This article explores the limitations of automated accessibility testing for alt text, highlighting that existing tools often validate presence without ensuring descriptive quality. The authors describe their approach to building a custom plugin for the GitHub Accessibility Scanner that differentiates between objective markup checks and subjective quality suspicions, ultimately integrating LLM-driven insights to flag potential content issues.
Key points
Automated accessibility tools frequently validate the existence of attributes but cannot inherently judge the quality or context of the provided descriptive text.
Categorizing checks into 'objective proofs' versus 'subjective suspicions' helps manage false positives and prevents developers from ignoring automated feedback.
Integrating LLMs into accessibility pipelines can help identify generic or non-descriptive alt text that traditional static analysis ignores.
Successful accessibility tooling should focus on facilitating human review for complex visual content rather than attempting to reach absolute, automated accuracy.
