Alt Text Is Editorial Work

01What You Are Actually Writing

When a sighted reader encounters an image in a book, they do not see raw visual data. They see a photograph of a polling station, or a diagram of the inner ear, or a graph comparing two decades of rainfall — and they understand it in context. The surrounding text has already told them what to look for. The image confirms, extends or complicates that meaning.

Alt text must do the same thing, and this is why writing it is an editorial act rather than a transcription one. The question is never "what is in this photograph?" The question is "what is this photograph for?" A picture of a crowded street means something different in a history of urban planning than it does in a travel memoir. The words that describe it must reflect that difference, or they describe the wrong thing entirely.

This is not a subtle distinction. A functioning alt text for a bar chart in an economics textbook will include the precise values, identify the axes, and state the trend the author drew the reader's attention to — because the image is carrying data. The same bar chart reproduced on a chapter-opening page as visual punctuation might warrant two words, or none at all and a null alt attribute that tells assistive technology to skip it entirely. Marking a decorative image with an empty alt attribute is not laziness; it is the correct call. A screen reader announcing "decorative swirl, page divider, flourish" interrupts the reading experience without adding meaning.

02Where Automation Fails

Image-recognition systems have become genuinely good at labelling the contents of a photograph. They will tell you there is a person, a table, a window, outdoor lighting. What they cannot tell you is which of those things the author cared about, or why the image was placed where it was. A neural network describing a photograph of Marie Curie in her laboratory will likely produce something accurate about a woman in a room with equipment. It will not tell you that the image is there to make the point that women were present in scientific spaces at a time when the text argues they were excluded. That interpretive layer requires someone who has read the chapter.

The failure is not a technical one that better models will eventually fix. It is structural. Automated description operates on pixel data. Alt text operates on editorial intent. Those are different inputs, and no amount of training on image-caption pairs from the open web closes the gap, because web captions are themselves often wrong, context-free, or written by people who also had not read the surrounding text.

Publishers who treat alt text as a metadata field to be auto-populated at the end of the production pipeline are not cutting a corner. They are producing something that will actively mislead a reader who cannot see the image, substituting an irrelevant description for the meaning the author intended. That is a content failure, not a formatting one.

03Who Should Be Writing This

Alt text belongs to the same part of the editorial process as caption writing, figure numbering, and the decision to include an image in the first place. It is the author's territory, or the editor's. A production assistant working from a finished file after the editorial team has moved on cannot reconstruct why a specific photograph was chosen for a specific page. That context evaporates.

The practical implication is that alt text should be drafted when images are commissioned or selected, reviewed with the same attention as any other text, and treated as a mandatory deliverable, not an afterthought. Organisations building born-accessible workflows have been making exactly this argument: the accessible version is not a retrofit but a discipline that changes when decisions are made. Alt text is perhaps the clearest case where the retrofit model simply cannot work, because by the time the book is finished, the editorial knowledge needed to write it well may no longer be in the room.