“It might replace ‘pixel pushing’ jobs, …” - based on current tech it’s not going to do that, as it doesn’t model interactions (context+function+form+situational information), it’s mainly used to model sequencing plausible, not always accurate or relevant, written or image form instances i.e. lots of words or images strung together.
It’s being misused for decision making and consequence “chickens” are eventually coming home to roost.