Large Language Models Fine‑Tuned for Accessibility Audits Pass Human Review in Early Trials

AI · 5 min read

Several consultancies and in‑house accessibility teams have begun fine‑tuning large language models on corpora of annotated accessibility reports, issue templates, and remediation patterns. When given HTML, component code, or screenshots with contextual notes, these models produce audits identifying missing landmarks, improper ARIA use, and potential keyboard navigation problems.

In blind evaluations, model-generated audits matched senior accessibility engineers' findings about 80–90% of the time for straightforward issues. Where models excelled was in triage: prioritizing issues by impact, suggesting typical remediation steps, and linking to code snippets. This speeds up the first pass, letting human experts focus on complex interactions and platform-specific quirks.

But the models still produce false positives and occasionally recommend inaccessible 'fixes' that misuse ARIA or rely on fragile scripts. Accessibility leads stress that LLMs should be used as assistive tools—improving throughput and consistency—rather than as autonomous auditors.

Organizations are building guarded workflows: automated audits feed PR comments and flagged issues, but each suggested fix requires an engineer or accessibility specialist to confirm. This hybrid approach is becoming the pragmatic standard for scaling audits across many products without sacrificing quality.