Eric1d ago
Here's some thoughts and feelings about needing to vibe code to do accessibility design work. https://ericwbailey.website/published/the-case-for-an-accessibility-designer-vibe-coding-when-all-his-coworkers-are-also-vibe-coding/
The case for an accessibility designer vibe coding when all his coworkers are also vibe coding

Am I letting my own personal beliefs and biases affect the outcome I ultimately want?.

Show threadMark T. Tomczak

@eric I hadn't considered AI for accessibility before... It may plug an unfortunate gap I've seen, which is that developers really, really, really don't want to support a11y in most firms and management doesn't care enough to force the issue.

But if an LLM can reduce the problem of identifying issues to one analysis pass... That's not a small thing. And I'm generally seeing LLMs are better at that than traditional heuristic approaches, because they end up trained on real-world examples of what goes wrong.

Jun 15 at 3:08pmWeb
Show threadEric1d ago
@mark Unfortunately, my experience has been a high degree of wildly incorrect fixes when it comes to a LLM performing autonomously. It takes a high degree of scoped human intervention and verification, much to my boss’ disappointment.
Show threadDrew Mochak1d ago
@eric @mark Very much this--I can get pretty good results out of it, because I (not to toot my own horn,) am a genius. I know precisely how it must function and test it riggorously before submitting If you just say, "here's a thing, fix it" it usually doesn't work the first time. But if you stick with it you'll either get it to a good place or realize your problems were deeper than you thought, which is still useful.
Show threadEric1d ago
@prism @mark It's funny how every domain specialist realizes this. Then you extrapolate and 🫠
Show threadGerard K Cohen1d ago

@eric @mark #Facts The amount of manual effort to get AI to do what everyone thinks it can do is severely understated. And even after all the setup, my experience was that it did not accelerate anything. Output was barely equal, but mostly lagging behind humans. (yes, this was actually measured)

And the reality is, thats "ok?" if everyone agrees and understands the effort and output. It's not magic. It's more artificial than it is intelligent. It's definitely not cheaper.

Show threadMark T. Tomczak1d ago

@eric The approach I've seen work best is AI as an automated reviewer in the pull request workflow. It can false positive and false negative, but in practice the false-pos rate I'm seeing is like 1-5% (and much of that corner case stuff, like "yes you're right but that is autogen code that cannot be fixed at this layer").

It needs human in the loop, but it doesn't get bored in the same way humans do.