Stanford Says 13%, Yale Says No Disruption: Why Both AI Jobs Studies Hold

In August 2025, Stanford economists reported a 13% relative decline in employment for workers aged 22 to 25 in the most AI-exposed jobs. Six weeks later, the Yale Budget Lab reported that the broader labour market showed no discernible disruption since ChatGPT launched. The Stanford AI jobs study and the Yale study covered almost the same months, and neither is wrong. They were built to see different things.

The short answer

Both studies can be right because they ask different questions. Stanford asks whether young workers in AI-exposed jobs are falling behind similar workers elsewhere, using payroll records. Yale asks whether the whole economy's job mix is changing unusually fast, using a household survey. A sharp decline inside one small group can be real and still be invisible economy-wide.

Stanford AI jobs study vs Yale: same window, opposite headlines

The two papers landed weeks apart and were reported as if one had refuted the other. The Stanford paper by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen was published as a working paper in August 2025. The Yale Budget Lab analysis, written with researchers at Brookings, followed on 1 October 2025.

The coverage windows overlap almost exactly. Yale's analysis began in November 2022, the month ChatGPT launched, and ran to the July 2025 Current Population Survey release, a span of 33 months. Stanford's later update dates its first estimate to the July 2025 data vintage.

"Stanford finds AI is taking young people's jobs" and "Yale finds AI is not affecting jobs" are both bad summaries. Read the papers rather than the coverage, and the authors largely agree on what the data can and cannot show. They disagree mainly on where to point the instrument.

Because the Yale paper arrived second, much of the coverage cast it as a rebuttal. It was not written as one. The Yale report cites the Brynjolfsson paper directly and says its own recent-graduate pattern "could be consistent with" the Stanford result. It is the most useful sentence in either paper for anyone reconciling them.

What the Stanford ADP payroll study measured

The Stanford team used payroll records from ADP, the largest private payroll processor in the US. Payroll data records who is actually paid by an employer each month, which makes it a measurement of jobs held rather than a survey answer. Its weakness is that it covers ADP's clients, not a designed national sample.

The researchers sorted occupations into five groups, or quintiles, by how exposed their tasks are to generative AI. The August 2026 update says the exposure scores come from Eloundou and coauthors' 2024 measure, which rates tasks by whether a large language model could perform them. They then tracked employment by age group within each quintile.

The headline result was a 13% relative decline for 22 to 25 year olds in the most exposed occupations. "Relative" is the key word. It compares those workers with similar young workers in less exposed jobs. It also controls for shocks hitting the whole firm, so a company shrinking across the board does not count as an AI effect.

The checks the Stanford team ran

The working paper reports that the pattern survives when technology firms are excluded and when occupations suited to remote work are excluded. The August 2026 update adds controls for exposure to interest-rate increases and for firms entering or leaving the sample. Those checks matter, because the obvious rival explanations are a tech hiring hangover and higher borrowing costs.

The paper also found that adjustment showed up in employment rather than pay. Declines were concentrated where AI tends to automate tasks rather than augment them. Where AI complements workers, employment was flat or rising, particularly for experienced staff.

What Stanford says it did not prove

The authors are explicit. The August 2026 update calls its results "descriptive patterns, not causal estimates of the effect of AI." It notes that some differential trends appear before generative AI was widely used. It also concedes the gaps shrink when education is accounted for, and are larger in the ADP sample than in national survey benchmarks.

What the Yale Budget Lab labour market analysis measured

Yale started from a different question. Rather than track one age group, it asked whether the economy's overall mix of occupations has been changing faster since ChatGPT than it did during earlier technology shifts. If AI were reshaping cognitive work at scale, the share of workers in each occupation should be moving unusually quickly.

To measure that, the team adapted a dissimilarity index from the sociologists Duncan and Duncan. Each month's occupational shares are compared with the starting month, and the absolute differences are summed across all occupations. The data is the monthly Current Population Survey, the household survey the government uses for the unemployment rate.

The benchmarks were 1984 to 1989, when personal computers spread, and 1996 to 2002, when the internet arrived, with 2016 to 2019 as a quiet control period. By 2002 the occupational mix was at most around 7 percentage points different from 1996. The post-ChatGPT path was running only about 1 point higher.

The verdict and its exact wording

The report's conclusion is that the broader labour market "has not experienced a discernible disruption since ChatGPT's release 33 months ago". It adds that the faster shift began in 2021, before generative AI was released. The Information, Financial Activities and Professional and Business Services sectors moved more than the aggregate, but their trends also started before ChatGPT.

Yale also tested exposure against unemployment directly. It used OpenAI's exposure scores and Anthropic's usage data. Unemployed workers came from occupations where about 25% to 35% of tasks could be done by generative AI, regardless of how long they had been out of work. There was no upward trend.

Yale did look at young workers, comparing recent graduates aged 20 to 24 with older graduates aged 25 to 34. The gap in their occupational mix has rarely left a 30% to 33% range since January 2021. It rose slightly faster in recent months, which the authors say could reflect AI or simply a slowing labour market, and they warn about small samples.

The Stanford and Yale AI labour market methods side by side

Each row is drawn from the papers themselves. Neither method is simply better.

Read across the "Population" and "Measure" rows and the apparent contradiction mostly dissolves. Stanford holds the economy constant and zooms in on one cohort. Yale holds the cohort constant at "everyone" and asks whether the whole distribution moved. Those are complementary designs, not competing ones.

The same pattern of instruments disagreeing because they count different things shows up in announced layoffs versus measured separations, covered in the comparison of Challenger layoff counts and payroll data. It also shows up in graduate hiring plans versus graduate unemployment, laid out in the analysis of conflicting entry-level hiring data.

Why both AI impact findings can be true at once

There are three mechanical reasons the studies can disagree without either being in error. They follow from the arithmetic of the two designs.

A small cell inside a large index

Yale's dissimilarity index sums changes across every occupation and every age group. A 13% relative drop among 22 to 25 year olds in the most exposed quintiles moves a small slice of total employment. Summed with hundreds of occupations that did not move, it can leave the aggregate index near its historical path.

Yale's own sample sizes make the point. Recent graduates are about 1,100 observations a year, roughly 3% of the survey. If the 18% high-exposure share Yale reports for the whole workforce held for that group, only about 200 recent graduates a year would sit in highly exposed jobs. That estimate is mine and illustrative.

A relative gap is not a level change

Stanford reports gaps relative to a comparison group. A 13% relative decline can come from the exposed group shrinking, the comparison group growing, or both. The August 2026 update shows it is both. Employment for 22 to 25 year olds in the two most exposed quintiles fell about 11% from November 2022 to June 2026. The same age group in the three least exposed quintiles grew about 10%.

Those two figures reproduce the headline. Divide 0.89 by 1.10 and you get 0.81, a gap of about 19%, which is the number Stanford published. So the widely quoted gap is roughly half a fall in exposed jobs and half a rise in less exposed ones. Yale's index would register some of that as normal reshuffling between occupations.

Payroll records and survey answers diverge

The third reason is the data itself. Stanford's update states that its gaps are larger in the ADP sample than in national survey benchmarks. It traces part of that difference to education, health care and public administration, and part to recent Census Bureau revisions of the population weights in the American Community Survey.

So the Stanford team has already conceded that a national survey would show a smaller effect. Yale is using a national survey. The two results converge once you accept that ADP's client base over-represents the cell where the effect is strongest. I would treat the true national figure as somewhere below Stanford's estimate and above zero.

The Stanford 13% has become 19%, and the Yale verdict has not moved

Both teams kept publishing. That makes this more than a one-off disagreement, because each update tests whether the original reading was noise.

Why 13% and 19% are not a clean series

Be careful reading the jump from 13% to 19% as six points of fresh damage. The August 2026 update gives the gap as 15% at the July 2025 data vintage, not 13%. The authors note data pipeline changes that left estimates directionally consistent but more sensitive to specification choices.

The 13% was also a specific estimate after controlling for firm-level shocks, while the 15% and 19% compare young exposed workers with similar workers in less exposed jobs. The like-for-like change on the update's own basis is 15% to 19% in under a year. That is still a widening gap, and it is the honest version to quote.

In February 2026, Fortune reported on a Budget Lab update using November and December survey data. It quoted the report saying the picture "largely reflects stability, not major disruption at an economy-wide level". Executive director Martha Gimbel told Fortune there did not "seem like there's major macroeconomic effects here".

The Budget Lab's tracker, updated on 14 October 2026, now adds a synthetic difference-in-differences test, a method that builds a weighted comparison group for exposed occupations. It reports that this test "does not yet clearly indicate an AI-related labor market footprint". The word "yet" appears throughout Yale's work, and I think it should be read literally.

Where this argument is weakest

The reconciliation above is tidy, which is a reason to distrust it. There are three places it could be wrong.

First, it leans on Stanford's controls being adequate. Stanford itself reports trends that started before generative AI and gaps that shrink once education is included. If the young-worker gap mostly reflects a weak graduate market that happened to hit exposed occupations first, then the "both are right" framing flatters Stanford. In that case Yale is right and Stanford has found a correlation with a different cause.

Second, it may flatter Yale. A dissimilarity index built on a survey with about 1,100 recent graduates a year is close to blind to a cohort effect. "No discernible disruption" is then partly a statement about the instrument's resolution, not the economy. Yale says as much when it warns that exposure-based analysis likely underestimates potential disruption.

Third, the illustrative 200-observation estimate assumes the workforce-wide high-exposure share applies to recent graduates. Graduates may well be more concentrated in exposed white-collar work than the average worker, which would raise that number.

Finally, I could not open the full PDFs of either the August 2026 Stanford paper or Yale's latest tracker during this research. The claims above come from the papers' published summaries, abstracts and report pages, and from one Fortune report. Sample sizes for the ADP data are not stated here because I could not verify them from a primary page.

What the two studies mean for AI and early career development

For anyone planning graduate hiring, the practical lesson is that both studies point at the same mechanism. Stanford's update says the adjustment works primarily through reduced hiring of young workers, not layoffs. A hire that never happens is exactly the kind of change a whole-economy survey index struggles to see.

If you quietly stop opening junior roles in exposed functions, you are the mechanism behind the Stanford number, and no national statistic will flag it for years. The longer-run cost of that choice for engineering teams is traced in the analysis of the junior developer pipeline and the piece on the missing junior rung.

Stanford's other finding cuts the other way. Experienced workers in occupations that rely on tacit knowledge, the judgement people learn on the job rather than from documents, saw employment rise. If your seniors are getting more valuable because of what they learned as juniors, cutting the junior intake is a bet that you will never need to grow more of them. The case for that skill premium is made in the analysis of the tacit knowledge premium.

Reading the next AI jobs study

Before you act on any new AI labour market study, ask four questions. What population does it cover? Is its measure relative or absolute? Is the data payroll records, a survey or job postings? And does the authors' own caveat change the headline? The same discipline applies to economy-wide exposure claims, examined in the review of the MIT Iceberg Index method.

Frequently asked questions

Where to start

Pick one function on your team where AI now handles tasks a junior used to do, and pull your own hiring records for it since late 2022. Count the junior requisitions you opened each year. That internal series is the version of the Stanford measure that actually applies to you, and no outside study can produce it.

Then decide, in writing, whether those roles were paused because of AI or because of budget. If it was AI, name who will become your experienced staff in five years. Bookmark Yale's tracker as well: when its index finally moves, it will be a signal the effect has stopped being confined to one cohort.

Related reading