[overfeed]

> I'd argue back that LLMs likely have a better understanding of a11y conventions than I do as well.

Only because you've not invested time[1] to learn a11y, but that is par for the course: LLMs are pretty good at impressing people without a deeper understanding of a field with its statistically average outputs.

My hot-take: one shouldn't oversee an LLM task whose output they are unable to evaluate.

1. And you don't have to! We have limited time, accessibility is broad, there are numerous standards, and screen readers can be unergonomic, to put it lightly. But a human has common sense, and can care. I doubt there's enough a11y training data to match a caring human testing and thinking through interactions. Sadly, human software QA is now threatened with extinction.