Your alt text passes automated checks. That doesn't mean it's any good.

GitHub’s new plugin for the Accessibility Scanner aims to improve alt text quality, which automated tools often overlook. According to WebAIM’s 2026 WebAIM Million report, 16.2% of images on popular websites lack alt text entirely, while another 10.8% use vague or duplicated descriptions like 'image' or filenames. Standard checkers flag missing alt text but fail to assess whether the provided text is meaningful, as strict quality rules risk false positives that developers may disable. The plugin introduces five default rules to detect unhelpful alt text, such as exact matches for non-descriptive terms, while avoiding over-punishing valid cases like decorative images with empty alt attributes.

The plugin also addresses repetitive alt text, which can confuse screen reader users by repeating identical descriptions. Initially, the tool flagged repeated alt text based on markup order, but this approach incorrectly targeted unrelated images, such as a header and footer logo. The updated rule now considers visual layout, extending runs of repeated alt text only when images are positioned closely on screen. This adjustment ensures the checker aligns with how users actually experience the page, reducing false positives while still identifying problematic repetition.

For more nuanced evaluations, the plugin offers an optional rule that uses a vision model to assess alt text quality in context. This model examines surrounding page elements, such as headings, page titles, and nearby prose, to determine if the alt text provides sufficient detail. For example, 'a smiling person' may suffice in a generic context but fail under a heading naming a specific individual. The model also considers whether an image functions as a link, as the alt text then becomes the link’s accessible name, requiring a description of the destination rather than the image itself.

The plugin’s implementation includes safeguards to balance accuracy and usability. Findings retain real page URLs and HTML for debugging, while sensitive data is redacted before reaching the model or logs. Cost is managed by limiting model calls to scheduled scans rather than every commit, as image-heavy sites can incur significant expenses. The tool’s design emphasizes practical trade-offs, prioritizing reliable checks that developers are likely to enable over overly strict rules that may be disabled.