How LLMs Are Becoming On-Call Accessibility Auditors for Design Systems

AI · 5 min read

Over the past year design teams at both startups and enterprise organizations have started integrating large language models into their design system workflows to perform lightweight accessibility audits. Connected to design files and component libraries, LLM agents analyze component documentation, ARIA usage, color token combinations, and copy to surface likely problems like low contrast, missing keyboard focus states, or unlabeled controls. The promise is clear: faster discovery of common issues before engineering handoff.

But teams are learning that LLM-generated recommendations must be curated. Models can hallucinate ARIA roles or propose fixes that conflict with platform-specific conventions. Successful pilots pair model suggestions with automated checks — contrast ratios and automated keyboard tests — and require a human-in-the-loop signoff for any change to a canonical component. That hybrid approach reduces noise while keeping accountability in the design system team.

Vendors supplying these integrations are responding with features tuned for accessibility: explainability prompts that show why a fix was suggested, auto-generated acceptance tests, and versioned change proposals that feed into component changelogs. Accessibility leads caution that teams embed guardrails such as testable assertions, provenance on suggestions, and periodic audits to avoid over-reliance on LLMs as sole arbiters of accessibility.