AI Exposure vs. Pay by College Major (2026) - codebook
MajorMatch.us - https://majormatch.us/ai-exposure-by-major-2026/
License: Creative Commons Attribution 4.0 (CC BY 4.0). Credit: MajorMatch.us, AI Exposure vs. Pay by College Major (2026).

SOURCES AND VINTAGES
- AI exposure: Eloundou, Manning, Mishkin & Rock (2024), "GPTs are GPTs", Science 384(6702); occupation-level file occ_level.csv from github.com/openai/GPTs-are-GPTs (ratings collected 2023 on O*NET task statements). Score = 100 x mean of dv_rating_beta (GPT-4 rater) and human_rating_beta; beta = share of tasks rated E1 (directly exposed) + 0.5 x share rated E2 (exposed with additional software). O*NET-SOC sub-occupations averaged to 6-digit SOC 2018.
- Pay and employment: U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics (OEWS), national, May 2025 (national_M2025_dl.xlsx), detailed occupations.
- Majors and jobs: U.S. Census Bureau, American Community Survey 2020-2024 5-Year PUMS (person files psam_pusa-d.csv, released January 2026).
- Cross-check: Felten, Raj & Seamans (2023) Language-Modeling AIOE (LM_AIOE.xlsx, github.com/AIOE-Data/AIOE).

ai-exposure-by-major-2026.csv (one row per Census bachelor's field of degree with >= 1,000 sample records; 151 rows)
- rank_ai_exposure: 1 = most exposed.
- major_acs_field_of_degree / acs_fod1p_code: Census field-of-degree name and FOD1P code (first-listed bachelor's field).
- ai_exposure_score_0_100: person-weighted (PWGTP) mean exposure score of the occupations held by employed civilians aged 25-64 with a bachelor's degree or higher in this field.
- pct_working_in_high_exposure_jobs: % of those workers whose occupation scores 50 or more.
- pct_working_in_low_exposure_jobs: % whose occupation scores under 20.
- median_earnings_fulltime_2024usd: weighted median of PERNP x ADJINC (2024 dollars) for full-time, year-round workers (35+ hours/week, 50+ weeks/year, earnings > 0) aged 25-64, all degree levels; rounded to nearest $100.
- median_earnings_fulltime_bachelors_only_2024usd: same, bachelor's as highest degree (SCHL=21).
- ai_exposure_score_bachelors_only: exposure score, bachelor's as highest degree.
- pct_with_graduate_degree: % of the field's workers (all employed 25-64) holding a master's, professional or doctoral degree.
- rank_felten_lm_aioe_check: the field's rank (1 = most exposed) when the same method uses Felten-Raj-Seamans LM-AIOE instead of the Eloundou score.
- three_most_common_occupations: Census SOCP occupation titles, most common first.
- acs_sample_records / weighted_workers: unweighted records and weighted workers behind the exposure score.

ai-exposure-by-occupation-2026.csv (one row per BLS OEWS detailed occupation with an exposure score; 800 rows)
- soc_2018_code, occupation_bls_oews: BLS code and title.
- ai_exposure_score_0_100: combined score; exposure_gpt4_rater and exposure_human_raters: the two components.
- median_annual_wage_may2025, employment_may2025: BLS OEWS May 2025. Blank wage = BLS publishes hourly wages only (5 performer occupations).
- exposure_match: "exact" = same SOC code in the Eloundou file; otherwise the O*NET codes averaged for an OEWS aggregate code.

ai-exposure-summary-2026.csv: every headline aggregate on the page.
- occupation_pay_fifths: occupations sorted by median wage and cut at each 20% of employment; employment-weighted mean score; pct in high-exposure (50+) jobs.
- occupation_pay_bands: employment-weighted mean score by occupation median-wage band.
- workers_by_education: ACS 2020-2024 employed civilians aged 25-64, person-weighted mean score by highest education (SCHL).
- majors_by_pay_quarter: simple average of major scores within quarters of the 151 majors by median earnings.
- correlations_spearman: rank correlations behind the page's statements.

MAPPING NOTES
- OEWS: 830 detailed occupations; 771 exact SOC matches; 29 OEWS aggregate codes take the mean of the O*NET codes sharing their first 6 characters (e.g. 31-1120 = 31-1121 + 31-1122); 30 residual "all other" codes have no score. Matched occupations = 98.9% of OEWS employment.
- ACS SOCP: exact SOC match where available; ACS aggregate codes (trailing X/Y or 0) take the OEWS-employment-weighted mean of the matching detailed occupations; ACS detailed codes that OEWS publishes only as aggregates use the Eloundou score directly; 10 ACS "all other" codes have no score (0.5% of degree holders; 0.6% of all workers).
- Exposure measures what AI could speed up, not job loss. See the page's limitations section.