AI Changes the Value of Schooling vs Experience (Returns)

In occupations most exposed to AI after ChatGPT's release, the return to a year of schooling fell by 0.55 log points while the return to experience rotated, rising early in careers and declining later, according to a study by Eduardo Levy Yeyati, Darío Judzik and Juan Ignacio Barbiero posted at arxiv.org/abs/2610.36008. The experience premium rises at around three years of potential experience, is unchanged around fifteen years, and declines by about twenty-five years, crossing zero near 18.1 years under the baseline. The authors frame the finding as a repricing of the wage structure inside AI-exposed occupations, not just a shift in overall wage levels.

The Short Answer In occupations most exposed to AI after ChatGPT’s release, the return to a year of schooling falls and the return to experience rotates—rising early and falling later. The study reports a 0.55 log point drop in the schooling return in exposed jobs. So what: in AI-heavy roles, extra credentials may deliver a smaller incremental wage premium, while early-career practice becomes more valuable at least for a while . Career moves and learning plans may need to shift toward experience that transfers to the new tech. Caveat: experience returns are not flat across the whole career—the experience premium rises at around three years of potential experience, is unchanged around fifteen, and declines by about twenty-five, crossing zero near 18.1 years under the baseline. AI Changes the Value of Schooling vs Experience Returns Generative AI isn’t just changing jobs—it’s changing which human skills are scarce. That’s the core idea in new research based on “returns to schooling and experience” in a world where AI makes some knowledge cheap and widely available while other know-how still has to be learned through doing. The study, by Eduardo Levy Yeyati, Darío Judzik, and Juan Ignacio Barbiero, uses new evidence tied to the public release of ChatGPT and asks a surprisingly old question with a brand-new twist: what happens to the wage premium of school and experience when generative AI expands the supply of codified knowledge? See the original paper at https://arxiv.org/abs/2610.36008. Here’s the headline result: in occupations most exposed to AI capabilities, the return to a year of schooling falls , and the return to experience “rotates” —it rises early in a career and falls later. Importantly, these changes show up in the prices of schooling and experience the gradients in wages , not just in overall wage levels. In other words, the wage structure itself is being repriced inside AI-exposed occupations. This is a big deal because “schooling vs experience” is one of the most familiar ways economists describe how careers get rewarded. The paper provides evidence that AI changes the bargain: schooling still matters, but the incremental payoff of extra schooling is smaller in AI-exposed roles after late 2022—while early-career experience becomes more valuable at least for a while . Why This Matters Right Now for Anyone Making Career Decisions This research is significant right now because the most common AI conversation is still framed like this: “AI will automate tasks, so some jobs will pay less.” That’s not wrong—but it misses a subtler labor-market shift economists care about just as much: AI can change what kinds of skills the market rewards. When the payoff to schooling shifts relative to experience, people can end up investing in the “wrong” thing—or at least the wrong order. A real-world scenario you can map onto today: imagine two candidates applying to the same AI-heavy field—say, analytics, compliance, or some legal/professional work. One candidate has more education credentials; the other has fewer credentials but has been doing the work for a couple of years. This paper’s results suggest that, in the most exposed occupations after ChatGPT’s release , the market premium for extra schooling compressed, while early-career experience premium increased around the first few years . That doesn’t mean “school is worthless,” but it changes how you should think about competing for roles when AI tools can instantly deliver a lot of codified content templates, explanations, drafts, standard procedures . How does this build on previous AI research? A lot of earlier evidence focuses on average productivity, employment, or earnings levels. This paper instead looks at the structure of earnings: the returns to education and to experience. That’s an important upgrade because wage levels can move for lots of reasons that have nothing to do with skill pricing. By examining how the education and experience gradients shift differently , the study connects generative AI to a clearer mechanism: codified knowledge becomes less scarce, while practice-based judgment becomes relatively more valuable—at least early on. What the Researchers Actually Measured: Wage Gradients, Not Just Wage Levels The paper’s object of interest is pretty specific: changes in the return to schooling and the return to experience in occupations with different levels of pre-existing AI exposure. The data: CPS earnings across 2018 to June 2026 The authors use the U.S. Current Population Survey Outgoing Rotation Group CPS-ORG from 2018 through June 2026 —so nearly five years pre-treatment and about three and a half years post-treatment . Their worker sample includes people: - age 16 to 64 - with a valid hourly wage - with earnings cleaned using wage validity rules and minimum floors notably, they impose a minimum hourly wage of $3.625 for non-hourly workers and $1 for hourly workers That leaves 655,919 observations . The treatment: occupation-level AI exposure, activated by ChatGPT Treatment is assigned at the occupation level using the AI Occupational Exposure index AIOE from Felten, Raj, and Seamans 2021 . Crucially, this index was built before ChatGPT—so it’s not fitted to outcomes. Then they define the shock date as ChatGPT’s public release in November 2022 , and compare changes in skill prices between: - Top quartile of AI exposure treated occupations - Bottom quartile of AI exposure control occupations Occupations in the middle quartiles are dropped from the main analysis. The key outcome: how schooling and experience “price” inside occupations Instead of just asking “did wages change?”, they estimate changes in the wage equation that includes: - years of schooling - potential experience modeled with a quadratic And—this is the clever part—they look at difference-in-differences in the skill returns : - the change in the education gradient - the change in the shape rotation of the experience profile So the question becomes: after ChatGPT, do the slopes linking schooling and wages, and experience and wages, change differently in exposed vs less exposed occupations? Why this design focuses on gradients The paper is very explicit that it’s not claiming “AI uniformly raises or lowers wages.” Instead, it’s showing a repricing of skills inside AI-exposed occupations: codified and practice-based components trade value relative to each other. The Big Findings: Schooling Returns Fall, Experience Returns Rotate The study’s results are organized around two main facts after ChatGPT’s release: 1. The return to schooling declines in AI-exposed occupations. 2. The return to experience rotates —it rises early and falls later. Education premium: a clear drop after ChatGPT In the baseline specification, the treated-vs-control differential change implies: - The return to one additional year of schooling falls by 0.55 log points standard error 0.19 in exposed occupations after November 2022. - The pre-shock return to schooling in exposed occupations was about 7.5 log points . - So that decline corresponds to about a 7.3% reduction relative to its own pre-shock level as the paper computes . Meanwhile, in control occupations, the schooling return is basically unchanged across the same periods. Experience premium: early-career up, later-career down For experience, because the wage-experience relationship is quadratic, the effect varies by career stage. They evaluate the treatment effect at several points. Under the baseline quadratic: - At 3 years of potential experience: the return to an extra year of experience rises by 0.26 log points SE 0.10 - At 15 years : the effect is statistically indistinguishable from zero - At 25 years : it falls by 0.12 log points SE 0.04 Using the quadratic form, the crossover where the experience effect changes sign is about 18.1 years of potential experience. That’s not the same as saying “experience stops mattering.” It says the marginal value of additional measured experience shifts across the career—consistent with a rotation rather than a uniform shift. The pattern is tightly event-based They also run event studies year-by-year relative to the release to check timing. - Schooling: pre-release coefficients are close to zero; the break appears in the first post-release year - Experience: the ends of the profile diverge after release—positive at early-career points, negative at late-career points That divergence matters: a simple wage level shock wouldn’t naturally generate this “pivot” shape. The Mechanism in Plain English: Codified vs Tacit Knowledge, and Why That Changes Career Payoffs This paper’s story is built on a familiar idea from tacit knowledge: some expertise is easy to write down and spread , and some is only learned through repeated practice in real conditions. Two kinds of knowledge - Codified knowledge : stuff you can express clearly—procedures, templates, explanations, “the standard way” to do something. Schooling is an imperfect proxy for this. - Practice-based / tacit knowledge : judgment, routines, context sensitivity—learned by doing and getting feedback. Experience is an imperfect proxy for this. Generative AI mainly attacks the codified side by making it cheap to generate and distribute. That changes scarcity. And if the market starts treating codified components as abundant, it compresses the value of schooling’s incremental contribution. Why schooling returns fall in AI-exposed occupations In their short-run framework, machine-supplied codified knowledge is a close substitute for the codified part of worker human capital. So after AI capability increases in exposed occupations: - the worker’s share of the codified input falls - the wage gradient attached to schooling declines That’s why the model predicts the education return should drop even if average wages don’t necessarily collapse . Why experience returns can rise early and fall later Here’s the twist: if codified knowledge and practice-based judgment are complements , then when codified knowledge becomes more available, the market can reward people who know how to apply it—especially when people are early in their careers and learning through practice is fast. But there’s a second force: - some of what experienced workers learned may be technology-specific - when tasks and routines change, that older experience can depreciate obsolescence Put together, this can produce: - complementarity dominates early-career - obsolescence dominates later-career - so you get a rotation with a mid-career crossover That is exactly the shape the data show: early career up, later career down. If you want the original model discussion, the paper lays it out formally in the framework section and then connects it directly to the empirical “rotation” patterns again, see https://arxiv.org/abs/2610.36008 . How They Test Whether This Is Really “AI Repricing” and Not a Measurement Trick Whenever someone reports a post-ChatGPT break, you should ask: is this causal—or could it be artifacts like changing wage censoring, pandemic shocks, or composition shifts? The paper does a set of checks that are pretty important. 1 Parallel pre-trends: no break before ChatGPT Their event study checks show flat patterns pre-release for schooling, and support the timing for experience rotation. For late-career points the evidence is weaker, but the overall rotation pattern is the robust feature. 2 A pandemic placebo They re-date the shock to January 2020 and estimate on 2018–2021 . - They find basically no schooling effect - experience effects are null - pre-trend tests don’t reject So the pattern doesn’t look like a generic “big labor market shock” story. 3 Top-coding and earnings censoring: a real concern CPS earnings are top-coded. And in this dataset, top-coding is much heavier in treated occupations: - about 11.5% of observations in high-exposure occupations - about 0.7% elsewhere Plus, the Census Bureau changed top-coding rules around April 2023. That could mechanically distort education gradients. They address this by: - using an interval-censored model not a naive imputation rule - re-running with harsher/common censoring rules - re-running checks where dynamically top-coded values are treated as fixed at assigned monthly means In these variations, the core results remain: - schooling effect stays close to baseline around -0.53 log points under common censoring - experience rotation remains 4 Allow each occupation its own baseline schooling/experience slopes They run a specification where each occupation can have its own pre-shock returns—so you’re not just picking up permanent differences in how occupations price schooling/experience. Under that more flexible approach: - schooling effect is still negative about -0.58 - early-career experience effect remains positive about +0.26 at 3 years - late-career effect remains negative about -0.11 at 25 5 Sector layoffs around ChatGPT: not driving it Because ChatGPT coincided with layoffs, they check by removing certain sectors e.g., computer and mathematical occupations . - The schooling decline and experience rotation largely persist. A useful comparison table: baseline results and robustness qualitative Here’s how to think about the checks at a high level: | Check | Main purpose | What happens to the headline pattern? | |---|---|---| | Event study timing | Verify break occurs after Nov 2022 | Flat pre-period; break appears post-release | | Pandemic placebo re-date | Rule out pandemic-driven changes | Null effects | | Common censoring rule | Remove top-coding regime artifacts | Schooling drop persists; experience rotation persists | | Occupation-specific baseline slopes | Remove fixed differences in Mincer slopes | Results remain similar magnitudes | | Drop tech sector | Control for layoffs coincidence | Pattern largely unchanged | What This Means for People and Institutions: Practical Implications This paper doesn’t tell you “go back to school” or “stop training.” It suggests something more realistic: AI is changing the returns trade-off . For workers: sequence and portfolio of skills may matter more than credentials alone If schooling’s incremental payoff compresses in AI-exposed roles, people may benefit from: - coupling education with early hands-on practice - building judgment and routines that let them extract value from codified AI outputs - staying alert to technology-specific obsolescence, especially later in a career The paper’s model also implies that learning through practice becomes especially valuable when codified content is abundant—so early-career mentorship and real task exposure matter more in AI-heavy work. For employers: hiring signals and training pipelines may need redesign If wages reflect changes in skill prices, employers can reduce misallocation by: - not over-weighting credentials as a proxy for the codified component since AI supplies parts of it - still valuing structured early-career practice where complementarity can kick in - updating training and processes to avoid locking people into routines that depreciate as tools change For educators and training providers: think “practice + judgment,” not “content only” Since AI can generate much codified content, curricula that only teach information may face weaker wage signaling. Programs that emphasize: - application, - feedback loops, - case-based judgment, - and tool use tied to real work contexts, may align better with what the market seems to reward after ChatGPT. Key Takeaways - In AI-exposed occupations, the return to schooling fell after ChatGPT’s public release in Nov 2022. The estimated drop is -0.55 log points for an additional year of schooling SE 0.19 . - The return to experience didn’t move uniformly—it rotated. It rises around early-career stages e.g., +0.26 at 3 years and falls later e.g., -0.12 at 25 years . - The estimated experience effect crosses zero around 18.1 years under the baseline quadratic model. - The mechanism is a knowledge re-pricing : generative AI expands codified knowledge substituting for schooling’s codified component , while practice-based judgment becomes more valuable early—then parts of technology-specific experience depreciate later. - The results survive multiple credibility checks, including event-time pre-trends , a pandemic placebo , and rigorous handling of earnings top-coding especially important here because censoring is concentrated in treated occupations . - This isn’t a long-run forecast of lifetime returns. It’s evidence that AI can change the market prices of existing human capital in exposed occupations—and it likely foreshadows how training, hiring, and career strategies should adapt. If you’d like, I can also translate these findings into a “career strategy checklist” for different reader types student, early-career worker, mid-career professional, and employer/HR using only what the paper’s results can support. Sources Used This article is a plain-English breakdown of the following peer-reviewed preprint. Read the original for full methodology and results: - Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience https://arxiv.org/abs/2610.36008 — arXiv - Authors: Authors: Eduardo Levy Yeyati, Darío Judzik, Juan Ignacio Barbiero