Depends what you mean by “AI”.
Machine learning trained to find cancer can be (and AFAIK is) helpful in identifying early stages of cancer.
Hallucination machines, aka fancy autocorrect, aka LLMs, have no place in any setting where factual accuracy is of utmost importance, such as peoples’ health.
I think the view can be described as two axes: doesn’t work <-> works well, and is not fun <-> is fun. “It’s functional” is hardly a ringing endorsement, so let’s put it just above zero on the “works” axis.
It’s still “not fun” on the “fun” axis, so you’d expect the overall review to be negative as an average of “neutral” and “bad”. Yet, for some reason, OP is finding that these reviews weigh “it works” so heavily that the score is positive, not neutral or negative.
If anything, I would expect most people to weigh “fun” more strongly than “it works” (where “fun” means engaging, not necessarily “happy”)