AI-Exposed College Majors at Labor-Market Entry

Sources#

Summary#

The vault's entry-level evidence measures workers by the job they hold: ADP payroll by occupation (Canaries), postings by occupation (Indeed), Revelio seniority at adopting firms (Seniority-Biased AI Adoption: The Junior Share at Adopting Firms). All of them condition on an outcome. A graduate who wanted a software job and ended up stocking shelves drops out of the exposed occupation's count and adds nothing to its pay statistics. Orr, Tucker & Warren, Graduating into Disruption (US Census Bureau CES working paper 26-56, September 2026, empirical) fixes exposure by college major, before graduation, and follows each graduate through employment, non-employment, industry and job switches.

Against the bottom six deciles of exposure, and relative to 2022:

Initial employment (any earnings in the quarter after graduation) falls ~5pp for the most-exposed decile. Deciles 7 and 8 fall about 2pp.
Initial earnings (log full-quarter earnings in the dominant job, two quarters out) fall ~13% for the top decile. By 2025 the gap is 14 log points. The paper puts this "comparable in magnitude to the effects of entering the labor market during a large recession," which is 9–10% in the annual-earnings literature. The measures are not directly comparable (quarterly and conditional on full-quarter employment).
About half the earnings loss is sorting. The top decile's share in Information fell by more than 3pp and in Professional, Scientific & Technical Services by nearly 6pp. Its share in Accommodation & Food Services and in Retail Trade rose by more than 2pp each. Less-exposed graduates' industry mix barely moved. Without the industry shift, top-decile earnings in 2024 would have fallen nearly 8%. With it, they fell 15%, to levels last seen before 2016.
Persistence splits by decile. Deciles 7–9 are back near zero on earnings within one to two years. The top decile is still −4.2pp employment and −8.1 log points earnings at four quarters, and −2.0pp and −3.3 log points at eight quarters (2024 onward averages).
Timing. The employment coefficients are flat before 2020 and break right after November 2022.

The top decile is "composed largely of computer science, information systems, and other degree fields closely related to software development." In the major-level regressions (Table A3, 2024 coefficients, verified against pdftotext -layout), Computer Science is −0.061 employment and −0.190 log earnings, Computer Engineering −0.068 / −0.120, and Computer & Information Systems-General −0.054 / −0.142. Setting aside a couple of very small fields, those three majors have both the largest employment declines and the largest earnings declines.

Evidence note. empirical, in the voice of a government statistical agency. The data are administrative: PSEO graduation records linked to LEHD unemployment-insurance wage records covering over 95% of jobs in participating states. The event study uses institution×major, institution×quarter and major×quarter-of-year fixed effects, with SEs clustered by major. It is a CES working paper, which the Bureau says has "not undergone the review accorded Census Bureau publications." Four limits travel with every number. (1) Sample. 6.67M bachelor's graduates from 27 PSEO partners (over 350 institutions in 24 states), about 29% of US bachelor's degrees in 2016–2024. The institutions are 92.7% public, against 53.4% nationally, and skew toward large, urban, research-intensive schools. Alaska and Michigan are excluded. (2) Employment means UI-covered wage work. Self-employment and graduate school count as non-employment. (3) Exposure is Eloundou GPT-4 β mapped to majors through ACS occupations of 21–29-year-old bachelor's holders in 2018–2022, fixed at that pre-period. (4) The post-period is short. Only the last few quarters of the longer-horizon outcomes include post-ChatGPT graduates. Co-author Tucker's own QWI paper (CES-WP-26-27) is cited for the industry-concentration and monetary-policy checks, so those are in-house corroborations rather than independent ones.

Why fixing exposure at the major matters#

The paper claims to reconcile two literatures. Occupation and industry studies of hiring find large declines for young workers in exposed jobs (Brynjolfsson, Chandar & Chen; Hosseini Maasoum & Lichtinger; Tucker). Studies following incumbents in exposed occupations find little extra unemployment (Eckhardt & Goldschlag at EIG; Massenkoff & McCrory at Anthropic). Both kinds of study assign exposure by "an occupation or industry that is itself an ex post labor market outcome." Graduates from exposed majors "remain strongly attached to the labor market, but they adjust to weaker demand at entry": later employment and worse sectors. So "modest employment or unemployment effects can coexist with substantial deterioration in job quality and early-career opportunities."

The same argument settles a wage puzzle the vault has left open. Canaries Fact 6 finds young new hires' starting pay unrelated to occupational exposure. Indeed finds the entry-level offer gap in exposed occupations at ≈2 points. Here, the top decile's initial earnings fall 13%. These are consistent, because they measure different things:

Instrument	Conditions on	Entry-level pay finding
Canaries (ADP base pay)	Being hired into the exposed occupation	No relationship with exposure
Indeed (advertised offers)	A posting in the exposed occupation	≈2 points cumulative at entry
This paper (LEHD earnings by major)	Only the pre-college field of study	−13% at two quarters, about half from industry sorting

The price of an exposed entry job has not fallen much. Fewer graduates get one. The rest take jobs in lower-paying sectors, and occupation-conditioned designs cannot see that loss. The within-industry half (about 8% at 2-digit NAICS) is not necessarily a pay cut for the same work either. A 2-digit sector still mixes occupations, and the paper does not observe occupation in LEHD.

The measure question: mean exposure hides the tail#

A major's mean exposure can hide how much of its occupation mix sits at the extreme. The paper's example: Journalism and Accounting have nearly the same mean, but Journalism has 36% of its occupational distribution in the top occupational exposure quintile and Accounting has 16%. Every major has some top-quintile share, and 82% have at least 10%. Within deciles 7–9, where the two measures correlate only 0.16–0.29, a 10pp higher top-quintile share goes with 2.3, 4.0 and 4.8 log points lower initial earnings. The decile means are only −0.7 to −1.3 log points. Firms substitute at the extreme of exposure, not on average.

Alternative exposure measures agree up to the 80th percentile and split at the top (Appendix B):

The Eloundou and Eisfeldt measures both rank computer-science majors above Accounting and Finance. Both show a drop below trend around the 90th percentile.
Handa et al.'s usage measure (Claude traffic; the AEI lineage) reverses that order, putting business majors above computer science. Its smoothed curve dips and then returns to trend.
The automation/augmentation split does not separate majors either. In the top half of the exposure distribution, Handa's "automative" share of prompts sits in a tight 0.33–0.41 band, and business majors (high exposure, small effects) and computing majors (high exposure, large effects) cannot be told apart on it.

The intro's claim that the results are "robust to adopting an alternative measure" holds for the decile design. It does not hold for the ranking at the very top, which is where the effect is.

This is in tension with Canaries. There, the automation coefficient on AEI usage is the strongest age-graded predictor (−0.098 per SD at ages 22–25; see Codified vs Tacit Knowledge Exposure). Here, the automative share carries no signal across majors. Resolution (partial): averaging over a major's occupation mix compresses variation. The paper shows this for exposure itself: the 90th-percentile major (0.50) sits at about the 60th percentile of occupations (0.49). The automative share gets squeezed in the same way. So this is a low-power null at the major level, not a refutation at the occupation level. Both sources are empirical. For the automation/augmentation question, Canaries' occupation-level design is the better instrument.

Alternative explanations the paper tests#

Remote work. Lambert & Schindler (2026) argue that most "AI" declines are really work-from-home, and WFH correlates 0.88 with AI exposure at the major level. With a time-varying ACS WFH control added, the post-period coefficients are "not practically or statistically different" from the baseline. A replication of Lambert & Schindler's specification has visible collinearity (pre-period coefficients diverge in opposite directions) but leaves the gap against the long-run average unchanged. The authors concede that neither approach resolves the collinearity. The CPS evidence points the other way on the same collinearity: among recent graduates in summer 2026, only teleworkability is significant, not either AI measure (see the CPS section above).
Oversupply. Top-decile graduates grew 24% from AY2017 to 2022, and employment kept pace. From 2022 to 2024 graduate growth slowed to under 3% a year, while employed graduates fell 3%. No other decile shows that break.
Self-employment and graduate school. The ACS sample is ages 21–24, so it is not the PSEO sample. For the top decile, self-employment rose 1.01% → 1.49% and graduate-school enrollment rose 16.20% → 17.86%, while the bottom six deciles barely moved. Taken at face value, these "may explain up to half" of the 5pp employment drop. Self-employed top-decile graduates earn less than a third of their wage-employed peers, which reads as weak demand rather than new opportunity. This is the paper's largest unreconciled caveat on the employment number. It does not touch the earnings number, which conditions on employment.
Monetary tightening and the tech correction. Tucker (2026) finds that monetary shocks explain at most about a quarter of the early-career decline and do not reproduce its timing. Hiring declines tied to AI exposure also span many sectors. Even so, the effect is concentrated in software-adjacent majors, and the fixed effects "do not account for time-varying shocks to particular fields." A computing-specific shock is the confound this design handles worst.

The CPS null: no 2026 unemployment break for recent graduates#

Fairlie & Wu, NBER w35796 (UCLA, September 2026, empirical) looks at the same entry moment through the Current Population Survey. Their sample is ages 22–25 with a bachelor's as the highest degree, not enrolled in school, observed January 2022 – August 2026. They find no statistically significant rise in unemployment in summer 2026, the first graduating class they argue entered after workplace AI use deepened (Ramp median AI spend per employee more than doubled December 2025 → July 2026).

Levels. June–August unemployment was 7.3% in 2026, against 7.1% (2022), 6.3% (2023), 7.8% (2024) and 7.2% (2025). Summer runs about 2pp above the rest of the year, so any comparison has to be summer against summer.
"Sidelined" unemployment. The paper adds people outside the labor force who say they want a job. The expanded rate was 10.4% in summer 2026, the highest of five summers but only 0.3pp above the previous high (10.1% in 2024). It still shows no significant break.
Regressions. Against the de-trended 2022–25 summer average, the summer-2026 coefficient is −0.008 (SE 0.007) standard and 0.001 (0.009) expanded. The event-study coefficients against summer 2022 are also null. Two difference-in-differences designs are null as well: against college graduates aged 30–49 (−0.007 to 0.004) and against young non-graduates (0.002 to 0.013). The non-graduate comparison gives the largest point estimates, partly because non-graduate unemployment fell (7.5% → 6.8%). On the expanded measure, the graduates' advantage over non-graduates shrank from 2.2pp to 0.3pp, the smallest in five summers.
Occupational exposure. Interactions with the AEI observed-exposure measure and with Eloundou's theoretical measure are positive in all eight specifications and significant in none. Teleworkability is the only occupational measure that carries signal: +0.028 (p<0.1) with a linear trend and +0.040 (p<0.05) in the event study. The paper's prose calls this "statistically significant" in §5 and "marginally significant" in the introduction. Teleworkability correlates 0.58–0.66 with the two AI measures, so the authors do not claim to have separated the two explanations.

Why this does not contradict the Census result. The Census paper reports a −5pp initial-employment break for the top decile of majors. Both papers are empirical, and they measure different things in four ways:

Dimension	Census PSEO×LEHD (this page)	Fairlie & Wu (CPS)
Outcome	Any UI-covered earnings; earnings and industry	Unemployment (active search), plus "wants a job"
Who is affected	The top decile of majors by exposure, mostly computing	All bachelor's holders aged 22–25
Graduate school	Counted as non-employment	Enrolled people are dropped from the sample (22.5%)
Self-employment	Counted as non-employment	Counted as employed
Job quality	Measured: −13% earnings, sorting into retail and food service	Not tested (mismatch proxy 17.1% vs 15.6% in 2022, descriptive only)

Different outcomes. The Census paper's own caveat is that self-employment and graduate school "may explain up to half" of the top decile's 5pp drop. Those are the two channels the CPS design cannot see: the self-employed count as employed, and new enrollees leave the sample. The rest of the Census loss is lower earnings and sorting into other sectors, and unemployment does not measure either. The Census paper predicts exactly this pattern: "modest employment or unemployment effects can coexist with substantial deterioration in job quality."
Different populations. A 5pp effect on about a tenth of graduates is, as a rough upper bound, about 0.5pp in the aggregate. That is inside the CPS standard errors of 0.7–0.9pp. The CPS test of occupational exposure, the finer cut, also points in the positive direction, just without significance. It also drops the graduates most likely to be affected: 16% of the unemployed and 36% of the sidelined report no occupation, because they have never held a job.
Different instruments than Canaries. Canaries measures the relative ADP headcount of 22–25-year-olds in the most-exposed occupations against the least-exposed (a 19% shortfall by June 2026), mostly through hiring. A shift of young workers between occupations moves that ratio without moving the unemployment rate at all. Fairlie & Wu cite Canaries' age band as the reason for theirs, and their conclusion concedes the point: aggregate occupation counts "might not adequately capture the denominator."
Timing only partly explains it. The linear-trend specification absorbs a gradual rise that started in 2023. The event study compares summer 2026 with summer 2022, which is before ChatGPT, and is still null. So timing does not explain the null away. In that same event study, fall 2025 (+2.3pp, p<0.01) and winter 2025 (+1.2pp, p<0.05) were elevated relative to 2022, and the paper's prose does not discuss either result.

Net reading. Through summer 2026, AI's effect on US graduate entry shows up in which jobs graduates get and what they pay (Census, Canaries, Indeed). It does not show up as a rise in unemployment among graduates overall. The CPS null limits how large the headcount story can be in aggregate. It does not overturn the occupation-level and major-level findings. The authors call their results descriptive, not causal, and say the classes of 2027 and later "might be more affected."

Mechanisms named, none identified#

The paper lists four and settles none:

Direct substitution of entry-level tasks.
Job transformation: workers specialize in the remaining tasks, subject to their comparative advantage and AI's user cost.
Lost training. "If AI automates the work through which junior employees acquire tacit knowledge, firms may hire fewer new workers, reduce training, or reorganize career progression" (Garicano & Rayo 2025; Ide 2025). This is The Tragedy of the Cognitive Commons's Mechanism 1 in a Census paper's words.
Signal erosion. AI raises grades and compresses their distribution, so employers can no longer read skill from grades, assignments or work samples. Chirikov (2026) finds grades rise most in AI-exposed courses where homework carries more weight. The paper offers this as the reason graduate effects could precede firm adoption (Bonney et al. 2024), a timing puzzle Seniority-Biased AI Adoption: The Junior Share at Adopting Firms also meets in its late cohorts. It connects the classroom offloading evidence on Experimental Learning Impact of Generative AI to hiring: the same automation of assessed work that suppresses learning would also degrade the credential.

The job-switching result is the one that bears on career ladders. For pre-ChatGPT graduates who reached their two-year mark after the release, job switching fell to COVID-era lows (the average coefficient for the top four deciles is −5.9pp in 2024). It then rose quickly for post-ChatGPT graduates (−2.7pp in 2025). That fits a pattern where graduates take weak first matches and then leave them. The top decile's two-year earnings gap suggests those moves have worked less well for it.

What it does not measure#

Developmental content. A move into Retail Trade is the nearest thing yet to a measure of what first jobs are, and it is a sector code. Nothing here observes whether a first job teaches.
Occupation. LEHD has no occupation field. Exposure comes in through ACS, and outcomes are read by industry.
Other degree levels and institution types. The sample is bachelor's only. Heterogeneity by institution type is left for future work.
Major switching. Students who changed majors after ChatGPT could bias results in either direction. The authors plan to measure it.
Recovery. "We cannot yet assess when or whether exposed graduates will catch up."

Connections#

The Tragedy of the Cognitive Commons — the entry-slot argument measured at the moment of entry. The top decile's losses persist at two years while deciles 7–9 recover, and the paper names Mechanism 1 (lost tacit-knowledge training) as a candidate cause without testing it
Seniority-Biased AI Adoption: The Junior Share at Adopting Firms — the other side of the same event. At adopting firms the junior share falls mostly because seniors grow. Here, the juniors who did not get in show up in lower-paying sectors. Both papers find effects that appear to precede firm adoption
The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay — this design fills in the entry-level row of that page's reconciliation. Offers and realized starting pay in exposed occupations barely move, while graduates of exposed majors lose 13%, because the loss happens in which occupation they land in
Codified vs Tacit Knowledge Exposure — a field of study is the codified-knowledge credential, and this design holds the schooling level fixed (bachelor's only) while varying the field. Exposure still predicts outcomes, so the age gradient is not just years of schooling. The major-level automative share, though, carries none of the signal Canaries finds by occupation
Experimental Learning Impact of Generative AI — the signal-erosion channel. The classroom offloading that the RCTs show suppresses unaided learning is, on the paper's hypothesis, also what erodes the grade signal employers hire on
Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — three instruments (Eloundou, Eisfeldt, Handa) scored against the same realized entry outcomes. They agree through the 80th percentile and disagree about who is at the top
Anthropic Economic Index — the Handa usage measure is the AEI's first release. It moves business majors above computing majors, and there the top-decile dip disappears
Task Saturation: Broad but Shallow AI Diffusion — Canaries Fact 6 (adjustment through headcount, not pay) is reinterpreted here. Pay adjusts too, through sorting into other sectors, which an occupation panel does not record

Open Questions#

Does the top decile's earnings gap close by years three to four, or does it scar the way recession cohorts do (Oreopoulos et al.)? At eight quarters it is −3.3 log points. Deciles 7–9 have already recovered. Trigger: a later PSEO×LEHD vintage in which the 2023–2024 graduating cohorts reach 12–16 quarters after graduation.
Is the top-decile effect caused by AI, or by a computing-specific shock (the tech-sector correction) that happens to line up with the most exposed majors? Handa's usage measure moves business majors to the top, and there the drop vanishes. Falsifiable: an event study restricted to non-computing majors in the top quintile of any exposure measure. If Accounting, Finance, Journalism and Marketing graduates show no break at November 2022, the "AI-exposed majors" result is a computing result.
Does the grade-signal channel operate in hiring? Falsifiable: link course-level AI-driven grade inflation (the Chirikov / Hausman designs) to graduates' initial employment, within major. If graduates from programs whose grades inflated most see larger hiring declines at the same exposure, signal erosion is part of the mechanism.
Does unemployment finally break for later graduating classes? Fairlie & Wu's CPS null covers only the class of 2026, and the authors expect the classes of 2027 and later "might be more affected." Falsifiable: a significant summer-2027 coefficient, on the standard or the sidelined measure, for 22–25 bachelor's holders against 2022–26 and against young non-graduates. If a later CPS vintage stays null while the Census top-decile gap persists, adjustment through sorting, graduate school and self-employment is the whole story at entry. Trigger: CPS basic monthly files for June–August 2027.

Sources#

Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors — Cody Orr, Lee C. Tucker & Lawrence Warren (US Census Bureau), Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors, CES Working Paper 26-56, September 2026 (month only), 79pp, empirical; not reviewed to Census publication standard. Cited: Abstract and §1 (headlines, reconciliation argument), §2 (PSEO sample, Eloundou GPT-4 β, ACS major→occupation mapping, LEHD), §3 (Eq. 1, identifying assumptions), §4.1–4.2 (Figures 3–9 as described in prose: employment, earnings, industry shares, the Figure 6 decomposition, persistence, job switching), §5–6 (major-level results, Journalism/Accounting, top-quintile-share gradients), §7 (WFH, oversupply, self-employment and graduate school, monetary policy), §8 (mechanisms, recession comparison), Appendix B (Handa and Eisfeldt comparison, automative share). Table A2 (institution characteristics) and the Table A3 computing rows were cited after checking against pdftotext -layout. Parse warning: docling replaced the major names in the legend tables under Figure 10 and the A-series figures with private-use glyphs, so those legends are unreadable in the raw (the PDF text layer has them). None are cited. The PDF font drops "fi"/"ff" ligatures in body text ("fnd", "fve"). All figure values quoted here come from the prose, not from chart readings
The Early Impacts of AI on Employment among Recent College Graduates — Robert W. Fairlie & Jane Wu (UCLA), The Early Impacts of AI on Employment among Recent College Graduates, NBER Working Paper 35796, September 2026, 49pp, empirical. Not peer-reviewed; no funding and no disclosures. The authors describe it as descriptive, not causal. Cited: Abstract, §1, §3 (sample, the sidelined, underemployment and mismatch measures, occupation-missing shares, teleworkability correlations), §4 (Eqs. 4.1–4.4), §5 (summer levels from Figures 2a–4b as stated in prose, and Tables 1–6 as described in prose), §6. Table 1 event-study rows (Winter/Fall 2025) were checked against pdftotext -layout. Parse warning: docling shifted or split cells in Tables 2 and 3. Table 3 is cited only from the > [!note] transcription block below it in the raw, and Table 2 only from the prose. Table 1 fused Fall 2025 with the Observations row. Table 4 dropped its Summer 2023 × AI exposure row (−0.002 / −0.003, SE 0.008, from the PDF). Table 5 merged its Summer 2024 and Summer 2025 rows. Table 6 was hand re-split and matches the PDF. No individual row of Tables 2, 4 or 5 is cited here