SINCE CHATGPT, A DEGREE PAYS A LITTLE LESS

What a worker knows comes in two kinds. Part of it can be written down, taught and stored: formulas, rules, procedures. The other part is learned by doing and stays bound to the person: in the paper’s examples, a caseworker learns which clients need closer follow-up, a surgeon learns when tissue tension is wrong, a negotiator learns how to read a room. Generative AI makes the first kind cheap and abundant at the moment it is needed. What does that do to pay?

Two prices on a payslip

Eduardo Levy Yeyati, Darío Judzik and Juan Ignacio Barbiero, of Torcuato Di Tella University in Buenos Aires, look at two familiar numbers: how much more a worker earns for one more year of schooling, and for one more year of experience. Schooling stands in, imperfectly, for knowledge that can be written down; years since leaving school stand in for knowledge gained through practice.

Their model makes two predictions. If machines supply written-down knowledge, the extra pay for schooling should shrink. If that knowledge and practical judgment work best together, practice should become more valuable — especially early in a career, when people learn fastest. But some experience is tied to the old way of doing the job and may lose value, so the effect could turn negative later on.

Exposed jobs, sheltered jobs

The test uses the US Current Population Survey from January 2018 to June 2026: about 656,000 wage records. Jobs are ranked by an AI-exposure index built before ChatGPT existed. The most exposed quarter — software developers, administrative assistants, accountants, financial managers, university teachers — is compared with the least exposed quarter: truck drivers, warehouse labourers, cleaners, care aides, waiters. The clock starts in November 2022, with ChatGPT’s public release.

Before that date, a year of schooling raised pay by about 7.5% in exposed jobs, against 1.7% in sheltered ones. The analysis compares workers inside the same occupation, before and after, so it measures changes in these “prices” of schooling and experience, not in average wages.

A year of school, repriced

In exposed jobs, the pay bonus for one year of schooling fell by 0.55 percentage points relative to sheltered jobs — from about 7.5% to 6.9%, a drop of about 7% of its former value. In sheltered jobs it stayed at 1.7%. There was no drift before the release, and the break shows up within the first twelve months.

A career that tilts

Experience did not shift up or down as a whole; the curve rotated. One more year of experience pays 0.26 percentage points more for someone with three years behind them, unchanged at fifteen years, and 0.12 percentage points less at twenty-five. In the authors’ main model, the extra pay for one more year of experience switches from gain to loss at about 18 years — which does not mean that workers past 18 years earn less overall; with more flexible curves, the tilt remains but that exact point is less certain. The tilt is strongest among workers holding a college credential (a two-year associate degree or higher).

Added up, a worker with sixteen years of schooling and fifteen of experience in an exposed job earns about 3.2% less than the trend of sheltered jobs would predict.

Checks against coincidences

• Moving the start date to January 2020, the beginning of the pandemic, produces no effect.
• Applying one common rule to the highest salaries, which US statistics cap and record differently since April 2023, leaves the results almost unchanged.
• Removing computer and maths jobs, hit by tech-industry layoffs at the same time, still leaves the schooling drop at 0.49 points.

The late-career decline is the weakest part of the evidence: before the release, that end of the curve already wobbled a little.

Not a forecast

The authors do not claim that degrees have lost their worth or that graduates became less productive. Their result is a short-run repricing inside exposed jobs. Schooling choices, training and entry-level jobs will adjust. One question, they write, may matter most in the long run: whether workers can still acquire the practical know-how that becomes more valuable, if AI automates part of the learning process itself.