In my experience documentation generation has a lower error rate than code generation, and the costs of errors are lower too.
I'm not really a big fan of AI agents writing features end-to-end, but I can definitely see them updating documentation alongside pull requests.
AI ,because by default only sees the code, in general describes the functionality not the intent behind the code.
Of course, that's what your tests are for: To document your intent, while providing a mechanism by which to warn future developers if your intent is ever violated as the codebase changes. So the information is there. It's just a question of which language you want to read it in.
"Updating docs" seems pointless, though. LLMs can translate in realtime, and presumably LLMs will get better at it with time, so caching the results of older models is not particularly desirable.