Imagine a Friday deadline that does not exist but could belong to anyone. A woman of twenty-four, whom we will call Ivana, has two envelopes on her desk. One holds an offer from a bank’s analytics team, with a salary and a start date. The other is an acceptance to a one-year master’s degree in data science that would cost her a year of pay. She has spent a week asking which of the two a hiring manager will still pay for in five years. Her father, a bank clerk for twenty-five years, tells her that experience always wins. Her professor tells her that the credential is what opens doors. Both are speaking from a labour market that may no longer exist.

On 26 October three economists at the Universidad Torcuato Di Tella in Buenos Aires, Eduardo Levy Yeyati, Darío Judzik and Juan Ignacio Barbiero, posted a working paper that puts numbers on exactly this argument. Its title is “Practice Makes Perfect”, and its claim is that generative AI has already repriced human capital in the US occupations most exposed to it. The wage paid for another year of schooling has fallen. The wage paid for another year of experience has risen for people early in a career and fallen for people late in one. The paper is a preprint and has not been peer reviewed.

Two prices, one shock

The authors start from an old distinction. Schooling, they argue, is a rough stand-in for codified knowledge, the kind that can be written down, taught and stored. Experience is a stand-in for practice-based knowledge: the judgment a caseworker or a surgeon builds by repetition, which resists being written down. A model that supplies codified knowledge at the point of work should make the first kind cheaper, and the authors predict that this lowers what schooling pays. If the two kinds of knowledge complement each other, abundant codified knowledge should raise the value of judgment, most of all early in a career, where practice teaches fastest. Some experience is also tied to the tools on which it was earned, and that part can lose value.

To test this they use the Current Population Survey’s earnings files for 2018 through June 2026. They take the top and bottom quarters of US occupations ranked by an exposure index that Edward Felten, Manav Raj and Robert Seamans built before ChatGPT appeared. In the exposed group are software developers, financial managers and lawyers. In the comparison group are truck drivers, construction labourers and waiters. The authors compare each group before and after November 2022, when ChatGPT was released. They estimate changes in wage gradients, the price of a year of schooling or experience inside an occupation, and leave the question of whether average pay moved to a separate term.

What the wage data show

Before the shock, a year of schooling paid 7.5 log points in the exposed occupations, more than four times the return in the comparison group. After it, the authors find the return fell by 0.55 log points, roughly seven percent of its former size. In the comparison occupations it stayed at 1.7 log points. Because the estimate compares workers inside the same occupation, it does not simply reflect that exposed jobs are more educated.

Experience moved in two directions. At three years of potential experience, the return to one more year rose by 0.26 log points. It was indistinguishable from zero at fifteen years and fell by 0.12 log points at twenty-five. Under the authors’ quadratic model, the effect changes sign at 18.1 years. For a worker with sixteen years of schooling and fifteen of experience in an exposed occupation, the paper estimates a net wage change of minus 3.2 log points against the path the comparison group implies.

The timing supports a link to the release. For schooling, the four pre-release coefficients are small and a joint test finds nothing before November 2022; the break shows up in the first post-release year. A placebo that dates a fake shock to the start of the pandemic returns a schooling effect of minus 0.02. The results also survive a common top-coding rule across the whole period, which matters because the Census Bureau changed how it records top earners in April 2023 and 11.5 percent of observations in the exposed group are top-coded, against 0.7 percent in the comparison group.

Where the authors are cautious

The paper is open about its weak points. A second falsification, comparing the second and bottom exposure quartiles, neither of which was treated, produced a schooling effect of similar size but failed the pre-trend test, which the authors take as a warning that schooling composition can create a false break. The late-career decline is the weaker half of the experience result: its pre-trend test gives a p-value of 0.06, and the authors say they put more weight on the early-career rise and on the overall rotation than on every late-career number. Years of schooling do not capture field or quality, and potential experience counts years since leaving school rather than years spent on exposed tasks. Young workers who stay in exposed jobs may be a selected group. The schooling coefficient can also reflect signalling, and a technology that changes how fast employers judge workers could narrow the gradient without changing what school teaches. The authors read their numbers as a repricing of measured schooling and experience, and say they are not a forecast of long-run returns.

The same thread, from other directions

Other sources this month point the same way from different starting points. The Stanford Digital Economy Lab’s “Canaries in the Coal Mine?”, revised on 12 August 2026 by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, uses ADP payroll data on millions of US workers through June 2026. It finds no evidence of widespread displacement, but employment of workers aged 22 to 25 in AI-exposed occupations stands 19 percent below where it would be had it kept pace with less exposed peers, while experienced workers show no comparable gap. The gap operates mainly through reduced hiring, and the authors call their facts early, descriptive indicators and not causal estimates. The Federal Reserve’s Beige Book of 14 October reported that in five districts some firms had cut staffing through attrition, delayed hiring or expected less job creation after investing in AI. And on 16 October Anthropic’s economists reported that the junior share of its own technical hires fell from 27 percent to 9 percent.

None of these studies measures the same thing as the Argentine paper. The payroll data count jobs, the Beige Book collects anecdotes and the Anthropic account describes one company. The wage paper is about prices. What they have in common is a story about the bottom rungs of professional ladders, and the paper’s own caveat that early-career results fit a formation channel without testing it is a reminder of how little is settled. The authors end on a question they leave open: whether workers can still acquire the practice-based knowledge that is rising in value when AI automates parts of the learning itself.

Back to the invented Friday

Ivana does not find an answer in the paper, because it gives none. It reports that a year of schooling paid less in exposed occupations after 2022, that early experience paid more, and that a year of schooling still paid more than it did in the comparison group. She takes the bank job, and she makes a list of the tasks she is allowed to touch in her first year that no model could have learned from a document, since early practice is where the paper finds the return rising. The master’s offer is still open for a deferral, and she asks about it on Monday.

References

  • Brynjolfsson, E., Chandar, B., & Chen, R. (2026, revised 12 August). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. digitaleconomy.stanford.edu
  • Federal Reserve System (2026). Beige Book, October 2026: National Summary. federalreserve.gov
  • Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, 42(12), 2195-2217. doi.org
  • Levy Yeyati, E., Judzik, D., & Barbiero, J. I. (2026). Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience. arXiv. arxiv.org
  • Wilmers, N., Catipon, R., Massenkoff, M., & McCrory, P. (2026). How employment is changing at Anthropic. The Anthropic Institute. anthropic.com