AI Agents Can Now Catch Most Website Accessibility Problems

A team of researchers has published a new way to find the accessibility problems that make websites hard — or impossible — to use for people who are blind, rely on screen readers, or can't control a mouse. Their system uses AI "agents" (software that can browse a page and take actions on its own) instead of a single fixed checker. They built one specialist agent for each of 40 rules in WCAG, the international standard for accessible websites. Think of it as replacing one general inspector with 40 narrow experts.

The results are striking. Tested against 250 real problems already identified by professional human auditors, the agents found 86% of them. The tool most companies use today, axe-core, found only 36%. A general-purpose AI model looking at screenshots without guidance found 67%. That's a big jump — the difference between catching a third of the barriers on a page and catching nearly all of them.

Why this matters beyond the lab: accessibility is increasingly a legal requirement. Businesses face lawsuits and fines when their sites shut out disabled customers, and professional audits cost real money and take weeks. A tool that approaches expert-level detection could let companies find and fix problems early and cheaply — and could mean fewer frustrating dead ends for the roughly one in six people worldwide living with a disability.

The catch is honesty about accuracy. The agents caught more genuine problems, but they also raised more false alarms, and the researchers warn that some of their measurements are estimates rather than proven facts. They also note that a flagged issue doesn't always mean a real user is actually blocked. So this isn't a replacement for human auditors yet — it's a faster first pass that tells experts where to look.