AI and ML (and I'm not talking about LLM, but more about those techniques in general) have many actual uses, often when the need is "you have to make a decision quickly, and there's a high tolerance for errors or imprecision".
Your example is a perfect example: it's not as good as a human-generated caption, it can lack context, or be wrong. But it's better than the alternative of having nothing.
No, what I'm saying is that if I had vision issues and had to use a screen reader to use my computer, if I had to choose between
I'd take the latter. Obviously the true solution would be to make sure everyone thinks about accessibility, but come on... Even here it's not always the case and the fediverse is the place where I've seen the most focus on accessibility.
Another domain I'd see is preprocessing (a human will do the actual work) to make some tasks a bit easier or quicker and less repetitive.