Greetings, Earthlings. For two years, your debate about AI and jobs has run mostly on predictions and on surveys of what bosses think might happen. This autumn, economists at the Federal Reserve Bank of Dallas brought something rarer: official records of what actually happened to real graduates.
They found that after ChatGPT arrived, students who studied subjects most exposed to AI found it harder to get work, and earned less when they did. It is one of the clearest pieces of evidence so far that AI is affecting entry-level jobs. It is also narrower than the headlines suggest.
The short version
Dallas Fed economists tracked Texas public university graduates using official education and earnings records.
After ChatGPT's release in late 2022, graduates from more AI-exposed subjects became less likely to find a job within a year: 1.7 percentage points less likely for every 10-point increase in exposure.
Those who did find jobs saw first-year pay fall by about 5% between 2021 and 2024, compared with graduates from less-exposed subjects.
More of them went on to postgraduate study, but master's graduates in those fields saw similar falls in pay.
The study shows a strong association, not proof that AI alone caused it, and it covers Texas only.
What the researchers did
Samuel Dodini and Tucker Smith used administrative data from Texas: what students studied, which courses they took, and whether and where they were employed and paid afterwards. These are official records, not self-reported surveys, which is what makes the study stand out.
To measure "AI exposure", they scored each occupation by the share of its tasks that generative AI can do, using a task-based measure built from Anthropic's data on how people use its Claude assistant. Each degree subject was then scored by the jobs its graduates typically go into, based on job adverts from before ChatGPT launched.
The most exposed subjects were computer science, computer engineering and languages. The least exposed were nursing, education and psychology.
The researchers then compared graduates from more- and less-exposed subjects, before and after ChatGPT, using the May 2021 graduating class as the baseline. The important detail is that the trends were stable before ChatGPT arrived. The gap opened afterwards.
What they found
Fewer jobs. For every 10-point rise in a subject's AI exposure, its graduates were 1.7 percentage points less likely to find work in Texas within a year.
Lower pay. Among those who did find work, first-year earnings fell about 5% from 2021 to 2024, relative to less-exposed subjects.
A retreat into education. May 2024 graduates were 1.4 percentage points more likely to start postgraduate study for every 10 points of exposure. Two-thirds of the 2024 computer science graduates who went back to study chose more computer science.
Upskilling did not rescue pay. New master's graduates in exposed fields saw pay falls similar to bachelor's graduates.
Students are already adjusting. Undergraduate enrolment fell 4.8% for every 10-point difference in exposure between autumn 2024 and autumn 2025.
The authors conclude that the shift to AI has hurt many recent graduates. They suggest the next things to watch are layoffs and pay growth for experienced workers in the most exposed occupations.
What it does not show
This is good evidence, but it has limits, and the authors are careful about them.
It is an association, not proof. The paper's own title says AI "plays a role" in the weak graduate market, not that it is the only cause. Other things changed for tech graduates after 2022: the end of pandemic over-hiring, higher interest rates and big tech layoffs. In our view, computer science is both the most AI-exposed subject and the one most hit by the tech hiring cycle, and the study's design reduces this problem but cannot remove it entirely.
Texas only. Graduates who took jobs outside Texas are not counted as employed.
The exposure measure is fixed. It is based on pre-ChatGPT job patterns and one company's usage data, so it may miss how demand has shifted since.
The results are relative. They show exposed-subject graduates did worse than others. They do not say how many jobs disappeared in total.
It is early. The post-ChatGPT period is short, and the enrolment change is a single year.
What about the UK?
The data is different, but the mood is similar. Job site Adzuna counted 8,383 graduate vacancies in July 2026, down 45% on a year earlier and the lowest since it began tracking in 2016, according to analysis published by Employment Hero.
UK business groups, however, mostly point to higher employment costs and economic uncertainty, and no one has yet produced UK evidence as clean as the Dallas Fed's that isolates AI's share. The honest summary: the UK graduate market is weak, AI is a plausible contributor, and how big a contributor is still an open question.
Signal vs hype: our reading
Signal: the first hard, records-based evidence that AI exposure is linked to worse early-career outcomes, with a clear before-and-after pattern.
Hype: "AI is wiping out graduate jobs". The effects are measurable but modest, counted in percentage points rather than collapse, and tangled up with other economic forces.
What this means if you are studying or hiring
For students: exposure is not destiny. The least exposed subjects involve people, judgement and physical presence. In exposed fields, the routine tasks that used to train junior staff are exactly the ones AI now does, so evidence of judgement, real projects and domain knowledge counts for more than ever.
For employers: if AI is absorbing junior tasks, ask where your future senior people will learn the job. Cutting entry-level hiring saves money now and can create a skills gap later.
Frequently asked questions
Is AI taking graduate jobs?
Evidence from Texas suggests AI is contributing to weaker job prospects and lower pay for graduates in the most exposed subjects. It is one factor among several, and the effect is measured in percentage points, not mass job losses.
Which degrees are most affected by AI?
In the Dallas Fed study, computer science, computer engineering and languages were the most exposed. Nursing, education and psychology were the least exposed.
Does a master's degree protect you from AI?
In this data, not much. Master's graduates in exposed fields saw earnings falls similar to those with bachelor's degrees.
End of transmission. We will keep watching the data, not the predictions.