Use AI to write image descriptions that tell readers what matters

A photograph can show the same objects in two articles and need a different description in each. On a furniture shop’s page, the useful detail might be the chair’s shape and upholstery. In assembly instructions, it might be where a bracket fits. AI becomes more useful when you explain what the picture is helping the reader understand.

Alternative text, usually called alt text, carries an image’s essential information or function for people who cannot access it visually. A screen reader can speak it or a Braille display can present it. W3C’s guidance makes the author responsible for deciding what matters in the image’s particular context. Describing every visible object is rarely the entire task.[1]

Give the assistant the image and its job

Use an assistant that accepts pictures alongside text. For example, Claude’s documented upload route is the plus button, then Add files or photos. JPEG, PNG, GIF and WebP are supported. Attach the image itself: Anthropic says non-PDF documents are read as text, so a picture embedded in a Word file will not be interpreted through that upload.[5]

Add the page title, the nearby paragraph and any caption. Say whether the image is informative, a link, or decoration. Include the intended audience: a general reader may need an unfamiliar object explained, while a specialist may need its precise name. This is a documentation-based workflow, with original textual exercises below; no image-description product test is claimed.

Read the observations before accepting the polished sentence. If the assistant invents a third person, a brand or a legible price, correct that detail explicitly. A beautiful sentence can still preserve the wrong observation. When print is too small, supply a clearer crop or transcribe the relevant wording yourself; Anthropic recommends avoiding low-resolution images.[5]

The same picture can do different work

Imagine our fictional photograph shows a red bicycle beside a bakery window, with two loaves in its front basket. The shop name and opening hours already appear as normal text on the page. No person is visible. These invented details let us compare the editorial decisions without relying on an untested AI output.

The caption and the alt text do not have to compete to tell the whole story. A caption might identify the bakery and credit the photographer. The description can concentrate on the visual detail that the surrounding text has not conveyed. The US government’s authoring guidance similarly advises matching the context and avoiding redundant information.[7]

What the research adds

A 2024 research prototype tested this principle with 12 blind or low-vision people who already used AI description tools. They compared descriptions for six researcher-selected images and two images of their own choosing. Context-aware descriptions were preferred in 76% of comparisons for the six selected images and scored higher for relevance and other qualities. This was a short user study of GPT-4V-based descriptions, not a general accuracy rating for today’s assistants.[4]

There was also a revealing design problem: simply feeding the webpage text into the model could make it summarise the article instead of describing the picture. The researchers filtered relevant context before producing the description. Their error analysis still found invented objects, incorrect counts and mistaken positions across approaches. Supplying context helps establish relevance; the picture must still support the visual claims.[4]

Give a chart somewhere to explain itself

A complicated chart needs more than a compressed sentence. W3C recommends a short description identifying it, together with a longer text equivalent of the information. The longer version can sit in the page’s normal content, where readers can use headings and a data table. This also lets people inspect the values at their own pace.[3]

For an original example, suppose a chart compares 40 repairs in April, 55 in May and 50 in June. Ask AI for a short description and a separate account of the pattern: repairs peaked in May and ended above April. Supply the actual values for the longer description. If the chart says nothing about staffing or demand, neither should the explanation invent a cause.

Put the description where readers can reach it

Copy the approved wording into the image’s alternative-text field in your website or document editor. In current Microsoft 365 desktop guidance, right-clicking the object and choosing View Alt Text opens the editing pane; Picture Format → Alt Text is another route. Pure decoration has a Mark as decorative option. Automatically generated wording should be reviewed in that same pane.[6]

Keep the description with the particular use of the image. Reusing the same file in a product page and a repair guide can require a different description. A small record containing the image filename, page, purpose and approved wording makes that choice easier to preserve during later edits.

Finally, read the page’s nearby text and description together. Does the combination supply the important information once, in a sensible order? If the page is translated, translate its alt text too. The aim is a reader who can understand the image’s contribution and continue with the article, rather than stop to reconstruct what has been left out.[7]

Report an error or suggest a correction