Ask HN: What “AI” features can you predict enough to use?

In reading Jason Snell's iOS 16 review to get a sense for what I will eventually have to accept, I was struck by this: "This idea of applying more intelligence to email is an interesting one, but it’s also a bit of a black box: even when the features work as intended, they’re sometimes a bit of a surprise." Programs are helpful when their behavior is predictable, but who can predict the behavior of a billion-parameter model trained on a data set no one truly understands?

Autocorrect was useful when it fixed simple transpositions, e.g. turning "teh" into "the," and minor spelling errors, e.g. "concieved" into "conceived." Apple's increasingly aggressive autocorrect is unpredictable to the point of being harmful. God only knows when it will reach back and change multiple words with no quick way to undo the damage, or switch "its" and "it's" incorrectly.

What if the "singularity" is not robots achieving superhuman intelligence, but unfathomable uselessness?