Affordances for me, but not for thee
For years, people have tried hard to get websites to build accessibility affordances. Now developers are willingly building them for AI.
This is worth sitting with:
“The modern web, originally built for sighted humans using browsers, is now being redesigned for a new kind of user.
What these developers are offering their AI visitors is essentially an accessibility accommodation. […] But when the audience is a disabled person, it has historically been treated as an afterthought. Structured, concise text-based representations of complex content are almost exactly the kind of accommodation that blind and low-vision screen reader users have spent decades requesting from web developers, largely in vain.”
One of the oddest parts of the AI shift is that people are much more willing to do things for LLMs that they should have been doing for human beings all along. Accessibility is clearly an important one: 95% of websites have accessibility flaws, and convincing teams to allocate time for accessibility concerns can be like pulling teeth. But now that similar affordances are required for LLM use, people are leaping over themselves to implement them.
The same goes for specifications and documentation. Often, these have been afterthoughts; policies have been hand-waved rather than concretely written down in ways that people can point to. Sometimes it’s even made explicit that this is to preserve manager optionality. But now that LLMs need more concrete instructions in order to behave well, specifications, documents, plans, and policies have rocketed up the priority list.
It would be beautiful if these needs converged, but as the article notes, the affordances needed by screen readers and LLMs are different. Similarly, documentation and planning documents aimed at an LLM are coercive in nature: they’re designed to force the software to do the right thing, rather than to provide background as to why something is the case.
The simple truth is that there is clearly a perception, in some quarters, that there is a stronger productivity gain from doing this work to serve AI than doing it to serve real human people. That’s quite a dystopian idea, particularly as, even if you don’t care about people with disabilities or your own colleagues, doing those things for humans clearly actually has a real benefit. Making your site more usable allows more people to interact with your work and improves your search engine performance. Writing clear documentation and policies allows your colleagues to spend less time figuring out what to do.
But you can’t measure those things neatly. The cause and effect aren’t immediately tethered; managers don’t see a boost they can cleanly ascribe to this work. In contrast, you know pretty instantly whether the AI you’ve trained on your documentation is doing the right thing.
More importantly, whereas accessibility affordances provide new abilities for vulnerable people, an AI affordance provides new abilities for people with power. And that’s probably the heart of it.
