Dallas Fed: AI-Exposed Majors Pay a 1.7-Point Hiring Penalty in Texas

Computer science enrollment at Texas public universities fell 27.8% in a single year, from 1,628 students in fall 2024 to 1,176 in fall 2025. Nursing, the least AI-exposed major in the same dataset, grew 10.6%. Both numbers sit in a spreadsheet the Federal Reserve Bank of Dallas published on September 22, 2026, alongside a study arguing that generative AI has already cut the hiring odds and pay of graduates from automatable majors. The spreadsheet says something the article's prose does not: the pivot away from exposed fields is real, uneven, and much smaller than the CS headline suggests.

What the study measured

Samuel Dodini and Tucker Smith, economists at the Dallas Fed, linked three administrative sources: Texas Higher Education Coordinating Board records (major, course history, entrance scores), Texas Workforce Commission quarterly earnings, and Lightcast job postings drawn from more than 220,000 online boards. [Fact] Each college major gets an exposure score equal to the average AI-automatable task share of the occupations that asked for that major in job ads between Q1 2018 and Q3 2022 — deliberately pre-ChatGPT, so the mapping is not contaminated by the thing being measured.

The task share itself comes from Anthropic's index that maps O*NET tasks to what its Claude models are actually used for. Disclosure: the occupation pages on this site draw on exposure figures from the same family, and this draft was written with an Anthropic model. Nothing in the results changes because of that, but you should know the ruler and the reviewer share a manufacturer.

The comparison is an event study. May 2021 graduates — the last cohort to finish a full first year in the labor market before ChatGPT — are the zero line. Every other cohort's gap between more- and less-exposed majors is measured against them.

Four headline numbers, with the scaling the headline drops

Every effect is stated per 10 percentage points of exposure. That matters, because the gap between the most and least exposed majors in the file is about 30 points.

[Fact] Employment within a year of graduating, in Texas: −1.7 points per 10 points of exposure for 2024 graduates. The chart data give −1.77 for the 2023 cohort and −1.71 for 2024, with a 2024 confidence band running from −1.08 to −2.34.

[Fact] First-year earnings among those who did find Texas jobs: −5.4% in 2024 (band −4.1 to −6.7), after −3.2% in 2023. The 2022 cohort shows −0.2%, effectively nothing.

[Fact] Graduate-school re-enrollment within a year: +1.4 points per 10 points of exposure for May 2024 graduates (band +0.7 to +2.1). Two-thirds of returning 2024 computer science graduates enrolled in a computer science graduate program.

[Fact] Undergraduate enrollment by major, fall 2024 to fall 2025: −4.8% per 10 points of exposure. The workbook lists the regression directly: slope −0.48, intercept +7.98.

Now do the arithmetic the article leaves to the reader. Registered nursing scores 1.0% exposure in the file; computer science (CIP 11.07) scores 31.7%. That is a 30.7-point gap, so multiply by roughly three. [Estimate] A 2024 CS graduate's odds of first-year employment in Texas fell about 5 points relative to a nursing graduate, first-year pay about 16%, and the CS graduate was about 4 points more likely to be back in school. The +7.98 intercept means a major with zero exposure would have grown about 8% last year; at CS-level exposure the model predicts −7%.

What the chart data add

I downloaded the workbook behind Chart 4 — 170 four-digit CIP majors, 152 of them with exposure scores — and summed it myself. None of the following figures appear in the article.

Total enrollment across all 170 majors rose 4.1% (126,419 to 131,575). Majors with exposure of 20% or more fell 3.9% (8,031 to 7,717). Majors under 5% rose 6.0%. A ten-point swing between the top and bottom bands: that part matches the story.

But the high-exposure band is only 6.4% of Texas enrollment. The "pivot away from AI-exposed fields" is 314 students in a system of 131,000. The enrollment-weighted average exposure of a Texas undergraduate is 7.9%. The typical student was never in the blast zone.

Inside the exposed band the movement is contradictory. Computer science (11.07) is down 27.8%. But "Computer and Information Sciences, General" (11.01), a larger code with nearly the same exposure at 28.7%, is down only 3.7%. Computer engineering (14.09) — the second most exposed major in the entire file at 33.7%, and one the authors name among the most exposed — grew 26.5%, from 906 to 1,146. Linguistics, the single most exposed code at 43.7%, lost exactly one student. Combine both computer science codes and the decline is 11.0%, not 27.8%.

Students are not fleeing exposure. They are fleeing a label.

Two more that cut the other way: general engineering (15.8% exposure) fell 14.4%, and liberal arts (7.1%) fell 5.4%, both worse than their exposure predicts. Enrollment moves for reasons the score does not capture — admission caps, program renames, transfers between codes. [Claim] Reading the −4.8% slope as "students respond rationally to AI risk" is one interpretation. "One flagship code got a bad reputation in 2025" fits the same data about as well.

The master's-degree finding is weaker than it reads

The article's most quotable line may be that master's recipients in exposed fields "experienced similar relative declines in earnings as four-year graduates," implying that formal upskilling does not help. The 2024 point estimate is indeed −5.4%. But the same series shows −4.1% in 2013, −5.3% in 2014 and −3.2% in 2015 — all with confidence bands that exclude zero, all in years before generative AI existed. The bachelor's earnings series has no such pre-period wobble, and its bands are roughly half as wide. [Claim] A series that produced a −5% "effect" three times before the treatment cannot cleanly attribute a fourth one to the treatment. The defensible summary is "no evidence a master's helps," not "evidence it doesn't."

The bachelor's employment series has a smaller blemish of its own: 2016 reads −0.83 with a band from −0.09 to −1.57. One pre-period year out of six brushes the line. That is acceptable for an event study, but the pre-trend is "mostly flat," not "flat."

The limitation the authors state, and the one they do not

Stated: whether the AI-literacy courses Texas universities now offer prepare students for anything remains, in the authors' words, "an open question."

Less foregrounded: employment means employment in Texas. The outcome is built from Texas Workforce Commission records, so a computer science graduate hired in Seattle or San Jose counts the same as one who found nothing. CS graduates are among the most geographically mobile groups in the sample; nursing graduates are among the least. [Claim] Some fraction of the −1.7-point gap is out-migration rather than joblessness, and the study cannot say how much. The same applies to graduate school: only re-enrollment at Texas four-year universities is visible. The 5% earnings effect is cleaner, since it is conditional on having a Texas job.

How this fits what we already had

Our September 15 write-up of the Dallas Fed's job-postings study showed Texas postings for AI-automatable roles down 7-8%, driven by incumbent firms. This new paper looks at the other side of that market: the people who would have filled those postings. Together they close a loop that our June piece on the broken bottom rung and our March cross-analysis of entry-level hiring could only assert: fewer postings, then fewer hires, then lower pay, then a retreat into graduate school that may not pay off. The authors add one forward indicator worth watching — layoffs and wage growth for prime-age workers (25–54) in exposed occupations, which would show the effect spreading beyond new entrants.

What this means for your job

If you are a software developer, computer programmer, web developer or data scientist early in your career, the 2024 cohort in this study is your peer group. Three practical reads:

• A same-field master's is not a proven hedge. Returning CS graduates overwhelmingly chose CS again and saw no earnings protection the data can detect. If you go back to school, go back for something the model cannot yet do — the authors' word is "complementary."
• Computer engineering enrollment grew 26% while computer science fell 28%. Incoming students are betting that hardware-adjacent work is safer. The exposure score disagrees (33.7% versus 31.7%). Someone is wrong; watch which.
• If you are a translator or interpreter, languages sit in the most-exposed group here — the same signal we have seen in Korean and ILO data.

For registered nurses, elementary school teachers and clinical psychologists: your majors are the control group in this study. Enrollment is flowing toward you. That is good for your profession's pipeline and, a few years out, worse for its wage bargaining.

Source: Samuel Dodini and Tucker Smith, "AI plays a role in weak labor market for college graduates," Dallas Fed Economics, Federal Reserve Bank of Dallas, September 22, 2026. The chart-data workbook (0922data.xlsx) was downloaded and summed by this site; band totals, the 6.4% share, the 7.9% weighted mean, the combined-CS figure and the CS-versus-nursing scaling are our calculations, not the Fed's. The views in the source are the authors' and not the Bank's.

AI-assisted analysis: drafted with an AI model from the primary source and its published chart data, then checked against both.