I’ve presented my talk “Test your website’s accessibility starting…NOW!” twice this year at the Nonprofit Technology Conference and the Washington State Nonprofit Conference. Both times, I got a question from an attendee asking about AI and accessibility. My thoughts are brief and straightforward.
AI is a Mirror
“Large Language Models” (LLMs) are what run things like ChatGPT or the coding tool GitHub Copilot that many people, including myself, use. Boiled down, here’s how they work:
- Analyze billions of pages of text, images, code, etc.
- *Complicated math*
- Predict the most likely output (text, image, audio, video, code) in response to an input/prompt/question
Hold that thought as we turn to the WebAIM Million, a report showing us that 95.9% of the top 1,000,000 home pages have at least one accessibility error, and the average home page has 56.8 errors.
Do you see the problem? If you were to predict the “most likely” home page based on all the existing home pages, it will not be accessible because most home pages are not accessible. Garbage in, garbage out. (And I’m not the only person pointing this out.)
We live in an ableist society, so I see no reason to believe an LLM would parrot anything other than ableist outputs. We saw this happen exactly with the release of the v0 project from Vercel that produced code with obvious accessibility problems despite being declared “production ready” by the tool’s maker.
And I’m not exactly hopeful for the future either. AI tools are being built by massive companies that are focused on their bottom lines, not the impact on marginalized communities. (Twitter famously fired their entire accessibility team two years ago.) The only hope for accessible website AI is a tool built by a company that explicitly prioritizes accessibility, and I don’t know of any companies actually doing that.
Everything I just said about ableism and AI equally applies to racism, classism, ageism, sexism, and more. Meenakshi Das recently posted “A call for responsible AI that leads with care and imagination” that gets at these issues more broadly. I think all the things she recommends are right on, but I can’t think of a profit-driven company that is ever going to do those things. I really hope I’m wrong, but I’ll need to see a whole lot of evidence to convince me otherwise.
Where can AI help people with disabilites?
A quick aside: I think AI does have the potential to help people with disabilities in very specific ways when they are asked how they want to be helped and given the power to design and control that technology. I don’t think that happens very often, though.
It’s much more common that able-bodied people assert that AI will “solve” accessibility, as we saw with Jakob Nielsen’s embarrassing post asserting that “Accessibility has failed”. It’s not worth a link, but here’s a good takedown and here’s a list of links to other responses. Accessibility is not a problem to be solved, because humans are not a problem to be solved.