New Grads' Summer 2026 Unemployment Shows No AI Spike — Yet

In June 2026, 7.8% of American college graduates aged 22 to 25 who were in the labor force and not in school were unemployed. That looks like the AI jobs shock everyone has been warning about. It isn't, or at least not yet. Two years earlier, the June-to-August average for the class of 2024 was already 7.8%. [Fact]

That comparison is the core of a new paper by Robert Fairlie (UCLA and NBER) and Jane Wu (UCLA), released as NBER Working Paper 35796 on September 28, 2026. It is the first study to test the class of 2026, the first graduating cohort to enter the labor market after workplace AI use deepened, against the four summers before it. The answer, across every specification the authors ran, is no statistically detectable rise in unemployment. [Fact]

The more useful part of the paper is what it shows about how easy it is to misread this data, and about the graduates the usual measures do not see.

Summer is always bad for new graduates

Unemployment among young graduates has a seasonal shape that swamps most year-to-year movement. It sits around 5% from January to May, jumps in June when the class arrives, and drifts back down through the fall. In 2026 it went from 5.4% in May to 7.8% in June, then 7.4% in July and 6.8% in August. [Fact]

Averaged over June, July and August, the rate was 7.3% in 2026. The four earlier summers were 7.1% (2022), 6.3% (2023), 7.8% (2024) and 7.2% (2025). [Fact] The average of those four is 7.1%, so 2026 is 0.2 percentage points above it, and the spread between the best and worst earlier summers is 1.5 points. The May-to-June jump alone, 2.4 points, is bigger than that entire spread. [Estimate]

That is the practical warning. A headline built on one June number, compared against a spring month, will find a spike every single year.

The study uses the Current Population Survey (CPS), the monthly household survey behind the official U.S. unemployment rate. Its sample is people aged 22 to 25 whose highest degree is a bachelor's and who are not enrolled in school; graduate-degree holders and current students are excluded. The analysis runs from January 2022 to August 2026. One gap worth knowing: there is no October 2025 CPS, because data collection stopped during the federal government shutdown. [Fact]

Three ways of looking, one answer

The authors did not rely on raw averages. Controlling for age, sex, race, region and a linear time trend, the estimated summer 2026 effect on the standard unemployment rate is -0.008 (standard error 0.007), meaning slightly lower than expected, and not distinguishable from zero. [Fact]

They then compared recent graduates with two groups that should be less exposed to AI. Against college graduates aged 30 to 49, the summer 2026 difference is -0.007. Against people aged 22 to 25 without a degree, it is +0.002. Neither is statistically significant, and the matching estimates for the summers of 2023 to 2025 are close to zero too. No summer since 2022 stands out. [Fact]

In raw terms, the gap between young and older graduates was 4.5 points in both summer 2025 and summer 2026: older graduates went from 2.7% to 2.8%, young graduates from 7.2% to 7.3%. [Fact]

The results hold when 2025 is dropped from the baseline, when the time trend is removed, and when the age window is widened to 22-27. [Fact]

The number the headline rate leaves out

The official definition only counts you as unemployed if you actively looked for work in the past four weeks. A graduate who is reading job postings, moving to a new city, or taking a break with no offer lined up is classified as out of the labor force, even if they want a job. The CPS asks those people directly: "Do you currently want a job, either full or part time?" [Fact]

Fairlie and Wu add them back and call the result "sidelined unemployment." Over 2022-2026 as a whole it adds 1.9 points, taking the rate from 5.7% to 7.6%. [Fact]

This is where 2026 does look different. The expanded summer rate was 10.4%, the highest of the five summers, against 9.4%, 9.3%, 10.1% and 9.4% before it. Sidelined graduates added 3.1 points to the headline rate in summer 2026, compared with 2.2 points in 2025. [Fact]

Our own subtraction puts that in context. The sidelined add-on was 2.3 points in 2022, 3.0 in 2023 and 2.3 in 2024. So 2026 is the highest reading, but only 0.1 point above 2023, which nobody treats as an AI year. Measured against the four-summer average (9.55%), the expanded rate is up about 0.85 points, roughly four times the gap on the official measure. [Estimate]

The regressions still do not call that a break. The summer 2026 coefficient on the expanded measure is 0.001 with a standard error of 0.009. [Fact]

What "no effect found" can and cannot rule out

A null result is only as strong as its error bars, and the paper reports them. Taking the standard errors at face value and using the usual 95% band, the headline estimate is consistent with anything from about 2.2 points lower to 0.6 points higher unemployment. On the expanded measure the band runs from about -1.7 to +1.9 points. [Estimate]

Put plainly: the data rule out a large jump, of the kind a two-point spike would be. They do not rule out a smaller one. The comparison against non-graduates is the loosest. On the expanded measure its point estimate is +0.013 with a standard error of 0.011, which leaves room for a relative rise of more than three points. [Estimate] The authors themselves describe the comparison groups as not clean controls and the timing of AI adoption as unclear, and they say the paper is descriptive, not causal. [Claim]

The graduates who vanish from the occupation analysis

The obvious objection to an aggregate null is that it could hide damage in AI-exposed jobs offset by gains elsewhere. The authors test this with two occupation-level exposure measures: the GPT-4-based theoretical exposure scores of Eloundou et al. (2024), and the observed-usage measure from the Anthropic Economic Index. Across eight specifications, the interaction with summer 2026 is positive every time and statistically significant none of the time. [Fact]

The one measure that does show a positive link is remote-work feasibility: significant in the authors' main tables, though their own introduction calls it only marginally significant. Graduates in occupations that can be done from home fared relatively worse in summer 2026; those in jobs that cannot be done from home fared relatively better. That fits earlier evidence that remote work weakens the informal training junior staff rely on. [Fact] It does not settle the question, because in this sample remote-work feasibility correlates with observed AI exposure at 0.58 and with theoretical exposure at 0.66. Software developers, accountants and financial managers score as fully remote-capable; registered nurses and pharmacists do not. Those are largely the same desk jobs that exposure indices rate highest. [Fact]

A structural problem cuts deeper. You can only match a person to an occupation if they report one, and many new graduates without work have never held a job. The paper loses 15.5% of the unemployed and 35.8% of the expanded sidelined group from the occupation analysis for exactly this reason. [Fact] The group most likely to be hurt by firms quietly hiring fewer juniors is the group least visible in any occupation-level test, in this paper and in any dataset built on occupation counts.

That is also why this study and payroll-based findings can both be right. ADP's Canaries Dashboard shows employment of 22-to-25-year-olds in high-exposure occupations falling year over year for 35 consecutive months. Payroll data counts filled jobs at the firms that use ADP. The CPS counts people, including those who never got hired. Fewer young people in AI-exposed occupations does not have to mean more unemployed young people, if they end up in other work, stay in school, or are counted as sidelined. [Claim] The CPS detail backs this up: less than 17% of unemployed recent graduates lost a job through a layoff. Most simply had not found one yet, so job-loss data such as unemployment insurance claims will mostly miss them. [Fact]

Why this matters if you are job hunting in 2026

If you graduated this year and are still looking, this paper does not say your search is easy. A 7.3% summer rate, or 10.4% counting the sidelined, is hard. What it says is that the class of 2026 is not doing measurably worse than the classes of 2022 through 2025 did at the same point in the year. Much of the summer pain is the ordinary cost of arriving in a market all at once. [Estimate]

Three things follow from the data.

First, compare like with like. When you see a scary number about new graduates, check whether it is a summer month compared with a summer month. The seasonal jump is larger than any year-to-year change in this series.

Second, the tell to watch is not the headline rate but the sidelined share. If the gap between "unemployed" and "wants a job" keeps widening in the summer of 2027, that is an early signal the official rate would miss.

Third, the remote-work finding is worth taking seriously on its own terms. Whether the cause is AI, remote work, or both, early-career workers in fully remote roles did relatively worse. If you can choose, a role with in-person contact with experienced colleagues may buy you the on-the-job learning that junior positions used to provide by default. [Claim]

You can see how AI exposure is measured for some of the roles the paper names on our pages for software developers, computer programmers and financial managers.

The authors end on a caution we share. Their result covers one cohort and one summer. If workplace AI use keeps deepening, they write, the classes of 2027 and later could face a different market, and measuring it will require counting not only the jobs firms add or cut but the graduates who never get a foothold at all. [Claim] For a different angle on the same cohort, see our write-up of the Dallas Fed study of graduates from AI-exposed majors in Texas.

Sources

• Robert W. Fairlie and Jane Wu, "The Early Impacts of AI on Employment among Recent College Graduates," NBER Working Paper 35796, September 2026, DOI 10.3386/w35796 — https://www.nber.org/papers/w35796
• The same paper circulated as CESifo Working Paper No. 12994 (dated September 15, 2026), which was the full text we read

About this analysis

This article was written with AI assistance. The NBER full text was not accessible to us at publication, so every figure attributed to the authors comes from the identical CESifo Working Paper version, cross-checked against the abstract on the NBER page. The four-summer average, the sidelined add-on by year, the decomposition of seasonal versus year-to-year movement, and the 95% ranges built from the reported standard errors are our own arithmetic on the paper's published figures and are labeled [Estimate]. The reading of the payroll-versus-survey gap and the advice for job seekers are our interpretation, not the authors'. The NBER page states the authors have no disclosures and received no funding for this research. The paper uses the Anthropic Economic Index as one of its exposure measures; this article was drafted with a model made by Anthropic. No tables or figures from the paper are reproduced.