---
id: "lib-2026-10-26-practice-makes-perfect-tacit-knowledge-generative"
type: "library"
title: "Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience"
lang: "EN"
org: "arXiv"
source_type: "Paper"
date: "2026-10-26"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-26-practice-makes-perfect-tacit-knowledge-generative/"
original_url: "https://arxiv.org/abs/2610.36008"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "A working paper by Eduardo Levy Yeyati, Darío Judzik and Juan Ignacio Barbiero (Universidad Torcuato Di Tella) with a formal model, a difference-in-differences design on US survey data, event studies, placebos and hard robustness checks reported even where they weaken results."
summary_basis: "full-text"
original_minutes: "120"
tags: ["future of work", "returns to schooling", "tacit knowledge", "wages"]
frames: ["tacit-to-explicit"]
---

# Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience

Tier A (Strong evidence): A working paper by Eduardo Levy Yeyati, Darío Judzik and Juan Ignacio Barbiero (Universidad Torcuato Di Tella) with a formal model, a difference-in-differences design on US survey data, event studies, placebos and hard robustness checks reported even where they weaken results.

Summary based on the full text.

## In short

An economics working paper (arXiv, 26 October 2026) on how generative AI reprices two parts of human capital. The authors argue that AI makes codified knowledge, the kind taught in school, cheap at the point of work, while judgment learned through practice stays tied to the worker. Using US earnings data from 2018 to June 2026, they find that after ChatGPT's release the wage return to an extra year of schooling fell in highly exposed occupations, and the return to experience rotated: it rose early in careers and fell late in careers. They read this as practice gaining value where it transfers to the new tools, and technology-specific experience losing value.

## Key numbers

- 0.55 log points: fall in the return to a year of schooling in the most AI-exposed occupations after ChatGPT's release
- about 7.3 percent: that fall relative to the pre-shock return of 7.5 log points in exposed occupations
- 0.26 log points: rise in the return to a year of experience at three years of potential experience
- 0.12 log points: fall in the return to a year of experience at twenty-five years
- 18.1 years: point where the effect on the experience gradient changes sign (s.e. 2.1)

## What it reports

- Return to a year of schooling: down 0.55 log points in the top exposure quartile relative to the bottom, about 7.3% of the pre-shock return of 7.5 log points.
- Return to a year of experience: up 0.26 log points at three years, no significant change at fifteen, down 0.12 at twenty-five. Under the quadratic specification the effect crosses zero at 18.1 years; the authors stress this is the gradient changing sign, not a crossover in wage gains and losses.
- Event studies are flat before the release; the schooling break shows up in the first post-release year. A placebo dated to the pandemic (January 2020) gives no effect, and dropping the tech sector changes little.
- A common artificial censoring rule across the whole sample gives -0.53 for schooling and nearly the same experience rotation. Allowing each occupation its own pre-shock gradients gives -0.58, +0.26 and -0.11.
- Dose-response: the total wage effect for a college-educated worker rises from 0.9% in the second exposure quintile to 3.0% in the top.
- With a different, task-based exposure measure, the schooling decline is 0.42 log points; the early-career experience gain is much weaker.

## How they got it

US Current Population Survey Outgoing Rotation Group, January 2018 to June 2026, 655,919 observations after restrictions. Treatment is the top quartile of the Felten-Raj-Seamans AI occupational exposure index, built before ChatGPT; control is the bottom quartile. The estimating equation interacts schooling and a quadratic in potential experience with treatment and post-November 2022 indicators, with occupation and month effects, so the estimates are changes in skill prices, not wage levels. Interval-censored models handle top-coded earnings.

## Limits

- Top-coding is concentrated in exposed occupations (11.5% of observations against 0.7%) and the Census rule changed in April 2023; the harshest treatment cuts the schooling estimate to -0.09.
- A comparison between two untreated quartiles also gives a schooling estimate of similar size but fails its pre-trend test; the late-career event study is the weaker margin.
- Results depend in part on the exposure index; years of schooling and potential experience are blunt measures.
- Short window; the authors say it is not a forecast of long-run returns.

## Why it matters for companies

The paper gives numbers to a pattern many managers sense: in AI-exposed roles, credentials buy less and early hands-on practice buys more, while long tenure on old routines is worth less. For workforce planning this points to two actions. Pay and hiring frameworks that lean on degrees in exposed roles may be overpaying for knowledge the tools now supply. And because practice-based judgment grows most in the first years, organisations that cut junior roles or let AI absorb the learning tasks risk losing the pipeline that produces the experience the market now rewards.

Original: https://arxiv.org/abs/2610.36008