Alt text that doesn't describe its image

Why it matters

a11y.image-alt finds images with no alt at all, and a11y.image-alt-quality finds the mechanical failures: a file name, the word "image", a paragraph. What neither can do is read "Cake" and ask whether it describes this picture, in this page's context.

A model reading the page around the image can. It is still only reading the markup and the words — it cannot see the image — so it is judging whether the alt text is plausible and useful given everything else on the page, not whether it is accurate. It will miss alt text that is well written and simply wrong.

An opinion, labelled as one, never worse than a warning. It reports the single worst one it found rather than every image, because a list of fifteen is not something anybody works through.

How to fix it

The suggestion beside the finding is the model's rewrite, and here more than anywhere it is a guess: check it against the actual image. Describe what the image shows and why it is on the page, leave out "image of", and use an empty alt="" when it really is decoration.

Examples

These are the rule's own test pages, and what it says about each. They are run on every change to the rule book.

<img src="/cake.jpg" alt="Cake">

aiReview: {"model":"Claude","alt":{"ok":false,"selector":"img","comment":"“Cake” tells a blind visitor almost nothing about a photograph the page is built around","suggestion":"A three-tier wedding cake iced in ivory with sugar peonies"}}

Warning“Cake” tells a blind visitor almost nothing about a photograph the page is built around (Claude's opinion, not a rule.)

fail-one-word

<img src="/cake.jpg" alt="A three-tier wedding cake iced in ivory with sugar peonies">

aiReview: {"model":"Claude","alt":{"ok":true,"comment":"It describes the picture and why it is there"}}

PassedThe alt text describes the images: Claude agrees.

pass-described