessibility rule applies to your page

Because Article 50(5) points at existing law, the first question is which accessibility rule already binds the page. This is our reading of the sources, not a legal opinion.

EN 301 549 version 3.2.1 remains the reference. The AccessibleEU Centre said on September 7, 2026 that version 4.1.1, which follows WCAG 2.2, is not yet the legal reference standard. Our European Accessibility Act guide and the EU accessibility law page explain who the EAA covers.

The microenterprise exemption deserves a precise reading. Article 3(23) defines a microenterprise as "an enterprise which employs fewer than 10 persons and which has an annual turnover not exceeding EUR 2 million or an annual balance sheet total not exceeding EUR 2 million". Article 4(5) exempts such an enterprise, when it provides services, from the accessibility requirements. So for many small shops the EAA adds nothing binding to their labels. Yet paragraph 142 of the Guidelines builds accessibility into the meaning of "clear", and the Guidelines do not explain how the two fit together. A lawyer should answer that, and the practical point stands either way: a label that screen reader users cannot perceive is a weak label.

The WCAG criteria an AI label touches

The table maps the criteria in WCAG 2.2 that an AI label can meet. The mapping is our analysis of the Commission's wording; the Commission does not publish one.

The pixels problem

The Code asks that the label be "directly embedded into the content, unless equivalent alternatives to an embedded icon are available (e.g., a user interface overlay that for natural persons appears to be on the content)". The icons page adds that the icon "must be visible when content is reshared or downloaded." Both push toward a label drawn into or on top of the picture. A screen reader cannot read pixels, and WCAG treats an image of text as non-text content that needs a text alternative. Our guide to images of text covers the rule. We found no passage in the Code, the Guidelines or the icons page that says where the text equivalent of a label drawn into the picture itself belongs.

The workable design, again our analysis, has two layers. Layer one is the visual label that everyone who can see the image gets: the "AI" icon, ideally with a word such as "generated" or "modified", embedded or overlaid. Layer two is the same message in text: the icon's own alt text, an aria-label on an overlay, or a visible caption in real text beside the image. Use one text carrier, not three. If the visible label is real text, the icon beside it can have an empty alt, so a screen reader does not read the message twice.

<!-- Option A: an overlay icon that carries its own name -->
<div class="image-frame">
  <img src="/images/street-scene.webp"
       alt="Woman in a green coat crossing a rainy street">
  <img class="ai-icon" src="/images/ai-generated-black.svg"
       alt="AI-generated image" width="56" height="56">
</div>

<!-- Option B: a visible caption in real text -->
<figure>
  <img src="/images/street-scene.webp"
       alt="Woman in a green coat crossing a rainy street">
  <figcaption>AI-generated image</figcaption>
</figure>

Whether the host image's own alt text should also say it is AI-generated is a separate question that nobody has settled; we cover it in should alt text say an image is AI-generated. Our guide to captions, alt text and figcaption explains how the three work together. Check the zoom, lightbox and mobile views of every image, because a label added as an overlay may not be present in them.

A checklist for an accessible AI label

1. Decide whether the content needs a label. Article 50(4) covers deep fakes, meaning realistic image, audio or video content that "would falsely appear to a person to be authentic or truthful" (Article 3(60)), and some AI-written text on matters of public interest. Many AI images are neither. Our virtual staging and synthetic performer guides show how that question plays out.
2. Use plain words. The Code makes the capitalized "AI" the main element, and the Commission says performance improved across all measures when the basic icon was accompanied by a text label such as "modified".
3. Put the words in text. If the label has to be in the pixels, add the same words as alt text or as a caption.
4. Choose the color variant for the worst case. On a busy photo, the black or white version is a safer start than a 50 percent transparent one. Measure the weakest spot behind it with our image contrast checker, and read text over image contrast for the method.
5. Make it large enough to notice without zooming. The Commission gives no number, so test it at phone width.
6. Give a second layer a real control, a button with a name that the keyboard reaches. Avoid a label that times out, and if it must, keep it up long enough to read.
7. Test with a keyboard and a screen reader. Automated checks cannot tell you whether a label is noticeable to a person.

What a check can and cannot tell you

An automated check can read whether label wording exists in a page's text, alt text or captions, whether the element is hidden from assistive technology, and whether an icon has a name. It cannot tell whether an image is AI-generated, whether it is a deep fake in the legal sense, whether a given screen reader announces your chat greeting, or whether a person would notice the label. Our free AI label checker runs in your own browser on the live page and reports where label wording sits, and our accessible AI label generator writes label text and markup with the text equivalent included. Neither says your label is sufficient.