The Earnings Record Cannot See the AI Shock Yet

What has happened to recent graduates since 2022

The American Community Survey is the only public source that records field of degree for graduates who finished after November 2022. The 2019 and 2021 through 2024 one-year microdata were restricted to respondents holding exactly a bachelor's degree, not enrolled, aged 22 to 24, the closest survey approximation to the Census paper's new entrants. Majors were grouped by the Opportunity Data AI Exposure Index, carried from academic programs to the survey's 174 field-of-degree codes through a hand-built crosswalk, with deciles weighted by graduate counts.

The top 10 percent of majors by AI exposure, which is the computing fields, computer and electrical engineering, and a handful of quantitative social sciences, tracked the bottom 60 percent through 2022. The two series then separated. Under the Census paper's own exposure grouping the top decile moves from 5.5 to 9.0 percent over the same years, so the result does not depend on which measure assigns the majors.

Unemployment rate among bachelor's-only graduates aged 22 to 24, not enrolled. The Census paper grouping uses the Eloundou GPT-4 exposure deciles reported in the paper's Appendix Table A3. Sample sizes per cell are 2,000 to 2,700 respondents in the top decile and 9,200 to 10,000 in deciles 1 to 6 under the Opportunity Data grouping.

Pay for those who found work did not fall. Among full-time, full-year workers aged 22 to 27 in the top 10 percent of majors ranked by the Opportunity Data AI Exposure Index, pay at the bottom, the lower quarter, the middle, and the upper quarter of the group is within three percent of its 2019 level in constant dollars, and pay at the top tenth of the group rose ten percent, all of it in 2024. The adjustment is on the employment margin: graduates who got jobs in these fields were paid as before or better, and fewer got jobs. The Census paper's decomposition points the same way, with part of its earnings loss coming from graduates moving into lower-wage sectors rather than from lower pay within the field.

Annual wages inside one group: full-time, full-year workers aged 22 to 27 holding exactly a bachelor's degree whose majors are in the top 10 percent ranked by the Opportunity Data AI Exposure Index. Each column is a point in that group's pay distribution, not an exposure decile. 2024 dollars.

What the public record shows for the same fields

The Census Bureau's Post-Secondary Employment Outcomes series is the public program-level earnings record. Its latest bachelor's cohort graduated in 2019 to 2021. Grouping its 4-digit program cells by Opportunity Data AI Exposure Index decile and taking the graduate-weighted median of cell medians gives the following history of one-year earnings, in constant 2023 dollars.

Across the two decades the record covers, the most exposed decile gained a third in real one-year earnings while the middle deciles stayed flat, and the same ordering holds five and ten years out. Within that top group, earnings also compressed: the 25th percentile rose 39 percent and the 75th rose 17 percent, so the floor rose faster than the ceiling. Every cohort the accountability system can currently see was formed in the strongest sustained period computing graduates have had. The Census paper dates the break to the first cohort the record does not yet contain.

Students are already moving

The National Student Clearinghouse Research Center's preliminary fall 2026 enrollment report, released October 29, 2026, shows the field the record rates most favorably losing students faster than any other. Enrollment in Computer and Information Science programs fell 12.4 percent at four-year institutions, 12.3 percent at two-year institutions, and 10.4 percent at the graduate level in a single year, while total undergraduate enrollment rose 0.6 percent, Health rose at every level, and the skilled-trades field the Clearinghouse reports, mechanic and repair technologies, rose 6.1 percent. Community College Daily's same-day report on the release, drawing on the Clearinghouse dashboard, puts the two-year decline at 18.4 percent for undergraduate computer science at four-year institutions and 17.5 percent at two-year colleges, after a 5.9 percent drop at two-year colleges last fall.

National Student Clearinghouse Research Center, Preliminary Fall 2026 Enrollment Trends, October 29, 2026: about 7.92 million enrollments reported as of September 24, 2026 by 46.0 percent of participating institutions; fields are two-digit CIP families; estimates are preliminary and final figures are due in January 2027. Two-year cumulative changes and the mechanic and repair technologies figure are as reported by Community College Daily from the Clearinghouse dashboard.

The Clearinghouse report offers no explanation for the decline, and this page does not estimate one. The timing is what matters for the accountability design. Enrollment in the most exposed field fell in the same two years in which unemployment among its recent graduates doubled, which is consistent with students responding to conditions the earnings record does not yet show. A program that loses an eighth of its students a year will fall below the 30-completer threshold sooner, and when it does, the cohort-expansion rule scores it on completers from the seventh prior award year. The test will reach these programs years after their students did, and will judge them on graduates from before the change.

The first program-level earnings after the break

The College Scorecard field-of-study files are the one public program-level series with an earnings year after November 2022. The October 15, 2026 release carries three one-year earnings cohorts for bachelor's programs: completers of award years 2017-18 and 2018-19 measured in 2019 and 2020, completers of 2018-19 and 2019-20 measured in 2020 and 2021, and completers of 2020-21 and 2021-22 measured in 2022 and 2023. The last of these is the first program-level earnings measurement that reaches past the break, for graduates who entered the market just before it. Restricting to programs present in both the first and last cohorts, deflating to 2024 dollars, and weighting by completers working gives the change in real one-year median earnings within the same programs, by Opportunity Data AI Exposure Index decile.

Median of program medians weighted by completers working, real 2024 dollars; the Scorecard adjusts each cohort to the year after its last measurement year and the series here re-deflates to 2024 with CPI-U. The least exposed 10 percent of programs is dominated by nursing, which is why its level is high. Columns are groups of programs ranked by the Opportunity Data AI Exposure Index. The within-program row matches programs on institution and 4-digit CIP across the first and last cohorts (1,166 to 3,597 programs per group).

Across the three cohorts the most exposed 10 percent of programs is the only group whose real one-year median fell, from $72,000 to $69,700, while the groups between the 10th and 50th percentiles of exposure gained between 5 and 12 percent. Within the same programs the top 20 percent were flat and every other group gained, so the most exposed programs lost between 4 and 14 points of real one-year earnings relative to the rest. The Scorecard measures a full calendar year of earnings for Title IV aided completers, so this is a wage-and-hours margin rather than the survey's unemployment margin, and the most recent cohort is only half post-break. It is nonetheless the first time the program-level record has moved in the direction the Census paper describes, and the record that moved is the one the earnings test will read next.

Why the earnings test cannot register it

The One Big Beautiful Bill Act of 2025 created a low-earning-outcome test for degree programs, and the Department of Education's final rule of July 1, 2026 (91 FR 40136) set the mechanics. A program's completers must have median earnings, in the fourth tax year after completion and counting only those who work, above the median for adults aged 25 to 34 in the same state with only a high school diploma (34 CFR 668.403). A program that fails in two of any three consecutive award years loses Direct Loan eligibility (34 CFR 668.603). Three features of that design determine when, if ever, the test will register the change in the section above.

1. The measurement point. The fourth tax year after completion means a cohort entering the market in 2023 is measured in 2027 and scored in 2029. The first cohort that graduated entirely after ChatGPT, award year 2023, cannot be scored before the 2029 calculation; two failing years out of three puts any consequence at 2031 or later. For the agentic systems of spring 2026 the arithmetic runs one step further: the first cohort entirely after that break, award year 2026, is measured in tax year 2030, scored in 2031, and can face a consequence no earlier than 2033.
2. The cohort expansion. When a program has fewer than 30 completers, the cohort period reaches back to the fifth, then the sixth and seventh prior award years, and then pools programs in the same 4-digit CIP and credential level (34 CFR 668.2). A small computer science program evaluated in 2029 can be scored on completers from 2019.
3. The exclusion of non-workers. The survey evidence and the Census decomposition agree that the margin that moved first is employment. A median over workers only is, by construction, unresponsive to a rise in the share who are not working, and the benchmark on the other side of the comparison, the state median for high-school-only workers aged 25 to 34, is also a median over workers. The test is a floor on pay conditional on having a job. It was not designed to detect a change in the odds of having one.

Lag arithmetic under 34 CFR 668.403(b)(1) (fourth tax year after completion) and 668.603(a) (two failing years of any three consecutive), applied to the Department's published timeline, which scores award year 2021 completers on tax year 2025 earnings in early 2027. Earliest consequence assumes failing results in the first two years scored.

The rule already contains one acknowledgment that an earnings year can be non-comparable. Programs preparing students for tipped occupations receive neither a pass nor a fail for any year that uses tax-year 2025 or earlier earnings, because the federal tax treatment of tips changed (34 CFR 668.402(c)(3)). The Department has accepted, in principle, that a policy shock makes a measurement year unrepresentative.

How to read these numbers

• The exposure measure is not a task-exposure score. The Opportunity Data AI Exposure Index scores digital and analytical work intensity reported by incumbent workers, with human contact and physical anchoring as protective structure; it does not estimate the share of tasks a language model can perform. At the field-of-degree level it correlates 0.71 with the Eloundou GPT-4 measure the Census paper uses, and it explains about half as much of the paper's realized 2024 employment effects. Eight of its seventeen top-decile fields, including every computing field, are also in the Census paper's top decile; sociology, mechanical engineering, materials engineering, and international studies are in the Opportunity Data top decile and the paper's middle deciles. Results are reported under both groupings for that reason.
• The survey sample is not the Census paper's sample. Graduates aged 22 to 24 include people up to three years out, which attenuates an entry effect. The employment-rate difference-in-differences, top decile against deciles 1 to 6 and 2024 against 2022, is 3.3 points under the Opportunity Data grouping and 1.8 under the Census grouping, between a third and two thirds of the paper's quarter-after-graduation estimate. Widening the age band to 22 to 27 gives 1.4 and 1.0 points.
• A national study finds no pooled spike, and the two results are consistent. Fairlie and Wu (CESifo Working Paper 12994, September 2026) use Current Population Survey data on all bachelor's-only graduates aged 22 to 25 and find that unemployment in June through August 2026 did not rise significantly against the 2022 to 2025 summer average or against older graduates and young non-graduates; the summer 2026 coefficient is −0.8 points with a standard error of 0.7. Three things separate that result from this page. The CPS does not record field of degree, so a doubling confined to the top 10 percent of majors, with the bottom 60 percent moving half a point, pools to a change inside their standard error. Their baseline is the 2022 to 2025 summers, which already contain the 2023 and 2024 divergence shown here, so a level shift that had happened by 2024 is part of their control. And their exposure test sorts graduates by the occupation they hold, not the major they finished, so a computing graduate who took a retail job counts as unexposed. Their window, summer 2026, is also two years past the last survey year here; when the 2025 and 2026 ACS files arrive, this page will report whether the field-level gap held.
• Unemployment is one margin. The survey cannot distinguish a graduate who is unemployed from one who has left the labor force, moved into a job outside the field, or delayed entry for graduate school. Employment-rate changes are reported in the methodology note alongside the unemployment rates.
• PSEO cohorts are three-year bins and cover participating institutions only. Earnings are real 2023 dollars as published, and the series is the graduate-weighted median of 4-digit program cell medians, not a national median. Institutions that joined PSEO late can shift the composition across cohorts.
• The spring 2026 break is a dated assumption, not a measured effect. This page does not estimate the labor-market effect of agentic systems. It records that no public data yet observes a cohort that graduated after them, and that the Opportunity Data AI Exposure Index is built on that later frontier while the Census paper's measure is built on the earlier one.

Versioning and next validation

This page uses ACS one-year PUMS through survey year 2024, PSEO release R2026Q2, the College Scorecard field-of-study file of October 15, 2026, and the September 2026 release of the Opportunity Data AI Exposure Index. The enrollment figures are the Clearinghouse's preliminary fall 2026 estimates and will be replaced with the final January 2027 figures. The Scorecard series will be extended when the next field-of-study release adds the 2021-22 and 2022-23 completer cohort, the first measured entirely after the break. It will be rerun when the 2025 ACS one-year file is posted, which adds the first full survey year after the Census paper's data, and again when the 2026 file is released in late 2027, the first public observation of graduates who finished after the spring 2026 break. The next validation is the 2025 rerun: if the top-decile unemployment gap of 2024 persists or widens in 2025 under both groupings, the employment-margin finding stands; if it closes, this page will say so.

Methods and sources

Read the methodology note (PDF) →

@techreport{rowe2026earningsrecord,
  author      = {Rowe, Benazir},
  title       = {The Earnings Record Cannot See the AI Shock Yet: Recent Graduate Outcomes by AI Exposure and the Federal Earnings Test},
  institution = {Opportunity Data},
  year        = {2026},
  month       = {October},
  url         = {https://opportunitydata.org/recent-grad-ai-disruption.html},
  note        = {Methodology note: https://opportunitydata.org/briefs/recent-grad-ai-disruption-methodology.pdf}
}