Humans have "skin in the game". If they don't do "good enough", they could be reprimanded (ego-hurt), out of a job (stressful), or even in prison or dead (existential threat). This is not foolproof, but it's at least something. There is no incentive for AI to be correct, nor any incentive for it to become more correct.

To wit: I asked ChatGPT to write me a function to generate Hanukkah dates. It spewed a couple of facts about Hanukkah, and the generated function just returned the 8 days following Christmas. I said it was wrong, and it said "you're right!" and then gave me another wrong function (always returning March 25th, since it's actually in the 3rd month of the Jewish calendar, duh). I told it that it was wrong again and asked specifically for a Jewish to Gregorian calendar conversion function, and it gave me some more code that I have no idea is correct or not. So I asked for a table of 10 test cases of Jewish dates and their corresponding Gregorian dates, and not a single one was valid.

I actually have worked with some remote contractors who pull this kind of crap, just throwing garbage over the wall repeatedly until I give up because I just don't want to deal with them anymore. And no, they never get any better. They are not junior devs you can mentor or train or whatever. Their whole model of work is to generate something complicated enough to seem plausible and then bullshit you into accepting it so they can get paid. I have to inspect every line of code they write and do all the diligence myself anyway, and this is more time-consuming and less fun than just doing it myself in the first place. At best it's a nerd-snipe motivation in a "wrong answer on the internet" kind of way.

Strong No Hire. And if you hire them and stick me on code-review duty, I quit.