AI will happily write your alt text, caption your video and draft your copy. All three come out worse for the people who need them most, and for a reason worth understanding.
- AI describes what is in the picture. Screen reader users need to know why it is there. Those are different sentences, and the second one is the useful one.
- The mechanism: the model sees the image alone. It cannot see the headline above it, the caption below it, or what the page is for — which is exactly what a human uses to decide what to say.
- Research found what screen reader users want depends on the context of use — which is information the model never receives, because it is handed the image alone.
- Automated accessibility checks pass anyway. A description that is present and grammatical scores fine and can still be useless.
- Auto-captions drop the words that carry the meaning — names, jargon, anything unusual. Precisely what a viewer cannot guess.
- Image models do not know about contrast and will put pale text on a pale background all day.
Almost nothing written about AI output covers who can actually use it. The output is judged on whether it looks right, which quietly means looks right to someone who can see it.
This is not a compliance page. It is about the gap between output that passes a checker and output a person can use, because AI widened that gap rather than closing it.
Alt text: the wrong sentence, fluently written
Description is not purpose
Ask a model to write alt text and you reliably get a description of contents. "A woman in a blue jacket standing beside a bicycle on a cobbled street." Accurate, well-formed, and frequently the wrong thing to say.
What a screen reader user needs is why the image is on the page. If it illustrates an article about commuting, the useful alt text is about commuting. If it is a link to a product, the useful alt text is where the link goes. If it is decorative, the correct alt text is empty — and a model will almost never tell you that, because it was asked to describe something and it can always describe something.
The reason it cannot do better
The model is handed the image on its own. It cannot see the headline above it, the caption beneath it, the sentence that refers to it, or what the page is trying to do. Those are exactly the things a person uses to decide what to write.
So this is not a quality problem that improves with a better model — it is a context problem. A system given less information than the task requires will fail at the task no matter how well it writes."From Provenance to Aberrations: Image Creator and Screen Reader User Perspectives on Alt Text for AI-Generated Images", CHI 2024 (ACM, doi 10.1145/3613904.3642325). The paper reports that image generation models do not output alt text with their images; that what screen reader users wanted varied with the context of use; and that the prompt used to make an image cannot serve as its alt text. The draft of this page paraphrased it as users valuing "contextual and emotional information above literal accuracy". That wording is not the paper's finding; it was replaced before publication. Checked 28 Aug 2026
And the checker will not catch it
Automated accessibility tools check whether alt text exists. A fluent, irrelevant description passes. An organisation can score well and still be unusable, and the first real signal is people abandoning the task rather than a failed audit.
The annual WebAIM Million scan of the top million home pages found 16.2% of images missing alt text in its 2026 report, down from 18.5% the year before, with just over half of all pages affected. The problem predates AI, and automatic generation moves images out of the missing column without telling you whether what replaced them is any use — which is the column no scan counts.WebAIM Million, 2026 report, webaim.org/projects/million. The missing-alt-text figures are theirs. The claim that automatic generation shifts the failure from missing to unhelpful is this page's reasoning, not a WebAIM finding — they measure presence, not usefulness. Checked 28 Aug 2026
Use AI for a first pass if it saves you time, then ask one question before publishing: does this say why the image is here? If it only says what is in it, rewrite it. If the image is decoration, delete the text entirely.
Captions: the words that go missing are the ones that mattered
Automatic captions fail selectively
Speech recognition is good on ordinary sentences and unreliable on exactly the words a viewer cannot reconstruct: names, place names, product names, technical terms, acronyms, and anything unusual.
This is worse than random error. A caption track that is 95% right can be useless if the missing 5% is every proper noun in the video — the viewer gets fluent text with the load-bearing words wrong, and no signal that anything is wrong.
The fix is small: before generating captions, give the tool a list of the names and terms it will encounter. Most captioning tools accept one. Then read the transcript once, checking only the nouns.
Generated video has a second problem
AI video often has no dialogue to caption but plenty of meaning carried visually. If a clip communicates something the surrounding text does not, that content is unavailable to anyone who cannot see it — and there is nothing for a captioner to transcribe.
The answer is not a caption track. It is making sure the page says in words whatever the video says in pictures, which is good practice for search engines and for people on slow connections too.
AI writing reads worse aloud than it looks
What screen readers expose
Model output tends toward long sentences, stacked subordinate clauses, and subjects that arrive late. On a page this scans as competent. Read aloud at speed, it is hard to hold — by the time the verb arrives the listener has lost the subject.
Three habits fix most of it, and they improve the writing for everyone:
The quickest test costs nothing: have your device read a paragraph back to you. Problems that are invisible on screen are obvious within two sentences of hearing it.
Generated images do not know about contrast
A model optimises for looking good, not being readable
Ask for text in an image and you will often get something tasteful and low-contrast — pale grey on cream, thin type over a busy photograph. It looks designed. It is also unreadable for a large number of people, and unreadable in sunlight for everyone.
Two rules that avoid most of it:
Do not put important text inside a generated image. Put it in real text on top. It is then selectable, translatable, readable by a screen reader, and you can fix the contrast without regenerating anything.
If text must be in the image, ask for high contrast explicitly and check the result rather than trusting it. Contrast is a measurable thing, and free checkers will tell you in seconds.
The before-you-publish check
Six items. Under two minutes for a normal page.
"From Provenance to Aberrations: Image Creator and Screen Reader User Perspectives on Alt Text for AI-Generated Images", CHI 2024 (ACM, doi 10.1145/3613904.3642325) · WebAIM Million, 2026 report. Checked 28 August 2026.
Two things this page deliberately does not do. It does not give a percentage for how often automatic alt text is wrong — the figure varies by methodology and by what counts as wrong, and a precise number here would be false confidence. And it does not claim to be a full accessibility guide: this covers the specific ways AI-generated output fails, not the standard, which is much larger and maintained by people who do only that.
The through-line: AI produces output that passes automated checks and fails people. Those are not the same test, and only one of them has a person on the other end.