Three accessibility tasks are done badly almost everywhere because they are laborious rather than difficult. That is exactly the profile where a draft plus a check beats a good intention plus no time.
Why these three tasks and not others
Alt text, captions and plain-language rewriting share a shape: there is a large volume of them, each individual item is small, the standard for adequate is well defined, and a wrong one is visible to a person who knows the source. That combination makes drafting safe in a way most tasks are not.
They also map onto specific requirements rather than a general aspiration. The Web Content Accessibility Guidelines, published by the W3C and currently at version 2.2, cover non-text content under success criterion 1.1.1, captions for pre-recorded media under 1.2.2, and reading level under 3.1.5. Naming the criterion matters when you are asking somebody to fund the work.
Nothing here removes the obligation to test with the people affected. Automated and drafted output raises a floor; it does not tell you whether your product is usable by a screen reader user, and no checklist does. Where accessibility interacts with AI interfaces themselves — streaming text and screen readers, for instance — there are unsolved conflicts, covered in accessibility in AI interfaces.
Alt text
Good alt text depends on why the image is on the page, and that is information the model does not have. Supply it, and the quality changes completely.
Write alt text for the attached image.
Context: this image appears in [a blog post about warehouse safety],
immediately after the sentence "[most incidents happen during the
last hour of a shift]".
Rules:
- Convey what a sighted reader gains from the image in that context,
not everything visible in it.
- Under 125 characters.
- Do not begin with "image of" or "photo of".
- Do not describe people's appearance, age or ethnicity unless it is
the point of the image.
- If the image contains text, the text goes in the alt text verbatim.
- If the image is decorative and adds nothing to the sentence above,
say so and recommend empty alt text instead.
Give me two versions: one at 125 characters, one at 40.
The check. Cover the image and read the surrounding paragraph with the alt text in place of it. Does the passage still make sense, and does it make the same point? If the alt text describes something true about the picture that has nothing to do with the argument, it is wrong however accurate it is. That is a check anyone can perform in ten seconds and it catches the dominant failure.
The last rule earns its place: the most common alt-text error is describing decorative images at length, which makes a page slower and more tiring to listen to. Being told an image adds nothing is a useful answer.
Captions and transcripts
Automatic speech recognition produces a draft; captions are the draft plus editing, and the editing is not optional. Recognition quality is good on clear single-speaker audio in a common accent and drops sharply with overlapping speech, background noise, unusual names and specialist vocabulary — which describes most real meetings and most conference talks.
- Supply the vocabulary in advance where the tool allows it: names of speakers, product names, technical terms, place names. This is the single largest quality lever and it is usually ignored.
- Fix the names first. A misspelled speaker name recurs on every line they say and is the most conspicuous error in the file.
- Check every number and unit. Recognition errors concentrate in digits, and a wrong figure in a caption is quoted afterwards as though it were the transcript.
- Add the non-speech information. Speaker changes, laughter, a door slamming, the fact that a demonstration is happening silently. Recognition does not produce these and their absence is what makes a mechanical transcript unusable as captions.
- Check timing on the first and last minute, where drift is worst.
The check. Watch three randomly chosen thirty-second stretches with the sound off, reading only the captions. If you can follow what is happening, they work. If you cannot, no error-rate statistic matters.
Plain-language rewrites
Rewriting institutional prose into something readable is the task with the largest gap between how valuable it is and how often it happens. The trap is that simplification changes meaning, and in the documents most worth simplifying — benefits letters, medical instructions, terms — changed meaning is the whole risk.
Rewrite the passage below in plain language.
Rules:
- Keep every fact, condition, deadline, amount and exception. If
something cannot be simplified without losing a condition, leave
it and flag it.
- Short sentences. Active voice. Address the reader as "you".
- Replace jargon with the ordinary word, but keep any term the
reader will see again on a form or a letter, and explain it once
in brackets the first time.
- Do not add reassurance, encouragement or anything not in the
original.
- Keep the original order of information unless a different order
is clearer, in which case say what you moved.
Then list, separately:
CHANGED MEANING - anywhere your version could be read differently
from the original.
DROPPED - anything in the original that is not in your version.
[passage]
The check. Take the two lists seriously; they are the deliverable as much as the rewrite is. Then run the reverse test: give somebody the plain version and ask them three specific questions the original answers — what is the deadline, what happens if you miss it, who do you contact. If they can answer, it worked. Reading-level scores measure sentence and word length and cannot tell you any of this.
Document structure, which is the invisible one
A screen reader navigates by structure: headings, lists, table headers, link text. A document whose headings are bold text at 18 point has no structure at all, and a listener cannot skim it, jump to a section or tell where they are in it. This is the most common accessibility fault in ordinary office documents and it is entirely invisible to sighted authors, who see a heading and assume there is one.
The remediation is repetitive and rule-governed, which makes it a good candidate for drafting.
Below is the text of a document, with formatting noted in
brackets where I could capture it.
Propose a heading structure: for each line that should be a
heading, give the level and say why. Flag any place where the
visual hierarchy is inconsistent with the logical one.
Then list separately:
- lines that should be a list rather than paragraphs with dashes
- tables whose header row is not marked as a header
- link text that does not say where it goes ("click here",
"read more", a bare URL), with a proposed replacement
- headings that skip a level
Do not rewrite the content. Structure only.
The check. Read only the proposed headings, in order, as a list. If they read as a coherent outline of the document, the structure is right; if they do not, the document’s organisation was the problem rather than its formatting, and no amount of tagging fixes that. Then apply the changes with your word processor’s real heading styles rather than by reformatting text: the point is the underlying markup, and nothing about the appearance proves it is there.
What it is bad at
- Charts, graphs and diagrams. A description of a chart is not accessible; the data is. Provide the underlying table and a one-sentence statement of what the chart shows. Asking a model to read values off a chart image produces plausible numbers that are wrong, which is worse than no alt text — and confident misreading of images is a documented failure mode rather than an occasional slip.
- Images whose meaning is contextual or cultural. A photograph whose point is who is standing next to whom, or an image that is ironic. It will describe the surface.
- Sign language. Automatic interpretation is not a solved problem and a captioned video is not equivalent for people whose first language is a signed one.
- Judging whether something is accessible. It will confidently pronounce a page compliant. Compliance is determined by testing, including with assistive technology and with disabled users.
The rule that has to hold
Nothing produced this way is published without a person checking it against the source. The reason is specific to accessibility rather than general caution: the people who depend on alt text and captions are precisely the people who cannot detect that they are wrong, because they do not have access to the thing being described. Every other kind of content has readers who would notice. This one does not.
A wrong caption is worse than no caption, because it is trusted. Bad alt text is worse than absent alt text for the same reason. The drafting saves the time; the checking is what makes the saving legitimate, and a workflow that drops the check has not made accessibility cheaper, it has made it fictional.
Top comments (0)