AI Described the Picture. Did It Tell a Blind Reader What Matters?

AI can describe an image in a second. A blind reader needs more than a list of visible objects: they need the information or function that the image contributes to the page. The useful test for an AI-written description is whether it lets someone understand or act without seeing the image.

The Web Accessibility Initiative at W3C explains that text alternatives depend on an image's purpose. Its image tutorial distinguishes informative, decorative and functional images. An informative photograph, a decorative background and an image used as a button need different treatment. An AI model that sees only the pixels may not know that purpose, so a human author must supply the page context.

What is this image doing here?

Before asking AI for alternative text, answer that question yourself. If the image is decoration and adds no information, an empty alt attribute can let assistive technology skip it. If it conveys a fact, describe that fact. If it is a link or control, describe the action. W3C's guidance treats those as distinct cases, which is why a universal instruction to describe everything is inadequate.

Imagine a photograph of a person holding a phone beside an article about accessible banking. A literal description saying only that a person holds a phone may be accurate but unhelpful if the picture illustrates a particular interaction. Conversely, inventing that the person is using a banking app when the screen is unreadable would be misleading. The surrounding caption and article determine what can be stated honestly.

Give AI the missing context, then verify

Tell the tool what the page is about, what the image adds and whether it contains readable text or a control. Ask for a short draft. Then compare that draft with the actual image and page. Remove guessed identities, emotions, brands and events. If a chart carries data, provide the data or a meaningful summary outside the image; a vague sentence about coloured bars is not enough.

The W3C easy checks say alternative text should convey an image's purpose. A useful human test is to read the page with the image hidden and the alt text spoken aloud. Does the page still make sense? Does the description repeat nearby text without adding anything? Those questions matter more than whether the sentence sounds elegant.

Do not overload the description

A long catalogue of colours, clothing and background objects can make the main point harder to find. If a detail affects the meaning, keep it. If it does not, leave it out. This is an editorial choice about the page, not a competition to identify every pixel.

For complex diagrams, one short alt sentence may introduce the graphic while a nearby text explanation gives the important relationships and figures. That gives all readers a way to inspect the underlying information. It also avoids hiding essential content inside an opaque image.

Check the description with the people it serves

Automated accessibility checks can tell you whether an alt attribute exists. They cannot always judge whether it communicates the right meaning. Where possible, ask screen-reader users to try important pages and describe what they understood. Their feedback may reveal that a technically valid description is confusing in context.

Do not assume one review covers every later use of an image. The same photo can be decorative on a homepage, informative in a news story and functional as a link on a catalogue card. The correct alternative can change with that role.

A three-part edit

Consider a chart showing a rise over several years. An AI draft might list a blue line and a white background. That describes appearance but leaves out the point. If the chart is central to the article, the text near it should name the measure, period and main pattern, with the underlying values available where practical. Now consider a decorative photograph next to a heading that already says everything important. Repeating the heading in alt text adds noise. The same vision model may describe both images confidently; only the page author knows which role each image plays. This example is hypothetical and illustrates an editorial choice, not a measured test of a tool.

After AI drafts alt text, check accuracy, purpose and brevity. Accuracy asks whether every asserted detail is visible or otherwise supported. Purpose asks whether the reader gets the information or action the image adds. Brevity asks whether extra words obscure that purpose. If any answer fails, revise it or use an empty alt attribute when the image is purely decorative.

AI is valuable as a fast first observer. The author still decides what a reader needs to know. Good alternative text is not a picture inventory; it is a small piece of the page's meaning delivered in words.