Authoring

Automatic alternative text still needs an editor

Image recognition can identify visible objects and draft a description. It does not know why the author chose the image, what the surrounding text already says, or what the reader needs to do.

Purpose comes from context

The same photograph can identify a person, illustrate an event, support a purchasing decision, or decorate a page. Each purpose calls for a different treatment. A model looking only at pixels cannot determine whether the image should have a concise alternative, a longer explanation, or an empty alternative because nearby text already provides the information.

Pixels do not contain the publishing decision
Pixels do not contain the publishing decisionA useful alternative depends on the image, surrounding content, reader task, and information that would otherwise be lost.

Generated text can be confidently wrong

Automatic descriptions can misidentify people, objects, text, relationships, and charts. They can also add irrelevant visual detail while missing the point of the image. Treat the output as untrusted draft text and avoid publishing it silently at scale.

Use automation for routing and first drafts

Automation can detect missing alternatives, flag filename-like wording, group images for review, extract embedded text, and suggest a draft. Keep the image and page context visible to the reviewer. Record whether the final alternative was accepted, edited, replaced, or marked decorative.

Ask four editorial questions

  1. What information or function would be lost without the image?
  2. Does the surrounding text already provide it?
  3. Can a short alternative communicate it, or is a longer treatment needed?
  4. Does the final wording remain accurate without speculation?

Sources