Accessibility and generative AI: a still poorly negotiated partnership

Generating image descriptions or suggestions for accessible HTML structure with AI has become commonplace. It is a real time saver, provided you do not trust it blindly.

What AI genuinely helps with

• First drafts of alternative text for large volumes of images
• Automated detection of contrast or structural defects on large sites (an online contrast checker remains deterministic, unlike a generative model)
• Suggested corrections for semantic HTML markup

Where the risk appears

An automatically generated image description can be grammatically perfect and semantically useless to a screen reader user if it does not convey the information the image actually carries in its context.

Generated alternative text: what RGAA criterion 1.3 really requires

The RGAA (the French web accessibility standard) does not require an image to have a description, it requires it to have the right one. Criterion 1.3 of the official method puts the question this way: for each informative image with a text alternative, is that alternative relevant? Test 1.3.9 adds that it must be short and concise. Criterion 1.2 deals with the opposite case: a decorative image must be ignored by assistive technologies, and therefore say nothing at all.

This is precisely where a generative model goes wrong. It describes what it sees, whereas the criterion is about what the image means at that point on the page. Two typical cases:

• A logo placed in a link to the home page. Generated text: "Blue and white logo with a stylised circle". Relevant alternative: "Home" followed by the company name, since what matters is the link's destination.
• A chart showing sales trends. Generated text: "Coloured bar chart". Relevant alternative: a sentence that gives the trend, supplemented by a detailed description or a data table, which falls under criterion 1.6.

In both cases, the generated text is accurate. It is simply useless.

The right discipline

Treat any AI output related to accessibility as a draft, never as a final deliverable. Human review, ideally by someone trained in how assistive technologies are actually used, remains non-negotiable.

The problem the law does not forgive: non-determinism

One technical point changes everything for professional use of AI in audits: a generative model is not deterministic. Running the same analysis twice on the same day can produce two different verdicts on the same image or component, whereas an RGAA audit must be stable and legally enforceable. Generative AI can therefore serve as a quick pre-diagnosis across a large volume of pages, never as proof of compliance in itself.

Automated tool or generative AI: determinism makes the difference

The two are often filed under the same "AI" label, even though they do not produce the same type of result.

Type of tool	Reproducible	Value in an audit	Use I recommend
Rule-based tester (contrast, heading hierarchy, missing attributes)	Yes, same input, same result	Reproducible finding, not a compliance verdict	Systematic detection across an entire site
Generative model (alternative text, code correction)	No, two runs can diverge	None without review	First draft reviewed by a human
Overlay or corrective widget	Varies from page to page	None	None, see below

AI-powered accessibility overlays and widgets

Overlays promise to make a site compliant by adding a single line of script. The US consumer protection regulator has sanctioned this kind of promise. On 22 April 2025, the Federal Trade Commission approved a final order requiring accessiBe to pay $1 million, following a complaint and a proposed settlement announced in January 2025. The order prohibits the company from claiming, without evidence, that its automated products can make a website WCAG compliant or keep it compliant over time.

The decision is American, but the lesson applies in France. An overlay acts on the page displayed in the browser, not on the source code or on downloadable documents. It replaces neither an audit nor a fix in the code.

My human review protocol

For generated content, I review with five questions, in this order:

1. Is the image decorative? If so, the alternative must be empty, whatever the tool suggested.
2. What does the user lose if the image disappears? The alternative must convey that, not the appearance.
3. Does the alternative repeat the neighbouring text? If so, I shorten it or empty it.
4. Is the text short? A complex image points to a detailed description rather than to a three-line alternative.
5. Will the result serve as evidence? If so, I redo the check by hand and record it, rather than citing a model output.

This review costs less than defending a contested accessibility statement.

The role of the RGAA expert is shifting

AI does not erase the role of the RGAA expert, it shifts it: less time spent identifying obvious defects, more time devoted to settling the ambiguous cases that AI cannot decide on its own. There is also everything the RGAA does not cover, which neither AI nor the audit can see. Version 5 of the RGAA, which the DINUM (the French interministerial digital directorate) announces for publication by the end of 2026 incorporating WCAG 2.2 (a W3C Recommendation since 5 October 2023), will change nothing about this principle: compliance remains a tool-assisted human judgement, not an automatic score.

AI speeds up accessibility work. It does not remove the need to know what accessibility is actually meant to serve.

This human review is something I systematically build into my accessibility practice, including on AI-generated content.

Further reading: RGAA: the errors that come up in 9 audits out of 10

Frequently asked questions

Can AI make a website RGAA compliant?

No. It speeds up the detection of defects and the production of first drafts, but compliance is established criterion by criterion, on a sample of pages, by a person who knows how to settle ambiguous cases.

Is AI-generated alternative text acceptable?

As a draft, yes. Criterion 1.3 requires an alternative that is relevant in its context, which no generator guarantees without review, and criterion 1.2 requires a decorative image to say nothing at all.

Will RGAA 5 change the game for AI?

According to the DINUM announcement of 2 March 2026, it will incorporate WCAG 2.2, define criteria for mobile applications and office documents, and designate Arcom as the supervisory authority. Nothing in that announcement opens the door to an automatic compliance verdict.