AGI's uncertain effects on human wages and employment

[funkervogt]

This NBER paper, "Scenarios for the transition to AGI", was just published and contains very fascinating conclusions: https://www.nber.org/system/files/worki ... w32255.pdf

Using different sets of equally plausible assumptions about the capabilities of AGIs and constraints on economic growth, the same economic models led to very different outcomes for economic growth, human wages, and human employment levels. I think their most intriguing insight is that the automation of human labor tasks could, in its early stages, lead to increases in human wages and employment, and then lead to a sudden collapse in the latter two factors once a certain threshold were reached. That collapse could also happen before true AGI was invented.

Put simply, GPT-5 might increase economic growth without being a net destroyer of human jobs. To the contrary, the number of human jobs and the average wages of those jobs might both INCREASE indirectly thanks to GPT-5. The trend would continue with GPT-6. People would prematurely dismiss longstanding fears of machine displacement of human workers. Then, GPT-9 would suddenly reverse the trend, and there would be mass layoffs and pay cuts among human workers. This could happen even if GPT-9 wasn't a "true AGI" and was instead merely a powerful narrow AI.

The study also finds that it's possible human employment levels and pay could RECOVER after a crash caused by AI.

That means our observations about whether AI has been helping or hurting human employment up to the current moment actually tell us nothing about what the trend will be in the future. The future is uncertain.

The paper gets very technical, but I encourage you all to try reading it. It includes other economic scenarios and points that are fascinating, yet I don't have time to summarize them. It's striking to see a study about these topics (and it basically analyzes the prospects of an AGI-driven Singularity) appear in the National Bureau of Economic Research. Ten years ago, their colleagues would have laughed the authors out of academia for this.