Graduate data complicates the AI jobs narrative

On September 26, 2026, a new working paper from economists Robert Fairlie and Jane Wu added an important qualification to claims that artificial intelligence is already displacing new university graduates in the United States. Using Census microdata, the researchers found no evidence of significant, widespread reduction in hiring or rise in unemployment among recent college graduates through summer 2026. Their result does not describe an easy market for young people. It does, however, challenge a simple reading of the AI jobs story: difficult graduate job searches do not yet amount to clear statistical proof of broad AI-driven displacement.

A summer figure within the recent range

Fairlie and Wu examined recent college graduates defined as bachelor’s degree holders aged 22 to 25 who were not enrolled in further education. Their source was the US Census Bureau’s Current Population Survey, a broad household survey used to measure employment and unemployment. The researchers focused on summer because unemployment predictably increases when a new graduating class enters the labour market. They compared year-on-year and seasonal patterns from 2022, when employment had returned to pre-pandemic levels and ChatGPT was released, through 2026.

The summer 2026 unemployment rate for this group was 7.3%. That was neither a record nor an obvious break from recent experience: the comparable rates ranged from 6.3% in 2022 to 7.8% in 2024. The conclusion remained broadly unchanged when the researchers included graduates who said they wanted work but were not actively searching, and therefore were not counted in the official unemployed labour force. In their assessment, summer 2026 did not look unusually weak for recent graduates relative to earlier summers.

• Summer 2026 unemployment among recent graduates was 7.3%, within the 2022-2026 range.
• The study compared young graduates with same-age non-graduates and with college graduates aged 30 to 49.
• It also grouped occupations by potential AI exposure using a 2023 occupational study.
• Most differences in trends across the comparison groups were not statistically significant.

Comparisons did not isolate an AI effect

The CESifo paper tested whether the picture changed when young graduates were compared with people of the same age who did not have college degrees, or with older college graduates aged 30 to 49. It also split work by potential exposure to AI, based on a 2023 assessment of occupations whose tasks were more suited to AI systems. In nearly all of these tests, differences in trends between 2022 and 2026 were not statistically significant. That does not establish that AI had no effect on any individual employer, occupation or applicant. It means the available unemployment data did not reveal a widespread relative effect large enough to distinguish from normal variation.

The timing remains central to the argument. Fairlie and Wu reason that reductions in demand may first become visible through fewer hires for standardised entry-level office tasks, rather than dismissals of established workers. They also note signs that workplace AI adoption has accelerated: a Census survey recorded a recent increase in firms reporting that they were replacing a large number of employee tasks with AI, while AI spending per employee and ChatGPT Enterprise token use had risen over the previous 12 months. Those developments make graduate outcomes worth tracking, but they do not by themselves demonstrate job losses.

The researchers explicitly treat the 2026 result as an early test, not a forecast. If workplace use intensifies, they write, classes graduating in 2027 and later may be more affected. Several years of additional data will be needed to determine whether effects emerge as adoption deepens. This is a meaningful limitation for a technology whose capabilities and corporate deployment can change faster than annual labour-market patterns become clear.

A difficult market can have several causes

Other measures underline why the findings may feel at odds with graduates’ experience. The Federal Reserve Bank of New York reported that college graduates aged 22 to 27 had unemployment of 5.7% in June 2026, above the national average and one percentage point higher than for that cohort two years earlier. Its measure is not identical to the CESifo paper’s narrower age range and summer timing, so the figures should not be read as directly conflicting. The New York Fed also reported a 42% underemployment rate for recent graduates, meaning work in jobs not requiring a bachelor’s degree, compared with 33.7% for all college graduates.

The New York Fed has found a broader hiring slowdown in recent years, but not one specifically concentrated in entry-level jobs with greater AI exposure. Based on job-posting research, it said AI may be contributing to labour-market developments but was not the main driver of slower hiring. Employer survey work similarly suggested that firms mostly intended to incorporate AI through retraining, with limited effects on hiring. These findings leave room for AI to influence particular functions or recruitment processes while warning against assigning every weak outcome for graduates to automation.

A separate Stanford study, based on payroll data from HR company ADP, recently found that entry-level employment in occupations described as AI-impacted was lagging other fields. The apparent disagreement is partly methodological. ADP payroll records cover a substantial cross-section of employers but may omit parts of the population captured in the Census survey. The Stanford approach examines the supply of jobs in fields, while unemployment also reflects demand for those jobs. Employment in a field may contract even as shifts in labour-force participation or job-search behaviour produce a less dramatic unemployment signal.

Hiring itself is becoming more automated

For applicants, AI matters not only as a possible substitute for tasks but also as part of the route to a first interview. LinkedIn research cited by CNBC found that 66% of recruiters planned at the start of 2026 to increase their use of AI for pre-screening interviews. The same research found that 81% of job seekers had used or planned to use AI in their search. Automated filtering can make application processes feel more impersonal, even where it does not reduce the overall number of jobs.

Employers are also signalling greater demand for AI capability. The National Association of Colleges and Employers said AI skills appeared in 16.5% of job descriptions in spring 2026, up from 10.5% the previous fall, and that 28% of employers sought early-career candidates able to use AI at work. Handshake reported that graduating seniors in 2026 mentioned AI skills on résumés at twice the rate of the class of 2022; 74% of those mentions referred to real-world projects rather than coursework. Those figures describe changing recruitment signals, not proof that such skills guarantee employment.

The practical message for graduates is therefore more measured than either panic or complacency. The 2026 unemployment evidence offers no confirmed nationwide AI shock among recent US college graduates, yet it sits alongside elevated underemployment, slower hiring and more automated recruitment. For students, careers services and employers, the immediate task is to make skills and project evidence legible to both screening systems and human recruiters. For policymakers, the locally relevant US test will be whether future Census, payroll and job-posting data begin to converge on the same pattern as the class of 2027 enters a labour market shaped by deeper AI use.