Labour-market effects of machine learning

Assembled from 18 blocks · 5 evidence · 12 related

1. Story
2. 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.
3. Knowledge
4. Labour-market effects of machine learning
5. Connections
6. Machine translation
7. Machine-generated code
8. Measured adoption of AI by firms
9. Payroll data shows a relative decline in entry-level employment in AI-exposed occupations
10. A whole-labour-market analysis finds no economy-wide disruption yet
11. The second International AI Safety Report reports mixed labour findings
12. Radiologist workforce projections and employment outlook, against the 2016 prediction
13. Evidence
14. Studying the staggered introduction of a generative AI conversational assistant across 5,172 customer-support agents, access to AI assistance increased productivity by 15% on average as measured by issues resolved per hour, with effects varying substantially across agents and concentrated among less experienced workers. V55 verification basis: the article was not fetched, but search results carried the journal, volume 140, issue 2, pages 889–942, the 2025 publication and advance-access dates, the 5,172-agent sample and the 15% average effect and attributed them to this paper.
15. Using high-frequency administrative payroll data from the largest United States payroll software provider, workers aged 22–25 in the most AI-exposed occupations show a relative employment decline — reported as 13% in the circulated version and 16% in a later version — since the widespread adoption of generative AI, after controlling for firm-level shocks, while employment for more experienced workers in the same occupations and for workers in less exposed occupations remained stable or grew. V55 verification basis: no version of the paper was fetched. The 13% and 16% figures, the age band, the data source and the controlling strategy were carried in search results attributing them to the Stanford Digital Economy Lab paper; the divergence between the two figures reflects successive versions and is stated here rather than resolved, because the later version was not retrieved.
16. Comparing the pace of change in the United States occupational and industry mix in the 33 months following ChatGPT's release against equivalent periods during the diffusion of computers and of the internet, the Budget Lab found occupational dissimilarity, industry dissimilarity and exposure measures flat or within historical ranges, and no sign that measures of AI exposure, automation or augmentation were related to changes in employment or unemployment to date; the authors note that historical workplace technological disruption has unfolded over decades and that 33 months is a short window. V55 verification basis: budgetlab.yale.edu was not fetched. The findings and framing were carried in search results attributing them to the Budget Lab's own publications; the October 2025 date of the principal analysis comes from secondary reporting and was not confirmed against the publication itself.
17. Applying a task-based model and a version of Hulten's theorem to existing estimates of AI task exposure and task-level cost savings, the macroeconomic effects appear non-trivial but modest: no more than a 0.66% increase in total factor productivity over ten years, with the author arguing that even this is likely exaggerated and predicting gains below 0.53%. V55 verification basis: the article was not fetched, but search results carried the journal, volume 40, issue 121, pages 13–58, the January 2025 publication and both the 0.66% and 0.53% figures and attributed them to this paper.
18. The second edition of the multinational scientific assessment, with more than ninety authors and advisers from over thirty countries, reports that general-purpose AI effects on labour markets are so far mixed — early evidence of reduced demand for easily substitutable work such as writing and translation, increased demand for complementary skills, newer research indicating no significant effect on overall employment to date, and possible effects on junior workers in exposed occupations such as software engineering and customer service — alongside rapid capability improvement in mathematics, coding and autonomous operation. V55 verification basis: the report was not fetched, but search results carried the chair, the authorship scale, the February 2026 publication and these labour and capability findings, attributing them to the report and to its executive summary.