Automated alt text has improved enormously, and using it thoughtlessly still produces bad accessibility. The gap is not about model quality — it's about what alt text is for.
What Alt Text Is Actually For
Alt text conveys the function of an image in its context, not an inventory of its contents. The same photograph needs different alt text depending on why it's on the page:
- On a product page: "Walnut desk organizer, three compartments, front view."
- In an article about deforestation: "Walnut harvested from a cleared forest area."
- As a decorative header: an empty alt attribute, so screen readers skip it.
A vision model sees only the pixels. It cannot know which of these situations it's in, and that limitation is structural rather than temporary.
What AI Gets Right
- Object and scene identification — usually accurate for common subjects.
- Colour, count, spatial arrangement — generally reliable.
- Text in images (OCR) — often good, and genuinely useful since text in images is otherwise inaccessible.
- Speed at scale. For a library of ten thousand undescribed images, AI drafts are vastly better than nothing, which is the realistic alternative.
What It Gets Wrong
- Purpose and context. The core problem above.
- Over-description. Models tend to describe everything, producing long alt text where one clause was needed. Screen reader users hear every word.
- Redundancy with surrounding text. If the caption already says it, the alt text repeating it is noise.
- Starting with "an image of" — screen readers already announce that.
- Confident errors. Misidentifying a product, a species, a landmark or a person, stated with no uncertainty.
- Names and identities. A model cannot know that this is *your CEO* or *the customer's product*.
- Charts and data. "A bar chart" is useless; the alt text needs the conclusion or the data, which the model may misread.
- Cultural and sensitive content — descriptions of people's appearance, disability, ethnicity or religious dress require judgement models don't have.
A Workflow That Works
- Generate drafts with AI for the whole library — this solves the coverage problem.
- Triage by importance. Informative images that carry meaning (products, charts, diagrams, people) get human review. Purely decorative images get an empty alt attribute, which no model can decide for you.
- Edit for context and length. Cut to one clear sentence that serves the image's purpose on that page. Remove "image of", remove redundancy with the caption.
- Verify identifications for anything specific — product names, people, places.
- Write chart and diagram alt text by hand, or put the data in surrounding text and mark the image decorative.
- Spot-check with a screen reader. Hearing your own page read aloud reveals problems no checklist catches.
Where Automation Is Genuinely the Right Call
- User-generated content at scale, where the alternative is no description at all.
- Legacy archives being retrofitted.
- First-pass drafts for a human editor.
Treat AI alt text as a starting point that raises the floor, not as a finished answer. The failure mode to avoid is a site that technically has alt text on every image and is still unusable — which is a worse outcome than missing alt text, because it looks compliant.