Let an AI model patch your accessibility tree live, and the first thing it will teach you is how it fails. Not occasionally, reliably, in specific, learnable ways: confident but wrong ARIA states, patches that compile but break keyboard flow, fixes that pass an automated audit while making the actual screen reader experience worse.
This talk is not a victory lap for a clever pipeline. It's the failure modes first, then the architecture that exists specifically because of them. We'll look at why static linters structurally can't catch dynamic ARIA bugs, what it actually takes to make an AI-generated accessibility patch safe to apply automatically, and the much larger lesson underneath: the accessibility tree is downstream of component state, not of markup which means most AI-accessibility tooling is analyzing the wrong layer entirely.
You'll leave with a concrete mental model for the gap between markup-level and state-level accessibility analysis, a typed validation pattern for constraining what an AI agent is allowed to touch in your UI, and an honest account of where this approach still falls short because if a talk about AI and accessibility doesn't tell you where it breaks, it hasn't been tested enough yet.
This talk has been presented at React Advanced 2026, check out the latest edition of this React Conference.
