Labour-market effects of machine learning

node

The most consequential question about this technology and the one where the evidence is genuinely, not rhetorically, unsettled. Four serious findings coexist and do not resolve into one another. A randomised deployment to 5,172 customer-support agents raised issues resolved per hour by about 15% on average, with the gain concentrated among less experienced workers — published in the Quarterly Journal of Economics in 2025. Administrative payroll data from the largest United States payroll processor shows workers aged 22–25 in the most AI-exposed occupations with a relative employment decline reported at 13% in the first version of the analysis and 16% in a later one, while older workers in the same occupations held steady. The Budget Lab at Yale, looking at the whole labour market over the 33 months after ChatGPT's release, found occupational and industry mix changing no faster than in earlier technological transitions and no relationship between AI exposure measures and employment. And a task-based macroeconomic estimate puts the total factor productivity gain over ten years below about 0.66%, and argues even that is generous. These are compatible: a large effect on specific tasks, a visible effect at one entry point, no economy-wide signal yet, and a modest aggregate ceiling. They are also compatible with several different futures, and this Journey does not pick one.

A node is not a place. Drawing it on a map would assert something about the world that no stored fact supports.