Two accessibility vendors published reports this summer arguing that AI is widening the accessibility gap from two directions. Both companies sell accessibility software and services, and both reports rest on their own data. Read the numbers as directional.
AudioEye’s 2026 Digital Accessibility Index, released June 25, scanned 166,457 pages across 6,161 websites in seven industries with its automated scanner, testing against WCAG 2.2 Level A and AA. It found more than 10.3 million issues, an average of 62 per page.
Its AI finding is about where people land. The report says two-thirds of AI search sessions skip the homepage and go straight to interior pages. Those pages averaged more issues than homepages, 62.4 against 56.7. It also says 64 percent of 2025 web accessibility legal claims referenced interior pages. Government sites averaged the fewest issues of any sector, 48 per page.
Deque Systems surveyed 200 engineering leaders who use AI-generated code and AI agents. It reports that 88 percent trust that AI-generated code is accessible, and 64 percent name accessibility as a top driver of rework after release. The survey page doesn’t give respondents’ roles, company sizes, or when the survey ran.
AI is changing where people land and how code gets written. Neither change favors accessibility by default.
Trust in the code, rework after release
Deque’s two numbers sit uneasily together. Most respondents trust that AI-written code is accessible, and most also name accessibility as a top source of rework after release. A prompt that asks for accessible code isn’t a test that the code is accessible.
AudioEye’s finding matters most for public agencies. Under the ADA Title II rule, WCAG 2.1 AA applies to state and local government web content, and every page counts, not just the homepage. If AI search sends residents straight to a permit form or a benefits page, that page is now a front door.
What to do
- Audit the pages people actually land on, not just the ones you prioritized. Check your analytics for traffic from AI search.
- Put automated accessibility checks into the pipeline for AI-assisted code, before release.
- Test the output. Don’t count a request for accessible code as a check.
- Keep human testing with assistive technology for what scanners miss. AudioEye’s own report recommends combining automation with human expertise.
- Ask vendors building for you to show accessibility test results, not a statement that the code is accessible.
- Read vendor accessibility statistics, including these, as directional.
The vendors have an incentive to raise the alarm. That doesn’t mean the alarm is wrong.
Sources