That's certainly one theory, but as we are largely out of training data there's not much new material to feed in for refinement. Using AI output to train future AI is just going to amplify the existing problems.
I remember when I worked in a data center and there was a custom server order that needed something like 64 sticks per server, and procurement didn't bother to make sure that we had sets that were the same speed, timing, or brand. Thankfully I caught it before we wasted a ton of time troubleshooting.
Imagine how smart you think the average American is. Realize you are probably biasing too high based on your peers, and adjust that down a little. Now realize that >40% of the population is dumber than that.
Being a surgeon is hard work. Grifting is a lot easier and cushier.