The bottom line is that humans "hallucinate" (or just plain make mistakes) also -- and, contrary to what the other commenter says, aren't always aware of doing

The bottom line is that humans "hallucinate" (or just plain make mistakes) also -- and, contrary to what the other commenter says, aren't always aware of doing so (nor are they always honest about this fact). Especially when doing stuff on a repetitive basis, for minimum (or barely higher) wages. In particular it has been established that MTurk-like workers (unless you hire bonafide "elite", expert-level ones, which in general doesn't scale) make lots of and lots of mistakes. It's just a question of what your acceptable noise level is to turn a dollar on whatever task needs to be done. So if for the "dumb" 80 percent of use cases for which a GPT-like system can outperform an average MTurk, on flat, accuracy-per-dollar-spent terms -- that's what the route that most businesses will ultimately take. Am just guessing on the "80 percent" of course, but from all that's been published in recent months (and putting my finger to the wind) -- that's where these systems seem to be heading.