The hits keep coming for American higher ed.
Total college enrollment, which had climbed continuously for decades prior, surged during the Great Recession but stagnated thereafter as Americans soured on the importance of a college degree. Birth rates peaked around that same time and then crashed, meaning that fewer and fewer kids will reach college age starting around … well, now, actually. Next came AI, which threatens colleges both on campus, thanks to rampant cheating, and through the post-graduation job market, as the new technology can do a lot of the things that colleges spend so much time teaching people to do.
A new working paper digs into this last possibility. It finds that the “college wage premium”—the extent to which those with bachelor’s degrees out-earn high-school grads—has suddenly dropped off, just in the past few years, after rising for decades and plateauing in the mid-2010s. (If that’s not pessimistic enough for you, an earlier study with slightly different methods put the plateau around the turn of the century.) It further shows that the trend is worst in jobs whose tasks can most readily be done by AI.
Before college grads and anxious parents panic, they should note that, in the authors’ estimation, the wage advantage fell only from about 86 percent to 77 percent between 2022 and 2026.1 But that’s a pretty significant drop in a few short years. And as I’ve discussed in this space before, not all educational earning disparities are actually caused by education to begin with—to some extent, the gap simply happens because those who self-select into additional schooling are smarter, more motivated, and so on.
The authors’ calculations are pretty straightforward, as far as these things go. They rely on the Current Population Survey, which samples American households every month to check on their members’ employment situations, and limit their analysis to workers age 16 to 64. They present the raw hourly wage premium (with salaries converted to an hourly rate) alongside a more complicated calculation that adjusts for demographic changes, though both “display identical turning points after 2022.”
Next, they look at whether AI is a plausible explanation, given that 2022 happens to mark the debut of ChatGPT. They turn to a “task-based exposure index” developed in a previous study, which “quantifies the share of task requirements within an occupation that can be completed or accelerated by large language models” and tends to be highest in occupations requiring college degrees. In the years leading up to 2022, this measure bore little relationship with wage trends over time—but in the years since, jobs with higher exposure have fared progressively worse.
Statistically, the authors are able to explain only about 28 percent of the college-premium dropoff with this pattern, though it’s important to bear in mind that the metric is an estimate of automatable tasks, not a concrete measure of how much of each job is actually being done with AI. The authors also caution that their analysis isn’t ironclad proof of causation, but the fact that AI exposure suddenly became a useful variable when ChatGPT debuted is, in my humble opinion, quite striking.
Further, in analyses that will be of most interest to economics nerds, the authors bring some supply-and-demand modeling into the picture. Logically speaking, if the college wage premium declined, and this can’t be explained by a sudden surge in the supply of college grads, the demand for college labor must have fallen. That’s what their fancy models show as well.
What’s the takeaway? Colleges—especially lower-tier ones that can’t simply relax admissions standards when applications drop off—should be in a full-blown panic right now, with these declining rewards to a college degree piling on top of a birth dearth.
Students, potential students, and recent grads can be a little bit calmer. College still pays off for a lot of people, and we have no idea what these trends will look like a decade from now. But by the same token, these folks do need to think about what an education offers them, and what they might learn to offer employers, in a ChatGPT world.
From the Manhattan Institute
Jennifer Weber, Ray Domanico, and Carolyn D. Gorman on the mediocre state of New York State’s schools.
Other Work of Note
“The average U.S. worker, regardless of whether they use the technology at work, believes there is a 20% chance of losing their job to generative AI within two years, which is comparable to workers’ realized risk of job loss from any cause over a similar horizon.” And some more survey data about how workers are using AI.
Will AI reduce linguistic diversity? And how will it affect tax revenue?
Why has unemployment generally trended downward for decades, notwithstanding cyclical fluctuations?
High earners’ income has been increasingly at risk, and those with riskier incomes unsurprisingly save more.
Eric Kaufmann and two coauthors evaluate the “equity-excellence tradeoff” using data from a Canadian academic program.
A congestion experiment that used Google Maps to reroute traffic. And occupational licensing for Uber drivers doesn’t seem to do much good.
It’s bad if your house burns down.
The 2017 tax law’s changes to corporate taxes caused “increases in firms’ investment, sales, profits, employment, and payrolls, with earnings gains concentrated among highly paid workers. In the short run, 87 percent of private income gains flow to the top 10 percent of the income distribution.”
CBO on the taxation of Social Security benefits. And on automatic stabilizers.
The connection between gender ideology and left-wing authoritarianism.
Why do C-section rates vary from place to place?
Big new survey on which guns Americans own and what they do with them. Plus recent non-fatal firearm violence stats for 17 cities. And how gun laws affect suicide.
Prisons shrank and racial incarceration gaps closed from 2000 to 2019.
“Prosecutor-defendant race pairings show little evidence of conviction effects overall.”
A new tracker of fatal police shootings.
When do health-services professionals endorse withholding information from the public?
Some ideas for regulating gain-of-function research.
Public-health laws reduced mortality during the 20th century.
The final, published versions of two studies of universal-basic-income trials (which already generated considerable discussion as working papers) are now available. As is one on how high-performing poorer students suffer under test-optional college admissions policies.
How often do studies use causal language even when their methods don’t justify it?
Gender gaps in school readiness.
These numbers are exponentiated from the logarithmic differences reported in the paper, so that normal humans might understand them.
It has dropped off, but I’m telling you from experience that if you’ve been getting work without a degree in a field that generally requires a degree, if you ever have to apply for another job suddenly, you’re going to be in the bottom half of the list for interviews (or excluded altogether)…
That’s why I’m in college in my mid/late thirties. 😬
what i liked is they don't oversell it. the AI angle is only about a quarter of why the wage gap shrank, and the rest is just less demand for grads that they can't fully explain yet. so a degree's worth a bit less than before, but nobody's sure exactly why. there's a good piece in noema on the bigger picture, basically that the old promise of out-earning your parents is quietly disappearing, and this wage drop looks like one small part of it.
https://www.noemamag.com/the-end-of-upward-mobility