AI coding tools fail WCAG: all 15 test sites missed Level A

Ask an AI coding tool for an accessible website and you get a website that says it is accessible. On September 30, AudioEye published a test in which five AI coding tools each built three sites to the WCAG 2.2 AA standard. All 15 sites failed WCAG Level A, the most basic tier. VentureBeat reported the results the same day.

What actually happened

AudioEye gave tools from OpenAI, Anthropic, Google, xAI and Lovable the same brief: build a news site, an online store and a financial services site, all compliant with WCAG 2.2 AA. The prompt asked for accessibility by name. It did not help.

The 15 sites were scanned and tested for keyboard focus, dialog handling, form error announcements and screen reader output. The findings:

• 306 distinct issues, repeated more than 59,000 times across the pages
• An average of 55 issues per page, against 62 on a typical website
• 91% of issues rated medium or high severity

So AI-built sites came out only slightly better than the average human-built web. AudioEye's explanation is that the models are not hallucinating. They learned from an inaccessible web and reproduce it at speed.

AudioEye also surveyed teams that use AI: 81% believe AI-generated code meets accessibility standards, 73% say accessibility complaints have gone up since they adopted AI, and 46% have received demand letters or lawsuits.

One caveat. AudioEye sells accessibility software, so it has a stake in this result. The release does not break the results out per tool. Treat it as a strong signal, not a ranking.

Why AI coding tool accessibility matters for your business

"Make it accessible" in a prompt is not a test. It is a wish. The tool will write aria-label on some things and miss the form that does not announce its errors. Your checkout and booking forms are where that failure costs you a sale or a demand letter.

The confidence gap is the real risk. Eight in ten teams think the output is compliant. That is how inaccessible code ships: nobody checks because everybody assumes. If your site was built or rebuilt with Lovable, v0, Claude, Codex or similar in the last year, assume it has not been audited.

It is the same bug twice. Last week we covered AudioEye's agent study: shopping agents read the accessibility tree, and they fail on bad markup. AI writes the bad markup, and AI agents then fail to use it.

The fix is a gate, not a prompt. Put an automated scanner (axe-core, Lighthouse, Pa11y) in CI so a build fails on Level A violations. Then do a manual pass on the three flows that make money: keyboard-only, then screen reader. Automated tools catch part of the problem, not all of it.

Sources: AudioEye study announcement, VentureBeat.