Blog / AI & search /
A review method for AI-generated alt text
An AI tool can write alt text for hundreds of images in minutes. That’s brilliant until it confidently labels a decorative shape as a product feature, or repeats what the caption next to the image already says. Fluent isn’t the same as useful, and the review has to judge whether the text helps someone understand the page.
Alt text depends on context. The same photograph may need different treatment on a product page, in an editorial article or as a purely decorative banner. Automated suggestions can support a workflow, but whoever publishes the page remains responsible for the result.
Decide whether text is needed
First work out the image’s role. A decorative image that adds no information generally shouldn’t get a long description. In web markup, an appropriate empty alternative lets assistive technology skip decoration, depending on the implementation and context.
An informative image needs an alternative that conveys its relevant purpose. A product photograph might help identify colour, shape or included parts. A chart might show a trend that also needs describing in nearby text. A button image may need wording that explains the action, not its appearance.
Don’t write alt text just because an image file exists. Consider the surrounding content and the task. If a nearby caption already gives the necessary information, repeating it in the alternative makes the experience tedious.
Give the model useful context
A tool can’t reliably infer what the publisher intends from pixels alone. Supply the page topic, the image role, relevant product facts and anything it must not guess. If the image shows a named product or person, give the verified identity only when it belongs in the description.
Ask for concise candidate text and make uncertainty visible. The tool shouldn’t invent materials, dimensions, ingredients, model names, emotions or actions nobody can confirm. For an image it can’t interpret, “needs human description” beats a confident fabrication.
Keep the prompt separate from the final accessibility decision. A detailed instruction helps generate a candidate; it doesn’t validate it. I wouldn’t feed private or customer-sensitive imagery into a service without checking the data rules and permissions first.
Review with a consistent rubric
For each suggestion, ask four questions. Is the image informative in this context? Does the text convey the relevant information or function? Is every factual detail supported by the image or an approved source? Is it concise enough not to burden the reader?
I check names, numbers, direction and relationships carefully. Image recognition can confuse similar objects or miss a small but important label. If text appears inside an image, decide whether it needs transcribing or is already available elsewhere. For complex diagrams, a short alt attribute may not be enough; a longer description next to the graphic can serve better.
Watch for redundant phrases such as “image of” when the context already announces an image, and avoid keyword stuffing. Alt text isn’t a hidden place for search terms. An honest description helps more than unnatural wording aimed at rankings.
Test the published experience
Inspect the actual HTML attribute on the rendered page, not just the content-management field. Themes, page builders and media components can add or override values. Check that an empty alternative is truly empty where intended and that text isn’t duplicated by a caption or nearby label.
Use a screen reader or accessibility review process to sample representative pages, and listen to the sequence in context. Does the alternative add meaning, repeat information or distract from the task? Automated scans catch missing attributes, but they can’t decide whether an existing description is accurate and useful.
Review mobile and responsive variants if the image changes across breakpoints. A crop may leave out details the desktop image showed. If the alt text mentions a feature no longer visible, revise the implementation or the text.
Build a manageable review queue
Prioritise images on important pages and images that carry essential information. Review newly generated suggestions before publishing. For an existing library, sample by template and risk, then expand based on the issues you find. A blank alt field isn’t automatically top priority if the image is decorative; a misleading description can be worse.
Record the reviewer, the date and the source for factual details. Keep original and approved values if the system supports it, so future editors see what changed. Don’t let the queue become a permanent spreadsheet with no owner; build the checks into publishing.
Set a clear rule for exceptions. Complex charts, medical or safety information and identifiable people may need specialist review or permissions. The model shouldn’t bypass those requirements because it returned plausible prose.
Measure quality by usefulness
Track review coverage and recurring error types, not how many suggestions were generated. Common issues include redundant wording, invented detail and confusion between decorative and informative images. Use those findings to improve prompts and editorial guidance, and keep human review.
If the tool saves time, note where the time goes. A fast draft followed by heavy correction may not be a useful workflow. Compare the cost of writing by hand, assisted drafting and template-based rules for simple cases.
Related reading
- When an AI-generated FAQ adds no value to a page
- A review process for AI-generated recommendations
- Testing an AI workflow on a small sample before scaling it
Take a recent page with several images. For each one, can you say what a person using a screen reader needs to know, and verify that the AI suggestion says only that?