Was this forwarded to you by a friend? Sign up, and get your own copy of the news that matters sent to your inbox every week. Sign up for the On EdTech newsletter. Interested in additional analysis? Upgrade to the On EdTech+ newsletter.
AI is driving job losses for recent college graduates, or the “AI Job Apocalypse is a complete fantasy.” As with most complex questions, both emphases can be true even if overstated. And the question has enormous significance in higher ed, as institutions attempt to determine whether we have real and sustained changes in the labor force demand for certain majors, or is the issue more of needing to redefine program outputs?
I have mostly chosen the Brave Sir Robin approach, saying that it’s too early to tell while running away. More seriously, the data have not been compelling enough to know. There was already a tech-driven job correction that started before the November, 2022 release of ChatGPT 3.5 (the generally-accepted AI is here marker), and too much of the AI job apocalypse explanations ignored Covid, inflation, decreases in the confidence in college degrees, and market pullbacks.
We now know a lot more from recent research and analysis, however. Taken together, the emerging research makes it increasingly likely, in my judgment, that AI is adding to the deterioration in early-career opportunities. There was a working paper released this month by The Census Bureau that provides some of the strongest evidence yet by connecting college majors to actual employment and earnings outcomes. How much of the damage comes from AI remains unresolved, but there is a stronger basis for judgment than there was before.
The Report
Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors is the new report, and it links college graduation records (for bachelor’s degrees), including each graduate’s major, to actual employment and earnings records, following graduates wherever they find work—not just within their chosen field. This treatment includes the unintentional barista problem of our economy, and too many studies looking primarily at occupations miss this problem. By comparing successive graduating classes across majors with different levels of AI exposure, the researchers explore results such as delayed employment, lower earnings, and shifts into lower-paying industries. This report gives us perhaps the best understanding of the transition from college to work and strengthens the evidence that AI may be contributing to early-career job deterioration in many fields.
The study listed 142 individual majors in Table A3, each with at least 1,000 graduates during 2018–2022. It also included double majors and smaller fields in the statistical analysis. These majors were grouped into deciles based on how AI-exposed they were—the researchers estimate each major’s exposure using the jobs its graduates historically entered and the tasks AI could perform in those jobs. They group majors from least exposed (decile 1) to most exposed (decile 10), keeping those assignments fixed throughout the analysis. AI’s capabilities continue to evolve.
The report’s discussion summarizes the findings.
Overall, our results suggest that there was a fairly immediate and concentrated deterioration in the early labor market outcomes of graduates from the most AI-exposed fields following the introduction of ChatGPT-3.5. Our results point to a disruption that operates most strongly through labor market entry and the quality of the initial jobs obtained by graduates from the most AI-exposed fields. Following the introduction of ChatGPT, graduates from these fields became less likely to find employment immediately after graduation and experienced substantial earnings declines. Among those who became employed, an important part of the adjustment occurred through changes in the sectors they worked. Employment shifted away from relatively high-paying sectors such as Information and Professional, Scientific, and Technical Services and toward lower-paying sectors such as Retail Trade and Accommodation and Food Services. The differential effects become smaller as workers move further from graduation, and job-switching rates increase among the cohorts that entered the labor market after ChatGPT. These patterns are consistent with some exposed graduates initially accepting weaker matches and subsequently moving toward better jobs, although employment and earnings for the most exposed graduates do not fully recover within our measurement window.
Unfortunately, the unpacking job of determining how much of certain college major job prospect changes is caused by AI versus other explanations is not possible from this study. Furthermore, the report suffers from the typical problems of dense, inaccessible prose and poor chart labeling. Accordingly, I have recreated and relabeled the charts in Datawrapper and made them interactive. Explore the charts online to identify individual majors and see their employment and earnings changes.
Changes Over Time - Employment (Figure 3)
Figure 3 from the report looks at the “likelihood of having observed employment for any duration during the quarter immediately following graduation,” showing AI-exposure deciles 7 - 10. These charts show whether graduates in more exposed majors gained or lost ground relative to graduates in less exposed majors, using 2022 as the benchmark. The flat line during 2022 is built into the comparison; it does not mean nothing changed that year. I am showing deciles 7 and 10 only using interval bands and representative majors. The employment decline is largest in the most exposed group; smaller declines also appear in other exposed groups.
Changes Over Time - Earnings (Figure 4)
And you get similar differences when looking at actual earnings from figure 4. Note that the most exposed group’s earnings had already weakened before ChatGPT, with further substantial deterioration afterward
Prospects by Major - Employment (Figure 10)
Being a mostly higher ed newsletter, the natural question is how the data illuminates labor force prospects pre- and post-AI introduction by major. In other words, looking at the correlation between AI-exposure and labor market changes. And here the results are compelling in terms of there being some connection. The highest AI-exposed fields have the worst pre- and post- changes. Below zero means a major’s outcomes deteriorated more—or improved less—than the overall trend.
Looking at the same two metrics, figure 10 shows changes based on employment share post-graduation, with decile 10 majors (computer science and programming in particular) down and to the right. Higher AI exposure is associated with weaker employment outcomes.
Prospects by Major - Earnings (Figure 11)
And figure 11 shows the metric of earnings. Higher AI exposure is associated with weaker earnings outcomes.
What We Can Now Say
As noted, there are limitations with this study. The study does not include certificates or associate’s degrees or graduate degrees, and it draws predominantly from public research universities with a limited timeframe. Further, the AI exposure is inferred, not directly measured—the researchers do not observe each graduate’s occupation or whether their employer adopted AI.
But the biggest limitation is that the study cannot separate the impact of the tech industry pullback from the AI disruption. The fields most exposed to AI—particularly computer science and related majors—were also heavily exposed to the reversal of the pandemic-era tech hiring boom. Employers expanded rapidly on the assumption that the surge in digital activity represented a permanent change. As those expectations faded and inflation brought tighter financing, investors and employers shifted their emphasis from growth toward profitability, prompting layoffs and reduced hiring. That correction was underway before ChatGPT launched in November 2022. The central question is therefore how much graduates’ prospects would have deteriorated even without generative AI.
The study investigates alternative explanations, but it cannot fully separate AI’s contribution from this overlapping correction. Its controls for remote work and its discussion of interest-rate effects do not capture every consequence of overhiring, disappointed growth expectations, and changing employer priorities.
This study also fits with other evidence, although the pieces do not tell a simple AI-caused-everything story. The charts in Burning Glass Institute’s No Country for Young Grads show weakening demand for junior computer-related roles before ChatGPT, reinforcing the importance of the tech hiring correction. Meanwhile, Stanford’s research using ADP payroll records finds that young workers in AI-exposed occupations have increasingly fallen behind, primarily through reduced hiring, with the pattern persisting even outside technology firms and computer occupations. Neither establishes how much AI caused, but together with this new study, they strengthen my interpretation that AI is adding to an already difficult transition from college to work.
In my view, the accumulating evidence makes it increasingly likely that AI added to the deterioration already in place in 2022, especially for recent graduates, while the size of separate effects remains unresolved. The study substantially improves our understanding of what happened to graduates, even with these limitations.
The main On EdTech newsletter is free to share in part or in whole. All we ask is attribution.
Thanks for being a subscriber.
