Five AI tools were given the same brief: build three websites each that meet WCAG Level AA. None of the 15 resulting sites cleared even Level A, the lowest of the three. The test was written up on 30 September in VentureBeat, and it was run by a company that sells accessibility remediation tools.
Key takeaways
- Tools tested: OpenAI, Anthropic, Google, xAI and Lovable
- 15 sites, all below Level A — the lowest WCAG threshold
- An average of 55 issues per page, 91 percent of them medium or high severity
- 81 percent of developers believe their AI-generated code meets accessibility rules
- The write-up is authored by the CEO of AudioEye, an accessibility tooling vendor
Three levels, none reached
The brief went to tools from OpenAI, Anthropic, Google, xAI and Lovable. The W3C specification — WCAG — organises accessibility requirements into measurable success criteria.
The W3C specification arranges its requirements into three thresholds:
The result: 15 sites out of 15 failed to meet even Level A. The average was 55 issues per page, and 91 percent of them were classed as medium or high severity. These are not cosmetic defects — Level A is the bare minimum of accessibility.
Belief versus measurement
The most interesting number is not about code but about people. 81 percent of developers believe AI-generated code meets accessibility requirements. The measurement says otherwise.
| Metric | Value |
|---|---|
| Average issues per page | 55 |
| Share of issues rated medium or high severity | 91% |
| Developers who believe AI code meets accessibility rules | 81% |
| Organisations facing a complaint, demand or lawsuit within two years | 46% |
| Of those, cases involving AI-written pages | 71% |
| Teams that noticed more issues after adopting AI | 50% |
This is not a disagreement about code quality. It is a gap between what teams assume and what an audit finds — and it has started to cost money, because it ends in a demand letter or a lawsuit.
Who counted it
The piece is signed by Kelly Georgevich, CEO of AudioEye — a company selling tools that detect and remove accessibility barriers. That does not invalidate the measurement, but it frames it: the study is run by a party with a stake in the outcome. The article also cites WebAIM data showing accessibility issues on the top million homepages rose 10 percent year over year.
Why it matters
Accessibility is one of the few areas of front-end work where a defect carries direct legal consequence. Code generators speed up site production but do not take on the liability — that stays with the publisher. The gap between 81 percent confidence and a 15-out-of-15 failure rate shows where the real risk sits: not in code quality as such, but in the fact that nobody is checking for this.
What next?
- The W3C WCAG 2.2 specification remains the reference point and defines three conformance levels: A, AA and AAA
- WCAG 2.2 added nine new success criteria over version 2.1, three of them at Level AA
- The piece gives no URLs for the audited sites and no raw reports, so the result cannot yet be reproduced
