AI can generate code, write tests, identify some defects, and suggest WCAG-related changes. A growing number of products also promise an automated accessibility audit. The promise is convenient, but digital accessibility cannot be reduced to checking HTML.
AI is a useful assistant and a poor final judge. Even a capable model should not replace a manual audit performed with a keyboard and assistive technologies.
Automated tools see only part of the problem
AI may confirm that an image has an alt attribute without knowing whether “Product photo” communicates the information a blind shopper needs. It may accept several links named “Learn more,” even though a screen-reader user's link list becomes meaningless.
Evaluation tools can accelerate testing, but their output requires human interpretation. See W3C WAI's guidance on selecting evaluation tools.
AI does not understand context reliably
The correct treatment depends on purpose. An image may need a detailed description, a brief label, or an empty alt when it is decorative. A heading can look correct in isolation while disrupting the page hierarchy. A friendly message can still fail to tell someone how to proceed.
Patterns in code are not the same as understanding the task a person is trying to complete.
Focus order is more than keyboard reachability
A focusable element is only a starting point. An auditor must confirm that focus follows a logical sequence, matches the presented content, avoids hidden controls, moves appropriately after a view change, and returns to the trigger after a dialog closes.
These behaviors emerge over time and across components. Static code analysis rarely captures the full interaction.
AI cannot predict every screen reader
Browsers, operating systems, and screen readers form different accessibility stacks. NVDA with Firefox may announce a widget differently from VoiceOver with Safari. The only dependable way to understand the experience is to test representative combinations and listen to the result.
ARIA does not guarantee accessibility
More ARIA is not necessarily better. Incorrect roles and states can override useful native semantics and make an interface harder to use. A div with role="button" still needs keyboard operation, focus behavior, and state communication. A native button already provides much of that contract.
Complete journeys matter
A scanner may pass every screen separately while the overall process remains unusable. An audit must follow registration, checkout, payment, error recovery, and confirmation from beginning to end. It asks whether users know where they are, what changed, and what to do next.
AI can also invent requirements
Language models can cite the wrong success criterion, combine different WCAG versions, or state a best practice as if it were a conformance requirement. Always verify claims against the normative WCAG standard and its supporting guidance.
Where AI is genuinely helpful
AI can speed up preliminary code review, explain criteria, organize findings, draft test cases, identify suspicious patterns, and suggest remediation options. Every result still needs review in the product's real context.
Use AI to expand an auditor's reach—not to remove the auditor. Accessibility is ultimately about whether people can complete real tasks, and that requires human judgment and hands-on testing.