LANDMARK · 7 MIN · JUDGE
Accessibility impacts
After this landmark, you can judge an AI feature by what it does for the people it works worst for, not only by what it does for the average user.
AI has genuinely widened access. Speech-to-text gives live captions to someone who is deaf or hard of hearing. Image description gives a blind reader a sense of a photo nobody bothered to write alt text for. Plain-language rewriting helps someone with a cognitive disability get through a dense form. These are real gains, and dismissing them because the technology is imperfect costs actual people actual independence. But the same systems fail unevenly, and they fail quietly. Speech recognition is measurably worse on accented, atypical, and disordered speech — exactly the voices most likely to need it. Auto-generated captions state wrong words with the same confidence as right ones. Auto-generated alt text describes what a picture contains without saying what it is *for*. The judgment this landmark asks for is not “is AI good for accessibility?” but “good for whom, failing for whom, and who finds out when it’s wrong?”
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
Use these tools freely for yourself — captions, described images, simplified text. When you produce something other people rely on, don’t ship the machine’s first pass unchecked: read auto-captions on a video you publish, and rewrite alt text so it says why the image is there, not just what is in it.
MAKE A DECISION
Your organisation is about to rely on automatic captions for recorded all-hands meetings, replacing the paid captioning service. What is the responsible call?
CARRY THISTake one AI-generated accessibility output you already publish — captions, alt text, a plain-language summary. Check it against the source and note what a reader relying on it alone would have got wrong.