Back to Reference
Contents
Reference
816 words
Jun 25, 2026

Image alt-text curation

Author a real text alternative for every published image, so visitors who use a screen reader can understand what each image shows (WCAG 2.1 - 1.1.1 Non-text Content).

Why this exists

The digital accessibility coverage report (/admin/accessibility) found that published images carry essentially no genuine alternative text. The catalogue had no dedicated place to record alt text, so the report could only fall back to the embedded IPTC/XMP caption. This worklist closes that gap: it gives cataloguers and contributors a real place to write and curate a text alternative for each image.

Alternative text here is authored by people. Nothing is ever saved automatically. To speed the work you can optionally ask an AI vision model (through the AHG gateway) for a draft description to start from - you remain the author and must review and edit it before saving. See AI-assisted suggestions below.

Where to find it

Admin → Image alt-text curation (/admin/alt-text). It links across to the Digital accessibility report (/admin/accessibility), which it complements.

What you see

  • Curated alt-text coverage - how many published images now carry a genuine, human-authored text alternative in the working language, out of the total published images.
  • Working language - alternative text is authored per language, so the same image can carry a description in English, Afrikaans, and any other language. Switch the language to curate another language; the worklist then shows the images still missing alt text in that language.
  • Worklist - the published images that still have no curated alternative text in the working language. Each row shows the parent record (linked), the surrogate filename, and - where present - the embedded caption, which you can adapt as a starting point. Type the alternative text and Save.

Writing good alternative text

  • Describe what the image shows, concisely, for someone who cannot see it.
  • Convey the meaning the image carries in its context, not every visual detail.
  • Do not start with "image of" or "picture of" - the screen reader already says it is an image.
  • If an embedded caption is shown, you can adapt it, but make sure it reads as a description rather than a title.

AI-assisted suggestions (optional)

When the AHG AI gateway is configured for this instance, each image in the worklist shows a Suggest alt text button next to Save. Pressing it sends that one image to a vision model through the sanctioned AHG AI gateway (https://ai.theahg.co.za/ai/v1) and drops a draft description into the text box, labelled:

AI-suggested - review and edit before saving.

The draft is a starting point only:

  • It is never saved automatically. Nothing reaches the image_alt_text store until you review the text and press Save yourself, through the same human save path as a hand-written entry. You are the author.
  • Treat every draft critically. A vision model can misread a scene or invent detail it cannot actually see - correct anything wrong and remove anything guessed (names, dates, places).
  • The draft is requested in the working language, so you can draft in English, Afrikaans, or any other language you have selected.

When the button is missing or says "unavailable"

The button only appears when a gateway endpoint and key are configured. If the gateway is unreachable, no vision model is available, or the image file cannot be read, the suggestion simply fails with a calm "suggestion unavailable" message and the box is left untouched - manual curation always still works. The AI assist is a convenience layered on top of the human workflow, never a dependency.

How the image reaches the model

  • AI is reached only through the AHG AI gateway - never a direct GPU node port.
  • One image at a time, with a size cap on what is sent. Raster images (JPEG, PNG, GIF, WebP, BMP) are eligible; large TIFF/JP2 masters are skipped.
  • The request carries the image plus a short instruction to describe what is visibly shown for a screen-reader user, in the working language.

How it counts toward the accessibility report

A published image counts as having a text alternative in the accessibility report if either it has a curated entry here or, as a fallback, an embedded IPTC/XMP caption. Curated entries are the genuine WCAG 1.1.1 signal and are counted directly; the caption is only a fallback. As you curate, the report's Image alternative text area improves.

Scope and safety

  • Only published images appear in the worklist.
  • Saving writes only to the dedicated image_alt_text store. It never changes the original record or the digital object.
  • The optional Suggest alt text assist only reads the image and returns a draft for you to edit; it writes nothing. The single write path remains your Save.
  • Clearing the text removes the entry, which is a legitimate curation action and keeps the coverage figure honest.