Remediating a semester of course materials without hand-writing two hundred alt texts
The situation
My observation module has ninety images of classroom environments, my reader has thirty-two scanned figures, and the accessibility deadline is not negotiable. Writing alt text by hand is the most time-consuming accessibility task I have, and the honest truth is that most faculty simply do not do it.
Steps
-
Sort images into three bins before generating anything
A spreadsheet or a printed list
Split images into decorative (no informational content), informational (the description carries meaning), and complex (charts, data figures, multi-panel diagrams). Decorative images get empty alt text so screen readers skip them; complex images need a long description or a data table, not alt text.
What you only learn by doing it: Do this first, by hand. AI will never tell you an image is decorative — it will cheerfully describe a divider line, and every one of those is noise a screen-reader user has to sit through. Most faculty find a third of their images are decorative.
-
Give the model the surrounding context, not just the image
An LMS accessibility tool, Word/PowerPoint suggestions, or any multimodal chat
Paste the paragraph the image sits next to and state what the image is doing there. AI sees pixels; it has no access to authorial intent or the image's role, which is the failure mode every accessibility office names first.
What you only learn by doing it: The same photograph of a block corner needs completely different alt text in a chapter on spatial reasoning versus one on conflict resolution. Supplying the purpose is the entire value you add and it takes one sentence. Instructors who skip it get generic descriptions and conclude AI alt text is useless.
-
Edit every draft: strip the preamble, cut to length, verify the facts
Your LMS alt text field
Delete openers like “an image of,” cut to roughly 125 characters for simple images, and check every factual claim. An inaccurate description is worse than none, because a blind user cannot visually verify it.
What you only learn by doing it: Watch hardest for people. AI descriptions guess at age, race, gender and emotional state, and in materials full of photographs of children that is both an accuracy and a bias problem. If the description makes a claim about identity or feelings you cannot confirm, it comes out.
-
Handle complex figures with a data table, never with alt text
An HTML table or a linked long-description page
For charts, ask the model to output the underlying values as a table, verify the numbers against the source, and publish the table adjacent to the figure with short alt text pointing to it.
What you only learn by doing it: Verify the numbers. Extracting values from a chart image is exactly where models fabricate most confidently, and a wrong data table is worse than an undescribed chart — it is authoritative-looking misinformation embedded in your course.
-
Fix auto-captions on the words that actually matter
YouTube Studio, Canvas Studio, or your capture tool
Auto-captions are usually around 90% accurate and the errors cluster in the wrong places: discipline vocabulary, proper names, and anything said by a speaker with an accent the recogniser handles poorly.
What you only learn by doing it: Search the transcript for your key terms rather than scrubbing linearly; that finds most of the damage in a fraction of the time. And expect materially worse accuracy for accented English — in a course with multilingual speakers that gap is not a rounding error.
Where this breaks down
The American Foundation for the Blind's position is that AI should suggest, not author, and that human-written descriptions should remain primary. This workflow is a compromise with reality — most alt text currently does not exist at all — not a best practice, and you should hold it as such.
AI output does not by itself establish compliance. Running an accessibility tool's suggestions and clicking accept produces a green score and can still produce genuinely unusable materials, which is the quiet failure mode here.
Screen-reader users report being able to tell when descriptions are machine-written, and generic description degrades trust in your whole course.
Test with an actual screen reader at least once. Reading alt text with your eyes tells you almost nothing about how it lands in the ear.
Provenance: institutional guidance rather than a published faculty case study. James Madison University's digital accessibility guidance treats AI as “an assistive drafting tool, not the final authority” and names the failure modes used here; the AFB position is from “Beyond Alt Text: Rethinking Visual Description in the Age of AI.” The three-bin sort and context-injection step are our synthesis.