The Dallas Fed Put a Number on AI's Hit to New Graduates

Graduates of Texas public universities who studied subjects most exposed to generative AI have been less likely to find jobs and have earned less than graduates of less exposed subjects since ChatGPT was released in late 2022. The Federal Reserve Bank of Dallas published the estimate on 22 September 2026: a 10-percentage-point higher share of automatable tasks in a major is associated with a 1.7-percentage-point relative fall in the probability of employment in Texas within a year of graduating, and first-year earnings about 5 per cent lower from 2021 to 2024.

The work is unusual because it does not rest on a survey or on job advertisements alone. Samuel Dodini, a senior research economist at the Dallas Fed, and Tucker Smith, a research economist there, linked administrative student records from the Texas Higher Education Coordinating Board to quarterly earnings and employment records from the Texas Workforce Commission. That gives them a graduate's major, course history, entrance exam scores and demographics alongside what they were actually paid in their first four quarters out of university.

How a major gets scored

Exposure is built in two steps. The authors start from a task-based metric developed by Anthropic that maps job tasks in the O*NET database, which covers roughly 1,000 US occupations, to records of tasks Anthropic's Claude models have performed. The result can be read as the share of an occupation's tasks that generative AI can do. They then convert that into a score for college majors using Lightcast job postings data, drawn from more than 220,000 online job boards, by averaging occupation-level exposure across every job advert that asks for a given field of study.

One methodological choice matters more than the rest. The mapping from majors to occupations uses only postings from the first quarter of 2018 to the third quarter of 2022, before ChatGPT's release, so that later shifts in demand, or in what employers say they want, cannot quietly redefine which majors count as exposed. The ordering that falls out is intuitive: computer science, computer engineering and languages rank among the most exposed, while nursing, education and psychology rank among the least.

More study, same field, same problem

The graduates responded the way economics predicts. When first jobs became harder to get, the cost of staying in education fell, and May 2024 graduates from majors with 10 percentage points more exposure were 1.4 percentage points more likely to enrol in graduate study within a year. What they did not do was change direction. Two-thirds of 2024 computer science graduates who went back to university that autumn enrolled in a computer science graduate programme.

That appears to have been a poor hedge. New entrants to the labour market holding master's degrees in more-exposed fields recorded relative earnings declines similar to those of four-year graduates, which led the authors to conclude that the returns to formal upskilling inside an AI-exposed field may be limited. Deeper training in the thing being automated is not the same as training in something complementary to it.

Eighteen-year-olds are reading the signal

The clearest behavioural finding concerns people who have not graduated yet. Comparing Texas undergraduate enrolment counts by major, the authors find that for each 10-percentage-point difference in automatable task share, enrolment at Texas four-year institutions fell 4.8 per cent between autumn 2024 and autumn 2025. Parallax Nexus has reported on Stanford's revised 'Canaries in the Coal Mine' work, which used ADP payroll data through June 2026 to put employment of 22-to-25-year-olds in AI-exposed occupations 19 per cent below where it would otherwise have been, with the adjustment running through reduced hiring rather than layoffs. This is the next link in that chain: the hiring gap has begun to reshape what students choose to study.

The judgment worth stating plainly is that this is the most credible causal evidence yet that AI has moved graduate labour outcomes, and it is still narrower than the headline invites. Every figure is relative, comparing exposed majors with less exposed ones rather than measuring an absolute decline, so a uniform shock across all majors would not show up at all.

The limits run further. Employment is recorded only for jobs held in Texas, so a computer science graduate who took a job in Seattle counts the same as one who found nothing. The post-2022 window also contains a broad technology hiring retrenchment and a sharp rise in interest rates, and the design cannot separate the part of the software-hiring slump caused by models from the part caused by cheap money ending. The authors name the indicators that would settle it, pointing to layoffs and wage growth among prime-age workers aged 25 to 54 in occupations at greater automation risk as the series to watch next.