As AI increases productivity, it will create new jobs in other sectors, and probably not have a massive impact on employment, just like previous tech. The problem comes when it can also do/learn to do those new jobs more quickly than the displaced humans, then we’ll have an actual problem. (Zvi’s take)
Terry Taos recent interview on math and AI gave a nice concrete case of AI enabling greater divisions of Labor via rigidity. Might be a good general model?
People like to point out that AI won’t be a complement to human labor by pointing to the fact that chess AIs beat computer-human pairs pretty quickly. But human+AI might be competitive for longer (assuming in some sense the same rate of general capabilities improvements) in general tasks than in chess because general tasks have out-of-distribution robustness problems, while chess does not. This means humans will maintain an absolute and comparative advantage at checking AI outputs for a while at least.
From the abstract, this looks about right: https://arxiv.org/abs/2312.05481