AI Wrote Your Alt Text. A Screen Reader User Still Can't Use the Page.
A city posts a photo of a new wheelchair ramp outside a permit office. The AI-generated alt text reads: people on stairs. Technically, there's a slope in the photo. Functionally, a blind resident just got told the opposite of what's there.
There's a name for this failure mode now, coined by accessibility testers running real-world tasks against AI-written descriptions. They call it functional blindness: alt text that's technically present, sometimes even technically accurate, and still gives a screen reader user nothing they can act on.
1What the Testing Actually Found
WebAIM's 2026 analysis found more than one in four images on popular home pages carry missing, questionable, or repetitive alt text. That's before AI enters the picture. Layer automated generation on top, and the errors change shape rather than disappearing.
Testers fed AI models real government and commercial images and checked the output against what a sighted person would actually need to know. A data chart came back labeled, simply, "image of graph." No axis. No trend. No numbers. The wheelchair ramp became stairs. In one study, screen reader users attempting real navigation tasks with AI-written descriptions completed them at noticeably lower rates than with human-written alternatives, because the model emphasized the wrong details and skipped the ones that actually mattered for the task at hand.
2Why This Happens, Mechanically
An AI model describes what pixels it recognizes. It doesn't know why the image is on the page, what the surrounding text is trying to accomplish, or which detail a resident actually needs to complete a task. A photo of a building entrance means something completely different on a "contact us" page versus an ADA accessibility statement describing which door has a ramp.
A human writing alt text knows the context automatically, because they're the one who put the image on the page in the first place. That context is exactly what gets lost the moment generation happens without a person reviewing the result against the page it's sitting on.
3The Legal Exposure Nobody's Pricing In
Organizations that shipped AI-generated alt text with no human review have already run into settlement costs over it, because WCAG 1.1.1 doesn't ask for text that merely exists. It asks for a text alternative that serves the equivalent purpose of the image. "Image of graph" doesn't serve the purpose of a chart showing a 40% budget increase. A government site that auto-generates alt text and never checks it is one complaint away from discovering that the gap between present and useful is exactly where DOJ and advocacy groups look first.
4Where AI Alt Text Actually Earns Its Keep
None of this means the technology is worthless. A decorative stock photo, a generic icon, a repeated background image, these are exactly the low-stakes cases where a model's best guess and a human's best guess converge, because there's no hidden context to miss. AI drafts also save real time on volume: a media library with thousands of untagged legacy images is a project nobody budgets a human to do one image at a time, and a first-pass draft turns a blank-field problem into an edit-and-approve problem.
The failure mode shows up specifically where context carries the meaning: charts, diagrams, photos tied to a specific claim or process, anything where the "why is this image here" question has an answer that isn't visible in the pixels themselves. Knowing which bucket an image falls into, and not treating every image the same way, is most of the judgment call a reviewer actually needs to make.
5A Second Failure Mode: Over-Description
Functional blindness isn't only about missing detail. Testers also flagged the opposite problem: AI models that generate paragraph-length descriptions for images that need one short sentence, burying the useful detail in filler a screen reader has to read through regardless. A resident trying to confirm a building has a ramp doesn't need four sentences about lighting and architectural style before getting to the one fact that matters. Both failure modes come from the same root cause: a model describing what it sees instead of what the page needs the description to accomplish.
6The Approach That Actually Holds Up
Nobody serious is telling teams to abandon AI here. The accessibility experts who tested this landed on a specific pairing, not a rejection: AI drafts, a person confirms. Models are fast at flagging what's missing across thousands of images at once. They're bad at knowing which detail matters without being told.
AI finds the gaps
A scan across the entire media library surfaces every image missing alt text in minutes, work that used to take a department weeks to even inventory.
AI drafts a starting point
A candidate description an editor can accept, edit, or reject, instead of staring at a blank field for every single image.
A person confirms the purpose
Five seconds to check that the draft matches why the image is on the page. The step AI skips and the step that actually matters.
7Four Questions to Ask Any Vendor
Does a human see the draft before it publishes?
If the description goes live automatically with no review step, you have automation, not a solution to functional blindness.
Can an editor override it in under ten seconds?
If fixing a bad description requires a developer ticket, the review step exists on paper but nobody actually uses it.
Does it treat charts and photos differently from icons and decoration?
A vendor that applies the same model and the same confidence to every image type hasn't thought about where the risk actually concentrates.
Can you see a log of what was accepted, edited, or rejected?
That log is exactly the kind of dated evidence DOJ has started asking for in accessibility settlements, which we cover separately.
Final Takeaway
Unreviewed AI alt text isn't a shortcut. It's a different kind of missing alt text, one that passes an automated scan while still failing the person it was supposed to help. The fix isn't choosing between AI and humans. It's refusing to let either one work alone.
See the review step in action
Alt Text Manager drafts descriptions with AI and holds every one for editor review before it ever publishes, on WordPress and Drupal.
See Alt Text Manager