Research update · 18 September 2026
The national graduate AI unemployment alarm just met a hard counter-signal.
A new U.S. CPS study finds no statistically significant summer-2026 unemployment spike among recent college graduates. That is a serious brake on the claim that AI has already produced a national graduate unemployment crisis. It is not proof that the entry-level market is safe.
What the new study measured
Fairlie and Wu use U.S. Current Population Survey microdata for June, July and August 2026. They compare recent college graduates with prior summers, older college graduates and young adults without a college degree. In difference-in-differences and event-study interaction models, they report no statistically significant relative unemployment increase. The result remains null when they add people described as “sidelined unemployed.”
What the counter-signal actually rules out
It weakens a very specific story: that the first summer of 2026 produced a clear, broad relative unemployment surge for recent U.S. graduates. A public claim that AI has already made graduates nationally unemployed cannot treat the recent-graduate unemployment rate as settled proof.
That is not a reassuring forecast. It is a measured short-window result about unemployment status, not a count of job offers, time to first job, earnings, underemployment or people who left the labor force for reasons the study cannot assign to AI.
Why it does not erase the other entry warnings
A Census administrative-record study finds that graduates in the most AI-exposed majors had lower initial employment and lower full-quarter initial earnings after ChatGPT’s release. A Texas study finds a regional pullback in more-automatable online postings. Those findings concern a first match, pay and advertisements—not the CPS study’s summer unemployment comparison. Different outcomes can move differently without either study being false.
They also do not supply a common causal test. The new CPS paper does not observe employer generative-AI adoption. Its occupation interactions find some evidence that remote-work availability is positively related to 2026 unemployment, which makes a simple exposure-is-AI story less secure.
The limit is as important as the relief
Neither the new CPS result nor the Census graduate study can calculate your chance of finding work. Their samples, cohorts, windows, outcomes and comparison groups differ. Exposure scores are not firm adoption records, and a result after ChatGPT’s release can still reflect education, remote work, macro conditions or other changes that hit exposed work differently.
The claim to demand before acting
Ask whether the source measured unemployment, a first job, an advertised opening or pay—and whether it observed AI use at all. “No clear summer unemployment spike among recent U.S. graduates” is defensible. “AI has not made entry harder” is not.