KIKOMONO

AI TOOLS / BEFORE YOU PUBLISH

Check AI-generated alt text before you publish it

AI Toolbox

A vision model can draft alt text in seconds, and the draft usually reads well. But it describes pixels, not purpose: it cannot tell what the image is for on your page, it can invent a name, a number or a word that is not in the picture, and it will happily describe a decorative divider that should stay silent. WCAG 1.1.1 asks for a text alternative that serves the equivalent purpose of the image — and only a human who knows the page can judge that. Run the five-minute check below before the image goes live.

Check AI-generated alt text before you publish it: Function first; No invented details; Text inside the image; Decoration stays silent

The short answer

Alt text has one job: give a person who cannot see the image the same information the image gives everyone else. AI is good at the first draft of that job and unreliable at the rest. It sees objects, colours and scenes; it does not know whether the picture is the product, a link, a chart, or decoration. So treat every generated draft as unreviewed copy, not as a finished accessibility fix. The check below takes about five minutes per page, and it catches the four mistakes that matter: wrong function, invented details, text inside the image that was never transcribed, and decoration described instead of silenced.

1. What the rule actually requires — checked 2026-11-17

WCAG 2.2 Success Criterion 1.1.1, Non-text Content, is Level A — the minimum level. It states: “All non-text content that is presented to the user has a text alternative that serves the equivalent purpose, except for the situations listed below.” The key phrase is equivalent purpose, not equivalent description. The same photograph needs different alt text on a product page, in a news story, and as a link to a gallery, because its purpose changes. W3C: Understanding Success Criterion 1.1.1.

The criterion also names the exception that AI drafts get wrong most often: if non-text content is pure decoration, is used only for visual formatting, or is not presented to users, it should be implemented in a way that assistive technology can ignore it. In HTML that means an empty alt attribute, alt="". A screen reader then skips the image instead of announcing a description nobody asked for. A generator that describes every image it sees is not being thorough; it is adding noise.

2. Where AI drafts go wrong

Four failure modes cover almost everything a review catches. None of them is a reason to stop using AI for the first draft — each is a reason to check it.

It describes the scene, not the function. A photo of a person clicking a button in a screenshot becomes “person using a laptop” when the image is really the “Save settings” step of a tutorial. The pixels are described correctly and the purpose is lost. Function decides the text, not the scene.

It invents details. A fluent draft can name a brand, a place, a number or a word that is not in the image at all. Models predict plausible words; they do not verify facts. Names, figures and any text the model claims to read inside the picture are the first things to check, because they are the claims a reader cannot verify by listening.

It misses text inside the image. Infographics, screenshots, charts and quote graphics carry their meaning as baked-in text. A draft that describes the layout and skips the words has described the frame and lost the content. If the words matter, they belong in the alt text or in real text next to the image.

It describes decoration. Dividers, background flourishes, spacer graphics and repeated icons get full-sentence descriptions from a generator that was never told the image is decorative. Every such sentence is a small interruption for a screen-reader user, multiplied across the page.

3. The W3C decision tree in sixty seconds

The W3C Web Accessibility Initiative publishes a short decision tree for exactly this judgement, and it is the fastest way to make the call consistently, image by image. Walk it in order. W3C WAI: An alt Decision Tree.

  1. Does the image contain text? If the text is also present nearby as real text, or is only there for visual effect, use an empty alt. If the image is a link or button, the alt should communicate its function. If the text appears nowhere else, include it in the alt.
  2. Is the image a link or a button? If removing it would make the link or button hard to understand, the alt should name the destination or the action — “Smallville Times home page”, not “newspaper front page”.
  3. Does the image contribute meaning to this page? A simple graphic or photograph gets a brief description that conveys that meaning. A graph or complex image needs its information elsewhere on the page, in text or a table. If the image only repeats real text nearby, use an empty alt.
  4. Is it purely decorative? Use an empty alt attribute. If none of the questions above gives a clear answer, the decision tree itself says it does not cover all cases — read the full Images Tutorial for that image rather than guessing.

4. The five-minute check before you publish

  1. Decide decorative or informative first. Before reading any draft, ask what the image does on this page. If removing it would change nothing, it is decorative: empty alt, no description, no matter how good the draft sounds.
  2. Read the draft with the image hidden. Cover the picture and read only the alt text. Does a listener learn what the image was for? If the sentence could describe a hundred other images, it is too generic; rewrite it for this image on this page.
  3. Verify every name, number and quoted word. Check each specific claim in the draft against the pixels. If the image contains words, read them letter by letter — prices, dates, product names and web addresses included. A wrong digit in alt text is worse than a missing one, because the listener has no image to correct it against.
  4. Cut the filler. Remove “image of”, “picture of” and keyword stuffing; screen readers already announce that it is an image. One sentence is enough for most informative images. Match the alt text to the language of the page.
  5. Check functional images for the action. Buttons, linked thumbnails and icons need alt text that names what happens when you use them, not what they look like. A linked logo says where the link goes; a search icon says “Search”.
  6. For charts and diagrams, move the data out of the image. Alt text should carry the takeaway — “Sales rose from June to August, peaking in August” — while the figures live in a table or nearby text that everyone can read, search and copy. A long alt paragraph is a poor substitute for a data table.

One more habit is worth keeping: check the exported page, not just the editor. Alt text can be lost when an image is re-uploaded, replaced, or moved between a CMS and a static export. Open the published page, inspect the image, and confirm the alt attribute is the one you approved.

5. What good looks like

Four ordinary cases, each decided by purpose rather than appearance. The wording is illustrative; the rule behind each one is the decision tree above.

Alt text is one of the cheapest accessibility fixes on a page, and AI has made the first draft nearly free. The cost that remains is the review — five minutes, one image at a time, judged against what the image is for. A draft that passes that check is worth publishing. One that was never checked is a guess with a confident tone.

Two companion checks in this series cover the image itself rather than its description: check an AI image for a watermark before publishing it, and how free AI image credits actually work. Our free AI tools for images guide covers which tools generate the images these checks apply to.

Where the facts in this guide come from

Every rule above was read on a W3C page on 2026-11-17: WCAG 2.2 Success Criterion 1.1.1, Non-text Content (Level A — all non-text content presented to the user has a text alternative that serves the equivalent purpose, with the listed exceptions including pure decoration implemented so assistive technology can ignore it), and the W3C Web Accessibility Initiative alt decision tree (image text, link and button function, informative versus complex images, and pure decoration). The four failure modes are general, observable behaviour of vision-model drafts, not measured error rates — no study figures are quoted here, and no tool is named as better or worse than another. If your site must meet a legal accessibility standard, the standard and your jurisdiction's rules govern, not this guide. This guide has no affiliate links — if a link ever earns anything, we will say so in the sentence that carries it. For the companion guides, see Check AI image watermark before publishing, Free AI image credits and Free AI tools for images.

Frequently asked questions

Does AI-generated alt text meet WCAG 1.1.1 by itself?

No. The criterion requires a text alternative that serves the equivalent purpose of the image, and a generated draft is only a starting point. The draft can describe the wrong thing, invent details, or miss the image's function on the page. Human review decides whether the alternative is equivalent; the generator cannot make that call for you.

Should decorative images have alt text?

No. WCAG 1.1.1 excepts pure decoration: it should be implemented in a way that assistive technology can ignore it, which in HTML means an empty alt attribute, alt="". Describing a decorative divider or background flourish adds noise for screen-reader users instead of information.

How long should alt text be?

As short as the image's purpose allows. One sentence covers most informative images. If the image is a chart or diagram, do not stretch the alt text into a paragraph — put the data in nearby text or a table and let the alt text carry the takeaway. Length follows purpose, not a fixed word count.

What if the image itself contains text?

Follow the W3C decision tree: if the text appears nowhere else on the page, include it in the alt attribute. If the same text is already present nearby as real text, use an empty alt attribute. If the image is a link or button, the alt text should name the destination or action, not describe the picture.

Do I need to check alt text my CMS generated automatically?

Yes. The same check applies no matter which tool wrote the draft — a built-in CMS generator, a chatbot, or a dedicated alt-text tool. The failure modes are the same: wrong function, invented details, missed text inside the image, and decoration described instead of silenced.