AI is changing accessibility in useful and harmful ways at the same time. On the useful side: draft alt text, captions, speech transcription, real-time scene description, communication aids for non-speaking people. On the harmful side: auto-captions sold as accurate when they're not, generated alt text that's confidently wrong, voice biometrics that lock out people with atypical speech, and overlay products that promise compliance but don't deliver it.
The Overlay Fact Sheet, signed by hundreds of practitioners and disabled users, documents how overlay products typically fail and sometimes make things actively worse. The pattern is consistent: tools are sold to non-disabled buyers, the underlying problems go unfixed, and disabled users end up worse off. Models trained mainly on non-disabled people also work less well for disabled users, sometimes in safety-critical situations.
Treat AI output as a draft, not a finished product. Review generated alt text, correct auto-captions before publishing, and check AI summaries for accuracy. It's a tool to help with accessibility work, not a substitute for doing it.