Alt Text with AI: Review Before You Publish

A vision model drafts alt text in seconds but omits or invents details. How to decide whether an image needs a description and how to review what AI writes.

info Quick Definition

Alt text is the description a screen reader reads in place of an image. With AI, a vision model can write it, but deciding what to describe and how much to say remains a human call, because it depends on the image’s function on the page.

What should alt text say?

Alt text represents the image’s content and function, not only what is visible. WebAIM stresses context: an algorithm can recognize what an image depicts, but not what it means on the page. A photo of a dog in an article about adoption needs a different sentence than the same photo on an insurance page, and sometimes it should be empty.

WCAG success criterion 1.1.1 (level A) requires that all non-text content has a text alternative that serves the equivalent purpose. Pure decoration is an exception. To decide case by case, the W3C decision tree uses simple questions:

Situation	What to do with the alt
The image contains text that already appears as real text nearby	Empty
The image contains text that appears nowhere else	Include that text
It sits inside a link or button and would be hard to understand without it	Describe the destination or the action
It adds meaning and is a simple photo or graphic	A brief description of that meaning
It is a chart or complex information	Put the full information elsewhere on the page
It is purely decorative	Empty

WebAIM sums up the length: usually a few words are enough, and rarely a sentence or two. AI does not decide which row of the table your image falls into; it can only describe what it sees.

What does AI do, and where does it go wrong?

A vision model describes visible things well: objects, colors, people, scenes. It fails on intent and precision:

It adds details that don’t matter and leaves out the one that does for that page.
It invents elements or relationships that aren’t in the image, or assumes data.
It describes a chart with general phrases but without the key figure.

Two 2026 studies show what this looks like. Cardia, Angileri, Poggianti and Leporini compared human alt text with multimodal-model output on STEM images (diagrams, graphs). The models produced fluent text that seemed informative, but fell short of human alt text, especially at conveying formal semantics and essential structural details. Automatic metrics such as BLEU and METEOR also captured only part of the errors that matter for accessibility.

Raees and colleagues studied how 12 disability-services professionals wrote figure descriptions with ChatGPT-4o. About 65% of their interactions were requests to write the text, and they rarely went back to check it against the paper. Raters scored the AI-assisted, human-reviewed text higher than human-only text on accuracy, completeness and clarity, but still found errors the participants had not corrected, and those with less subject knowledge tended to accept the output without evaluating it. The authors conclude that human oversight is necessary.

The practical conclusion: AI saves time on the first draft, but the review is still your job.

How do you review what AI generates?

Before publishing, run three checks:

Function: does the text say what the image does on this page? If removing it wouldn’t change anything, it should be empty.
Accuracy: is everything it says in the image and true? Check names, figures and any text the model read inside the picture.
Brevity: can it be said in one sentence? Remove “image of,” “graphic of” and filler words.

To give the model context, tell it in the prompt where the image will appear and what the reader should understand from it. That avoids an inventory of objects and gets you text closer to the function.

For complex charts, the W3C asks for two parts: a short description that identifies the image (and says where the long one is) and a long description with the key information, available to everyone and not hidden. The long one can live in nearby text, on a linked page, or inside a <figure> with a <figcaption>, with a table if needed. The longdesc attribute is deprecated and WebAIM recommends not using it.

The same logic applies to images you generate yourself. If you produce an illustration for a page, the alt is part of the image design and should be written with the same intent as the rest of the piece. To learn what a vision model can and cannot interpret, see multimodal models. The levels and criteria framework is in WCAG explained.

Example

Illustrative example: a store publishes a photo of a blue backpack. The model proposes “Blue backpack with side pockets, on white background, with adjustable straps and a front zip closure.” The page only sells that backpack and the straps don’t affect the purchase, so the team keeps “Blue 22-litre backpack.” In the product photo carousel, the second image shows the interior, and it gets “Backpack interior with a 15-inch laptop compartment,” because that detail helps the buying decision.

Common mistakes

Publishing generated alt without reading it.
Starting the text with “image of” or “photo of.”
Describing everything visible when the page only needs the function.
Leaving alt empty on an informative image to save time.
Stuffing the alt with keywords that don’t describe the image.
Describing a chart without saying what it shows, or cramming all its data into the alt attribute.
Repeating in the alt the same text that already sits next to the image.

When NOT to use it

When the image holds critical or complex information, such as a medical diagram, an evacuation map or a document with sensitive data. Studies on technical images show the biggest AI gaps right there. Also when the image shows a real person and an automatic description could be inaccurate about their identity. In those cases, someone who knows the content should write the text, with supporting context in the page copy when needed.

Sources

Frequently asked questions

Can AI write alt text for my images?

It can draft it quickly, but it needs human review. A 2026 study on STEM images found fluent descriptions with gaps compared with human-written text, and another study found that accessibility professionals checked ChatGPT’s output very little.

When should an image have empty alt text?

When it is decorative, when the same text already appears nearby as real text, or when the image adds no meaning. Empty alt (alt="") tells screen readers to skip it.

What should alt text avoid?

Phrases like ‘image of’ or ‘graphic of’, since screen readers already announce an image. Also ’link to’ or ‘click here’ for images inside links.

How do I describe a complex chart in alt text?

With a short description that identifies the chart, and the full information in a long description that everyone can see, for example in nearby text or a table. The full data does not fit in the alt attribute.

What does WCAG say about alt text?

Success criterion 1.1.1 (level A) requires that all non-text content has a text alternative that serves the equivalent purpose, with exceptions such as pure decoration.