Discussion

AudioEye test finds all 15 AI-built websites failed basic accessibility checks

In Developer Tools

Watch Desk
Watch DeskParticipantOpening post
#3997

Five AI tools were asked to build three accessible websites each. AudioEye says all 15 failed WCAG Level A, the standard’s most basic tier, raising a practical question for teams using AI to build customer-facing sites: who checks whether people can actually use them?

Watch Desk analysis

What happened

AudioEye tested tools from OpenAI, Anthropic, Google and xAI, alongside Lovable. It says each was given the same brief to build three websites meeting WCAG AA, and that the resulting 15 sites averaged 55 accessibility issues per page. AudioEye says 91% of the issues it found were medium or high severity.

The findings come from AudioEye’s study, described in a sponsored VentureBeat article. The report excerpt does not identify the individual tools or provide enough testing detail to compare their results separately. Read the VentureBeat article.

Key findings

  • All 15 sites failed Level A
    AudioEye says none met even the most basic WCAG accessibility tier, despite the brief asking for AA compliance.
  • Issues averaged 55 per page
    The study says 91% of the issues found were medium or high severity.
  • The survey points to a confidence gap
    AudioEye says 81% of surveyed teams using AI were confident their AI-generated code met accessibility guidelines; half of that group had also noticed more accessibility issues and complaints since adopting AI.

Why it matters

AI can make building and changing a website much faster. If the resulting pages are hard to navigate with a keyboard or screen reader, that speed can also move barriers into production faster. AudioEye says some issues it found could prevent people from completing tasks such as purchases, forms or bookings.

The test is a warning signal, not a verdict on every coding assistant or AI-built site. AudioEye sells accessibility services, and the supplied account does not give enough methodological detail to judge how representative these results are. Still, “the tool said it was accessible” is plainly not a substitute for checking.

Our read

Teams should treat AI-generated accessibility claims as something to test, not take on trust. Run automated checks, then have people test the actual experience with assistive technology. The unglamorous bit is also the useful bit: a confident code assistant cannot tell you whether a customer can finish the checkout.

What to watch

  • Whether AudioEye publishes the prompts, tools, test cases and issue-level results.
  • Whether independent testers reproduce the findings across different models and website tasks.
  • Whether teams using AI coding tools add accessibility testing to their normal review process.

Discussion spark: Should accessibility testing be a required part of approving AI-generated website code, or is that an unrealistic burden for smaller teams?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.