More crudely, GPTs make it easier to produce disposable crap automatically by applying statistical data compression to large training sets.

They're purely mechanical and have no concept of semantics or subjectivity.

So if you ask GPT to write code, you may get something that mostly works, or you may get garbage. And if you're lucky enough to get something that mostly works you'll still need to refine and test it manually.

Human coding has the same perpetually-almost-there issue. But humans can understand a specification and also understand unstated requirements. So the almost-there is more likely to be fit for purpose.

Humans can also take the initiative with new classes of problems.

The real danger is that AI-ifying everything will cripple that imaginative ability, because we'll all be settling for AI worse-is-better instead of aiming higher.