Alt text your readers can actually use

A validator that asks “does this image have an alt attribute?” hands a clean report to a file where every image is described as “image”. Origami asks the second question, and reports the answer as a finding of its own: The images do carry a description, but it is the same generic word repeated — “image”, “photo”, a file name. It passes an automated “has a description” check and still tells a reader nothing about what they are missing.

It is the failure mode nobody measures, and it is everywhere: alt text added in bulk to clear a report, alt text inherited from a design tool that wrote out the layer name, alt text copied from the caption already printed under the image. Each of those passes automated validation. Each of them still costs a reader the picture — which is why Origami counts them apart from the images that have nothing at all.

AI writes what is in front of it. A documentary photograph, a diagram whose labels are legible, a chart whose numbers are printed on the chart, a screenshot, a portrait, an image it can recognise as decorative: for these a vision model produces a description a reader can use, and reviewing one is far faster than writing one.

A person writes the rest. The image whose meaning comes from the paragraph beside it. The illustration in a children’s book, where the description is part of the work rather than a label on it. The visual joke. The photograph whose caption contradicts what is in the frame. The image carrying cultural or historical context a model has no way to hold. And any image inside assessed educational material, where the description decides what the student is actually being asked.

So Origami generates where it can and refuses where it cannot. In PDF, the vision pass writes a description per figure and marks every one for review. In EPUB, Origami does not auto-fill alt text at all: it finds every gap, explains what each one costs a reader, and leaves the sentence to you.

A wrong description is worse than none: a screen reader announces it with exactly the confidence of a correct one, and the reader has no way to know.

The tools that come up when you search for an alt-text generator take one image and give back one sentence. A backlist is thousands of images across hundreds of files, and the question is not “what does this picture show”. It is “which of my titles can I sell in the EU, and what has to happen to the rest”.

Origami works on the catalogue. Add your titles to a workspace and run the audit across all of them, or drive the same tools from your own systems over the Origami API. Plans are counted in publications, not in per-image credits, so a title with four hundred figures costs what a title with four costs.

Once the descriptions exist, the file has to say so. accessibilityFeature, accessMode, accessModeSufficient and a plain-language accessibilitySummary in the OPF are what tell a store, a library platform and a reader that this book really does have alternative text.

Retailers already surface those fields. That is what turns alt text from a compliance chore into something a buyer can see before they buy, and it is why the work ends in metadata rather than in an attribute: the ONIX record you send out should say the same thing the file does.