Post

Funny, one of the reasons I left GitHub is because the powers that be think they can "solve" it by automating it away. That completely ignores the "functional but unusable problem." This creates a scenario where bots can operate something and the people that need the accessible experience cannot.

Thank you for the link! I enjoyed the overall thought behind the post, but this part killed the message for me 'it’s easier than ever to implement things in a “correct” and accessible way with AI tools. You can tell your agents to be WCAG-compliant, and it can just do it'

We've been saying for years that WCAG is the floor and it does not guarantee a website is usable by people. 😢 Not to mention AI's technical implementation limitations. Maybe it's not Cassidy's intention but this paragraph makes a11y sound like a fully automatable task, which it isn't.

One of the things I've found is that AI heavily indexes resources like caniuse and mdn. Then assumes that is all that is needed to use an aria pattern. aria-orientation is a big one that it assumes just works but has relatively limited screen reader support. Many popular libs hit this too like shad

I could be too pessimistic, but based on the annual WebAIM survey I think accessibility is getting worse, not better. That might point to a difference between perceived accessibility, as reported by developers, and actual accessibility, as experienced by people with disabilities.

Can Your AI Pass The Accessibility Test? is another good read about this, and emphasises including feedback from disabled users. In the era of "synthetic user research", I do worry that developers might use agents in place of user research with real people. www.aaron-gustafson.com/notebook/can...

aaron-gustafson.comCan Your AI Pass the Accessibility Test? :: Aaron GustafsonLast week at Microsoft Build, Jessie Lorenz, Carie Fisher, and I gave a short talk on a question every AI-enabled product team should be asking: can your AI pass the accessibility test? The core poin...

An idea, if you’re open to it: Let’s chat about this problem and share it out. You’ve got a huge reach, and it could potentially do a ton of good in terms of people better understanding the nuance of the issue.

I'm a bit annoyed by Svelte being the "bad example" in this post: Svelte is probably one of the few frameworks that deeply care about a11y. We warn our users for common a11y traps and while that's not enough to guarantee an accessible website it's a nudge in the right direction. While developing

our docs we tried to make them accessible (and if they are not, it's definitely our fault, and we would love to hear how to make it more accessible). The fact that we are also accessible for LLMs doesn't mean that we don't care for humans too. This specific sentence in the blog post is completely

backwords in my opinion: > When the audience is AI, accommodation is offered with a wink. Beep boop! We don't offer accommodation with a wink for our humans readers because that should be the absolute default. You don't have to visit a `/visual-impaired.txt` route, you should just browse normally!

Yeah this is very rude @jonathanzong.com @frank.computer! It's also nonsensical. I agree that ideally AI systems would navigate the world as it is, but even if llms.txt and semantic HTML were somehow mutually exclusive (which they're obviously not), both exist to ultimately serve end users.

Researchers should take care to evaluate the tools they are speaking against. Using svelte for 5 seconds and you will see it empowers accessibility from its core, and has for some time. Taking 1 line in the svelte docs and turning it into an entire diatribe against svelte is certainly... a choice.

I don't agree with the use of the word diatribe here, and also think it's bypassing the larger point: We extend accommodations to LLMs that we don't to the people who rely on said accommodations. It's the same energy as the forgiveness we extend to LLMs when they screw up compared to humans.