CEPR study maps how workers actually use generative AI at work

The CEPR paper cuts through a lot of hand‑wavy talk about “AI replacing jobs” by asking a simpler question: who is actually using genAI right now, and for what? The authors show that adoption is broad but shallow: many occupations touch genAI somewhere, but within a given job family, a minority of workers are the ones consistently using it. That means the near‑term impact of genAI is being driven by pockets of early adopters and power users rather than uniform automation.

For the race to AGI, this evidence matters because it tempers both hype and fears. Even very capable models do not automatically translate into economy‑wide transformation; they have to be integrated into specific tasks, workflows, and incentives. The paper’s finding that task‑level “exposure” scores are poor predictors of actual use also undercuts simplistic forecasts that just map technical feasibility to job loss. Culture, skills, and management decisions shape whether frontier models show up as copilots, invisible infrastructure, or not at all.

Strategically, labs and platforms that want to accelerate adoption need to think less about generic capabilities and more about where in the task graph genAI actually delivers net benefit without adding friction. For policymakers, the data suggests that retraining and diffusion policies targeted at specific occupations and tasks may be more effective than broad, sector‑wide interventions.