Does AI autonomy determine which workers gain from it?
Core argument: Non-autonomous AI, limited to acting as a co-pilot advising workers, benefits the least knowledgeable individuals the most, while autonomous AI, which can also function as a co-worker, primarily benefits the most knowledgeable.
The impact of AI on labor outcomes critically depends on its autonomy. Autonomous AI can function both as a co-worker and as [a co-pilot advising workers] while non-autonomous AI is limited to acting exclusively as a co-pilot. Our main finding is that non-autonomous AI tends to benefit the least knowledgeable individuals the most, in contrast to autonomous AI, which primarily benefits the most knowledgeable individuals. In fact, non-experimental evidence suggests that firms perceive current iterations of generative AI more as a basic co-worker than as a co-pilot. For instance, using high-frequency job posting data from LinkUp, Berger et al. document a marked decline in job postings for white-collar roles requiring low levels of knowledge, such as low-level office jobs, following the introduction of ChatGPT in November 2022. At the same time, job postings for high-knowledge roles, such as executive positions, increased. This trend indicates that firms view generative AI as a technology that complements high-skilled knowledge workers while substituting for their low-skilled counterparts.





