AI-exposed job growth needs a gender-and-education composition check

What the study found

Pew Research Center compared U.S. employment in 480 occupations before widespread generative AI use, using 2017–19 data, with January 2023–June 2026. It built an occupational exposure score from O*NET task descriptions and Current Population Survey employment counts, then compared growth for men and women within each occupation rather than treating occupations as inherently gendered. Axios separately covered the emerging gap and its workforce implications.

For each 1-point increase on Pew’s 0 to 3.5 exposure scale, men’s employment grew by an estimated 6.3% and women’s by 0.2%. The education split was more uneven: among workers with high school education or less, the estimates were 8.8% for men and -9.5% for women, although men in this education group were sparsely represented in highly exposed occupations; among workers with some college, including an associate degree, the estimates were 3.6% and -6%; and among those with a bachelor’s degree or more, 0.8% and 0.7%, neither statistically significant according to Pew.

What the comparison cannot establish

Pew explicitly says the relationship is not necessarily direct. The design compares periods and occupational exposure; it does not observe an employer introducing AI and randomly assign workers to exposure. Industry demand, the post-pandemic recovery, occupational mobility, hiring patterns, hours and exits can move the same employment counts. An exposure score also describes what AI could do or assist with, not how intensively a particular employer adopted it.

The headline therefore should not become “AI displaced women”. The useful signal is narrower: the composition of employment growth differs as exposure rises, and the divergence is concentrated below bachelor’s level. That is enough to justify investigation, not enough to diagnose a cause or prescribe a course.

The composition check

A workforce team should reproduce the comparison inside its own decision boundary. Start with occupation and business unit, then separate hires, voluntary exits, involuntary exits, internal moves, hours and contract type. Add the material tasks that changed and the date of each AI deployment. Where lawful and proportionate, break the measures down by gender and education, and suppress small cells that could expose individuals.

The check should also look for counter-patterns. If women’s exits rose before the deployment, if one shrinking industry dominates the result, or if hours changed while headcount stayed flat, a training intervention may miss the mechanism. If task access or internal mobility differs after a deployment, leaders can test whether job design, selection rules, schedule constraints or learning access explain the gap.

Keep the note tied to decisions. A quarterly review should name which mechanism remains plausible, what new data arrived and whether a previously approved response should continue. Compare the affected group with a relevant internal reference group, but avoid turning a small difference into a trend. Where representation is thin, use qualitative task interviews and case review alongside counts. Publish the limits with the result so a manager cannot reuse the exposure score as a performance or redundancy proxy.

Decision rule

Do not fund a targeted AI-skills programme from the exposure statistic alone. First require a signed composition note that names the denominator, time window, task changes, hiring and exit flows, plausible alternative explanations and the evidence that would distinguish them. Then choose the smallest reversible response and measure whether the relevant flow changes. This extends the earlier Skills Intelligence analysis of stage-by-stage talent evidence from representation counts to occupation and transition mechanisms.