A friend of a friend — I'll call her Jo, because I've not asked her permission to write this and I'd rather be careful — was diagnosed with motor neurone disease three years ago. One of the things a speech therapist told her early on, at a point when she could still speak more or less normally, was to bank her voice. Sit down and record a few hundred phrases so that when the disease takes her speech, the device she'll use to talk sounds like her.
Ten years ago that meant hours of recording in a clinical setting and a synthetic voice that was recognisably hers in the way a photocopy of a photograph is recognisably you. Now it takes a fraction of the recording and the result is close enough that her son, on the phone, hears his mum.
I've thought about that a lot. Not because it's uplifting — it isn't, particularly, it's a terrible illness and a good tool doesn't change that — but because it's the clearest example I know of AI doing something that could not be done any other way, for people who were not the target market of anything.
The unglamorous version, which is most of it
Voice banking is the story that makes people put their glass down at a party. The everyday version matters more, because it affects far more people far more often.
Live captions that are actually accurate, on any call, without arranging anything in advance. Image descriptions generated on the fly so that a blind person browsing a website isn't confronted with forty instances of the word "image." Speech recognition that copes with dysarthria — with a voice that slurs, or shakes, or stops — which mainstream systems were terrible at for years because they were trained on people who sound like newsreaders. Text simplified into plain language for someone with a cognitive disability or a brain injury, on demand, without having to ask a person and feel like a burden.
Every one of those is a thing that used to require either money, advance planning, or asking another human for help. Removing the "asking another human" step is underrated by people who've never had to do it eight times a day.
There's an old idea in design about dropped kerbs — cut into pavements for wheelchair users, and then used by everyone with a pram, a suitcase, a delivery trolley. Most of this is a dropped kerb. Live captions are used by disabled people out of necessity and by everyone else on a train with no headphones. The features arrive because of a small group and stay because of a large one, which is usually how the funding gets justified, and I've made my peace with that.
Where I get uneasy
Here's the part I'd want anyone building this to hear, because it's the failure mode I keep seeing and it's a subtle one.
Automatic accessibility can become an excuse to stop doing accessibility.
A team turns on AI-generated alt text across their site. The audit score goes up. Somebody screenshots the dashboard. And the alt text now says "a person sitting at a desk" for an image whose entire point was that the chart on the screen shows a 40% drop. Technically described. Functionally useless. Before the tool existed, someone might have written a sentence that mattered. Now nobody will, because the box is ticked and the box is what gets measured.
Same with captions. Automatic captions are genuinely good now, in good conditions, in a common accent, on ordinary vocabulary. In a lecture on pharmacology, with a lecturer from Newcastle and a room with an echo, they are not good, and the errors are confident and plausible, which is worse than errors that look like errors. A deaf student who can't tell which words were wrong is in a worse position than one who knows the transcript is unreliable.
So the test I'd apply is simple, and it's not a technical test. Does the disabled person get to check it, correct it, and be believed? A system that generates a description and lets the user say "that's wrong, describe it again, tell me what's on the screen" is a tool. A system that generates a description, publishes it, and offers no recourse is a compliance product that happens to sit between a person and the thing they were trying to reach.
And the obvious thing, which somehow needs saying
Ask them.
The number of accessibility features I've seen designed by people who have never watched a screen reader user navigate their product is not small. It's most of them. An afternoon with four people who actually use this stuff will tell you more than a quarter of internal workshops, and it will spare you shipping something well-meant and infuriating.
Jo's voice works, by the way. She says the synthesised version is a bit too polite — that it doesn't do her swearing properly, which she considers a significant regression. I think that's the most useful piece of product feedback I've heard all year.
