AI Falls Short as Accessibility Expert, SitePoint Analysis

SitePoint published an analysis arguing that large language models cannot replace trained accessibility auditors. The article, titled "AI Is Not Your Accessibility Expert: What LLMs Still Miss About WCAG," runs through the failure modes that keep surfacing when models handle compliance alone.

Where do LLMs fall short?

WCAG 2.2, the W3C's current recommendation published in October 2023, organizes accessibility into four principles: perceivable, operable, understandable, and robust.

Each principle breaks into guidelines, then into numbered success criteria — entries like 1.4.3 (Contrast Minimum) or 2.4.7 (Focus Visible).

A large language model can pattern-match against fragments of those criteria it has seen in training. What it cannot reliably do, the analysis argues, is apply every criterion to every interaction a real user will encounter.

What kinds of issues slip through?

The piece highlights problem areas that recur in remediation tickets across the industry:

• Color contrast that meets a 4.5:1 ratio at rest but breaks when a hover state shifts the background.
• Form inputs that pass the model's check but lack a programmatically associated label in the rendered DOM.
• Custom widgets — date pickers, comboboxes, disclosure patterns — that need ARIA roles and keyboard handling no static review can verify.
• Cognitive accessibility criteria added in WCAG 2.2, including 2.4.11 "Focus Not Obscured" and 3.3.7 "Redundant Entry," that remain thinly represented in training corpora.

None of these are exotic. They are the everyday fixes accessibility consultants bill against in any organization shipping at scale.

Why does this matter now?

The cost of shipping an inaccessible product has shifted. Lawsuits under the ADA continue to climb, and the European Accessibility Act took effect in June 2025, tightening procurement requirements across the EU.

A confident wrong answer from a chatbot — "this button meets WCAG" — can travel from a code review into production, then into a legal complaint.

The article's central claim is methodological. WCAG is a normative document. Each success criterion carries a normative test, often with exceptions and notes.

What should teams do instead?

The analysis does not recommend abandoning AI. Chatbots accelerate the obvious passes — generating alt text drafts, flagging missing landmarks, suggesting focus styles.

The argument is against using them as the final word. Pair chatbot output with:

• Manual keyboard testing through every flow.
• A screen reader session on each release candidate (NVDA on Windows, VoiceOver on macOS and iOS, TalkBack on Android).
• An automated scanner for the share of WCAG it can detect mechanically.
• A human reviewer for the remainder.

What does the spec look like next?

The W3C has begun work on WCAG 3.0, which moves toward a more nuanced scoring model rather than the binary pass/fail of the 2.x series.

As the standard evolves, the gap between what a model was trained on and what current guidance requires will widen.

Teams that treat AI as a junior assistant rather than an expert will ship fewer defects. Teams that skip the human step will keep finding out about them in court.