New research challenges assumptions about who faces the greatest reskilling pressures.
As technological advancement reshapes the workforce, much attention has focused on how technology might displace or eliminate jobs. But new research from James Evans (Faculty Co-Director of Novel Intelligence; Max Palevsky Professor of Sociology & Data Science; and Director of the Knowledge Lab) and colleagues reveals how technology is transforming the content and skills needed for the jobs that remain.
Published in Nature Communications, the study analyzed 167 million online job postings covering 721 occupations across nearly a decade. Challenging the conventional wisdom that technological disruption primarily affects STEM and high-skill workers whose jobs exhibit greater skill volatility, the researchers found that jobs with fewer skill requirements, lower wages, and less education experience the greatest skill change.
The researchers quantify skill change not only by the number of skills required but also by how far new skills deviate from existing ones. For example, a computer programmer learning a new coding language faces a smaller cognitive leap than a food batchmaker suddenly needing database administration skills. When this “skill distance” is factored in, the researchers found that lower-skilled jobs undergo far more radical transformations.
The analysis also revealed a “catching-up” pattern across geography and demographics: jobs at small employers and in smaller labor markets required larger skill upgrades to close the gap with large companies and major metropolitan areas, where skill change is more incremental. This creates uneven pressures across the workforce, disproportionately affecting women, minorities, and workers in smaller communities who are more likely to hold these positions.
However, the researchers note that these dynamics could also narrow gaps in job quality and prospects as lower-skilled positions incorporate more sophisticated capabilities, a trend consistent with recent wage growth among low-wage workers.
Using AI-constructed “skill embedding spaces” to geometrically represent how skills relate to one another, the researchers chart the direction of job evolution over time, validating that distances in this space correspond with the amount of training required to move from a known to an unknown skill. This approach could give policymakers and educators tools to anticipate workforce transformations before they arrive, enabling more targeted support for workers facing the steepest reskilling demands.
This figure integrates two datasets to visualize where future skill-adjustment pressures may concentrate. The Lightcast job-posting dataset (formerly Burning Glass) contains detailed U.S. job advertisements representing 1,060 ONET occupations across 52,979 locations. For this visualization, we use a two-month slice—January and October 2018—of the dataset to obtain a cross-sectional snapshot of local labor demand. The Occupation Automation Risk dataset (OC) provides automation-risk scores for 702 ONET occupations, inferred from a training set of 70 occupations labeled as “computerisable” by artificial-intelligence experts (Frey & Osborne 2013). We map 47,990 U.S. locations with at least three job postings in the selected months. For each location, we construct a local labor-market profile by calculating the occupation share distribution from its observed job postings, then compute its automation-risk score using a market-share-weighted average of OC values. Log labor-market size and automation risk exhibit a modest but statistically significant negative correlation (Pearson r = –0.11, P < 0.001), revealing that smaller labor markets face disproportionately higher exposure to computerisation—and thus greater potential upskilling pressure. In this version, instead of point-level mapping, we aggregate the data into 80 latitude bands and plot the smoothed longitudinal automation-risk profiles as stacked black-and-white ridgelines. Peaks represent regions with higher average automation risk, and the stacked ridgeline occlusion highlights large-scale geographic structure while preserving the underlying quantitative information.Mapping U.S. Labor Markets’ Upskilling Pressure Through Automation Risk. This figure integrates two datasets to visualize where future skill-adjustment pressures may concentrate. The Lightcast job-posting dataset (formerly Burning Glass) contains detailed U.S. job advertisements representing 1,060 ONET occupations across 52,979 locations. For this visualization, we use a two-month slice, January and October 2018, of the dataset to obtain a cross-sectional snapshot of local labor demand. The Occupation Automation Risk dataset (OC) provides automation-risk scores for 702 ONET occupations, inferred from a training set of 70 occupations labeled as “computerisable” by artificial-intelligence experts (Frey & Osborne 2017). We map 47,990 U.S. locations with at least three job postings in the selected months. For each location, we construct a local labor-market profile by calculating the occupation share distribution from its observed job postings, then compute its automation-risk score using a market-share-weighted average of OC values. Log labor-market size and automation risk exhibit a modest but statistically significant negative correlation (Pearson r = –0.11, P < 0.001), revealing that smaller labor markets face disproportionately higher exposure to computerisation—and thus greater potential upskilling pressure. Locations are color-coded by automation risk—brown for high-risk areas (risk > 0.65; top 25%) and cyan for low-risk areas (≤ 0.65; bottom 75%).
