Answers are grounded in this paper and cite their sources.
Cognitive accessibility issues like clutter, conflicting calls to action, and complicated workflows may be seen on screens but cannot be detected by markup-based checkers. Hence, we created a Chrome Extension that uses a Vision Language Model (VLM) to drive the browser via a screenshot-action loop. It then evaluates each screen based on a 46-item cognitive accessibility rubric. Ten VLMs were used from Amazon Bedrock to audit three seeded websites. The models were allowed to navigate freely. Kimi K3 and the three Claude models (Opus, Sonnet and Haiku 5.5) were the most successful across all three sites. They identified 71-100% of the violations. The Claude models completed the audits in one-third of the time or less than Kimi K3. On the two realistic sites, the leading models had a precision of 70-83%. Most models are limited by their ability to navigate, and the top performers are distinguished by their perceptual abilities. An auditor such as this could provide a first-pass compliance check prior to usability testing with end users.
No takes yet. Share an insight, caveat, or question.
Lobo et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: