Lesson 8 of 8 · 9 min read · last verified 2026-08-26
Doing this respectfully
In this lesson you will:
- Produce alt text and captions that are actually useful
- Avoid making decisions about someone's access on their behalf
The previous six lessons were about using these tools yourself. This one is about the other side: you will make and share things that other people have to get at.
That is most readers of this module, including those who will never need any of it personally.
Alt text, and why AI cannot finish the job
Ask an assistant to write alt text and you get a description of what is in the picture. That is a reasonable start and it is not what alt text is.
Alt text conveys what the image is doing here. The same photograph needs different alt text depending on why it was used:
On an article about a flood: “A car submerged to its windows in brown floodwater on a residential street.”
On a page about that car model: “A 2019 Astra, side view, in silver.”
Purely decorative, beside text that already says everything:
alt=""— an empty attribute, so a screen reader skips it rather than announcing noise.
Only you know which situation you are in. So use AI for the first draft and then edit it against the reason you used the image:
Write alt text for this image. It appears in [context], and the point I am making is [point]. Under 125 characters. Do not begin with “an image of”.
That last instruction saves a lot of tidying — screen readers already announce that it is an image.
Do not caption decorative images. Filling every alt with generated
description makes pages noisier, not more accessible. An empty alt is often the
correct, considerate answer.
Captions and transcripts you publish
E9·L5 was about consuming these. Publishing them has a different standard.
Automatic captions are a starting point, not a finished product. Before they go out, fix names, technical terms and anything the transcript garbled — the things L5 identified as failure points.
Uncorrected auto-captions on published material are widely regarded as not meeting accessibility requirements, and they are frequently worse than nothing because they look like provision while conveying something else.
The accessible version cannot be the worse one
The rule that matters most, and the one most often broken with good intentions.
If there is a plain-text version, it must contain the same information as the designed one. If there is a transcript, it must include everything said. If there is an audio description, it must convey what the visuals convey.
Common failures: the accessible version arrives three weeks later. It is a summary rather than the content. It omits the diagrams because they were hard. It exists but nobody links to it.
Each of these communicates the same thing to the person relying on it, and it is not what was intended.
Ask, do not assume
The active principle behind this whole module.
Do not decide what somebody needs. Do not announce that a tool will help them. Do not describe a person’s disability to an assistant to get advice about how to treat them.
Ask what would help, and then do that. People know their own requirements and have usually spent years refining them. The answer is often smaller and more specific than you would have guessed — a document format, a seat position, more notice before a meeting.
Two specifics worth stating:
Someone’s disability is their information. Do not put it into an assistant, or into a document, without their agreement. E4·L8’s paste test applies completely.
Do not use AI to assess or second-guess a request. Somebody says they need something; that is the end of the enquiry, not the start of an investigation.
Build it in, do not bolt it on
Retrofitting access is expensive and always partial. Doing it as you go costs very little: headings in order, alt text as you insert images, captions with the video, colour that is not the only signal, and text you can resize.
Most of it takes seconds at the point of creation and hours afterwards. That is the entire argument.
Try it now (7 minutes)
Take something you published recently with an image in it.
Write alt text using the context prompt, then edit it against why you used the image. If the image was decorative, change it to an empty alt — that is the right answer more often than people expect.
Check your understanding
Recap
Alt text conveys what an image is doing here, so AI can draft it and only you can finish it — and decorative images take an empty alt. Correct auto-captions before publishing. Never let the accessible version be shorter, later or worse. Ask what would help rather than deciding, keep people’s disability out of your tools, and build access in as you go, because retrofitting costs hours where doing it costs seconds. That completes E9.
🗂 3 flashcards from this lesson join your daily review.
Previous: Help with the exchange