AI-drafted alt text: review the image’s job on the page

Review AI-drafted alt text by asking what an image does in its particular page. A creature portrait, an image-only action button and a decorative flourish need different treatment. A fluent description of visible pixels can still fail if it leaves the reader unsure what the control will do or repeats information already available beside it.

Begin with the page, not the picture file

An image generator can deliver an attractive asset without explaining its job in a product. The same exported image might appear as a large editorial illustration, a small linked card or a decoration beside an existing heading. Treat those placements as separate editorial decisions. Do not approve one description and automatically attach it to every use.

W3C's Images Tutorial distinguishes informative, decorative and functional images. The central practical point is purpose: useful text alternatives convey the information or function that matters in context. That supports a review process, rather than a demand to narrate every colour, reflection and background leaf in an image.

This article concerns HTML images on game websites and companion web interfaces. It does not claim that writing an alt attribute makes an in-game interface accessible. Nor does an automated description establish accessibility certification. The QORPO studio publishes development exercises to make these distinctions concrete.

Give the AI the surrounding task

If you use AI to propose wording, provide the visible heading, nearby copy, destination and intended interaction. Ask for a short candidate and the reason for it. Then check the actual rendered result. A model that sees only the picture cannot know whether the image opens a creature profile or merely decorates a section already named in text.

Keep uncertainty visible. If a generated illustration suggests a battle that never occurred, the description should not promote it into evidence of gameplay. Our article images carry concept-art captions for that reason. The text alternative and caption have related jobs, but the caption can explain provenance without forcing every use of the image to repeat a lengthy production note.

Original example: one creature, three placements

Imagine a fictional creature website with a teal animal portrait. This is a proposed review example, not a description of a shipped QORPO control. In a species article, the useful information might be the creature's distinctive armoured shell. The nearby heading already supplies its name. Decide what the image adds to that paragraph before writing the alternative.

Now imagine the portrait is the only content inside a link to the species profile. A description such as “teal animal on a pale background” does not tell the reader the destination. Review the link's accessible name as a whole. The user needs enough information to choose the profile, not a guessing exercise about what clicking the picture might do.

In a third placement, the same portrait appears beside an already complete linked species name inside one card. Repeating the name and every visual detail may make the card cumbersome. Inspect the resulting name and reading order before deciding what the image contributes. The correct decision comes from the complete component, not from a rule applied to the asset library in isolation.

Listen to one complete card

An original acceptance exercise starts with a small set: the article hero, one listing card and one image-based control. Use a browser and an available screen reader to reach each in its normal reading order. Record what is announced, what the user can activate and where activation leads. If that setup is unavailable, record the missing validation instead of saying the experience passed.

Listen for repetition as well as omission. A card can announce a heading twice, read an unhelpful filename or describe a scene while never naming its destination. Those are different observations. Write the actual announcement in the report and include the relevant markup so a reviewer can locate the cause.

Keep this check separate from the keyboard route. Our dialog focus exercise addresses where keyboard interaction continues after closing a panel. Both checks may belong in the same review, but success in one does not establish success in the other.

Review edits in their final context

A useful edit can be small: replace an ornamental description with the intended action, remove a redundant phrase or add the missing information that the illustration supplies. After the change, return to the same component and repeat the same observation. Do not compare the revised hero with a different listing card and call that a before-and-after test.

Keep a simple editorial record linking the page, component, image and approved wording. Generated assets often receive descriptive working filenames, but filenames are not a substitute for a considered text alternative. If the image later changes, that record gives the next reviewer a reason to revisit the wording.

Frequently asked questions

Should alt text contain every SEO keyword?

No. It should serve the image's purpose for the reader. A keyword list can obscure that purpose and make a simple link much harder to understand. Our blog uses clear headings and article copy to establish topics.

Can AI make the first draft?

Yes, as a proposal. Give it context, check for invented details and review the result in the rendered component. Do not treat confidence or grammatical polish as evidence that the alternative works.

What should happen to decorative images?

W3C describes an empty text alternative for purely decorative images. First establish that the image really adds no needed information or function in that placement. A functional control is a different case.

Explore QORPO's live studio work and apply this exercise to one real page before attempting a bulk rewrite. One understood component is a better starting point than hundreds of unreviewed descriptions.

Sources: W3C Images Tutorial and W3C alt decision tree, checked 8 November 2026. All interface examples above are original proposed exercises; no screen-reader pass is claimed.