Researchers use AI to measure what job surveys can't — cognitive demands of work at a level of detail unavailable before

A team developed a method to use large language models to score the cognitive content of job tasks — reading thousands of job descriptions and rating them for how much AI could help or hurt workers. This creates a new measurement tool for labor economists studying which occupations face automation risk, at a granularity that traditional surveys can't reach.

Until now, labor economists studying AI's effect on work have relied on surveys and expert judgment — blunt instruments that miss the specific cognitive demands of tasks within a job. This method uses Claude to score 18,796 task statements from the Department of Labor's occupational database and validates the results against six existing AI exposure indices, showing strong agreement. The method separates two distinct dimensions: tasks AI might augment (make faster) versus tasks it might replace entirely. This matters because the difference between 'AI makes you faster at your job' and 'AI replaces your job' is huge for workers and policymakers planning retraining programs — and until now, the tools for measuring that difference at scale were poor. The practical implication: labor economists can now map occupational risk profiles with more precision, which changes what's knowable about which workers face genuine displacement versus which face productivity gains.