After ChatGPT, a Year of School Pays 7% Less in AI-Exposed Jobs

In the US occupations most exposed to generative AI, an extra year of schooling raised wages by 7.5 log points before November 2022. After ChatGPT shipped, that return fell by 0.55 log points, a drop of roughly 7% against its own starting level. [Fact] In the same jobs, workers three years into their careers saw their experience premium rise, and workers twenty-five years in saw theirs fall. [Fact]

The paper is "Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience," posted to arXiv on October 26, 2026, by Eduardo Levy Yeyati, Darío Judzik and Juan Ignacio Barbiero of Universidad Torcuato Di Tella in Buenos Aires; Levy Yeyati is also affiliated with the Inter-American Development Bank. [Fact] It tests a wage model on 655,919 Current Population Survey records from 2018 to June 2026. [Fact]

Two kinds of knowledge, one of them now abundant

The model splits what a worker knows into two parts. Schooling stands in for codified knowledge, material that can be written down, taught and stored. Experience stands in for tacit knowledge, the judgment a caseworker, surgeon or negotiator builds by doing the job. [Fact] Generative AI enters as machine-supplied codified knowledge, a perfect substitute for the schooled kind. [Fact] Two predictions follow: the schooling premium falls as AI capability rises, and the experience profile rotates, steeper early and flatter late, when young workers' know-how complements the machine while senior know-how is tied to routines it replaces. [Claim]

What the wage data show

Treatment is the top quartile of the Felten, Raj and Seamans AI exposure index, which contains software developers, financial managers and lawyers; the control is the bottom quartile, with truck drivers, construction laborers and waiters. [Fact] Before the shock, treated workers earned 65 log points more and had 3.6 more years of education. [Fact]

After November 2022 the return to schooling in treated occupations fell by 0.55 log points (s.e. 0.19) relative to controls, while the control-group return stayed at 1.7 log points in both periods. [Fact] The drop appears in the first twelve months after release, at 0.91 log points, and the pre-release trend test gives p = 0.90. [Fact]

The experience side rotates. The return to an added year rises by 0.26 log points at three years of experience and 0.23 at five, is indistinguishable from zero at fifteen, and falls by 0.12 at twenty-five. [Fact] The quadratic crossover sits at 18.1 years (s.e. 2.1); a spline puts it at 13.4, so the authors treat the exact year as a summary rather than a boundary. [Fact]

For a treated worker with sixteen years of schooling and fifteen years of experience, the combined post-release wage effect is -3.2 log points, around 3% of pay. [Fact] Across exposure quintiles the hit for a college-educated worker runs from about 0.9% in the second quintile to 3.0% in the top one. [Fact] Setting the paper's figures side by side, the treated schooling premium was 4.4 times the control premium before the release and about 4.1 times after it; the gap narrowed without closing. [Estimate]

What it means for which occupations

The treated quartile is management, business, legal and computer work. The paper names software developers, financial managers and lawyers as examples. [Fact] Dropping computer and math occupations, 14.8% of the treated group, leaves the schooling effect at -0.49, so the result does not come from the technology-sector downturn alone. [Fact] On the control side, truck drivers, construction laborers and waiters show no change in either return. [Fact]

What to prepare

The first five years in an exposed occupation now carry a higher premium on practice than before 2022. For workers past fifteen years, the question is whether their experience is transferable judgment or technology-specific routine: the rotation penalises the second, and the authors state it does not treat all senior experience as obsolete. [Claim] An additional generic year of schooling buys less wage gradient in these jobs than it did in 2021. The authors list curricula, mentoring, AI-assisted learning and the organisation of entry-level work as the margins that will move next. [Fact]

Limits

This is a working paper without peer review, and its estimates price generic years of schooling and potential experience, not fields of study or time spent on AI-affected tasks. The late-career pre-trend test gives p = 0.06, which the authors call the weaker of the two margins, and a harsh robustness check that assigns top-coded wages a single monthly value cuts the schooling effect from -0.55 to -0.09 while leaving the experience rotation intact. The authors describe the findings as a short-run repricing, not a forecast of long-run returns to education.

Sources

• Levy Yeyati, E., Judzik, D., Barbiero, J. I., "Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience," arXiv:2610.36008, working paper, first draft September 2026, this version October 2026: https://arxiv.org/abs/2610.36008

About this analysis

This article was written with AI assistance from the full text of the arXiv paper. The ratio comparison marked [Estimate] is our arithmetic on the paper's figures; occupation links point to our job pages.