What Should Your Kid Study in the Age of AI?

In its February 2026 figures, the New York Fed put unemployment for recent US computer engineering graduates, aged 22 to 27, at 7.8%. Fine arts graduates: 7.7%. Nursing: 2.1%.

The degree everyone called safe and the one everyone was warned off, a tenth of a point apart.

That number proves less than it seems. Samples for a single major are small: an analysis of the previous release found the computer engineering figure could have been anywhere from about 4% to 11%. And unemployment isn’t pay. But it’s a useful warning about anyone who tells you which degree is safe.

So here’s what the longer evidence says.

The fast-paying subjects are the fast-changing ones

David Deming and Kadeem Noray (Quarterly Journal of Economics, 2020) followed US graduates across their careers, and tracked how fast the skills in millions of job ads changed between 2007 and 2019.

In 2019, 47% of computer and maths job ads asked for at least one skill that was new since 2007. In education, law and social work, under 20% did.

The pay follows the same pattern. Computer science and engineering graduates earn about 45% more than other graduates early in their careers, and about 33% more by age 50. For business, the lead falls from about 38% to 20%.

The authors call it a race: every year you learn, but in a fast-changing field some of what you learned stops being useful. People leave, too. About six in ten computer science and engineering graduates work in computing or engineering at 26, and about four in ten at 50. Almost all who left went into management.

In the slow subjects, talent takes ten years to show up on the payslip.

One finding stands out. A computer science graduate with well-above-average ability scores earns about 24% more than an average classmate at 23, and 43% more at 34. An equally able graduate in another subject earns only 7% more at 23 — and 43% more at 34. Same place, reached later.

What this doesn’t say: that technical degrees are a bad bet. A third more pay at 50 is still a big lead.

The same shape, thirty years earlier

Eric Hanushek and colleagues (Journal of Human Resources, 2017) compared men with vocational and general education in 11 European countries, using surveys from 1994–98.

At 16, the vocational group was about 7 percentage points more likely to be in work. The gap closed by about 2 points a decade, and flipped at around 49 (54 in another version; the authors call the exact age imprecise). Pay flipped earlier. The pattern was strongest in Germany, Switzerland and Denmark, the apprenticeship systems the rest of the world copies.

Their best guess is that specific skills go out of date faster. But the general-education men also did more training later in life (37% against 30%), and the authors say they can’t separate the two explanations.

A different machine

Both studies were done before generative AI. The 1990s data describes automation that came for the hands: factory jobs, specific machines.

The newest evidence describes something else. Stanford’s “Canaries” study (Brynjolfsson, Chandar and Chen, updated August 2026) uses US payroll data. For workers aged 22–25 in the jobs most exposed to AI, employment fell about 11% between late 2022 and mid-2026. In the least exposed jobs, it rose about 10%. That’s a 19% gap. Experienced workers show no comparable gap.

The most exposed jobs include software developers and customer service. The least exposed include nursing aides and home health aides. The authors’ explanation: AI replaces codified knowledge, the “book-learning”, but not yet tacit knowledge, the know-how that builds up with experience. Where people use AI to do tasks for them, young workers’ jobs fell. Where they use it as help, jobs held steady or grew.

This cuts both ways for parents. A general degree is mostly codified knowledge. But so is the start of a software career, and the most job-ready degree of all is losing its entry jobs first. AI isn’t sorting by major. It’s sorting by whether you can write the job down.

The authors call these early, descriptive signs, not proof that AI caused the gap. The gap shrinks once education is accounted for, and it’s stronger in their payroll data than in government surveys.

And in New Zealand?

The squeeze is here too. Graduate roles have been just over 2% of Seek’s job ads, and unemployment hit 5.6% in the June 2026 quarter. But Seek’s own economist links the fall in graduate ads to a slower economy and a looser job market, not AI. Nobody has separated the two yet.

So what do you tell a sixteen-year-old?

Two answers, and the episode doesn’t pick between them.

Breadth. Two studies, decades apart, show the job-ready path winning early and fading. In both, the people who did best later kept learning. So pick the subject that makes the next one cheaper to learn, and expect to retrain.

Real work, early. If tacit knowledge is what protects you, and the beginner jobs that build it are drying up, don’t wait for an employer to provide the first rung. Apprenticeships, work placements, a demanding part-time job, a project other people actually use. This one is an inference from the Stanford explanation, not a tested finding.

---

The major matters less than most parents fear, in both directions. A technical degree still pays more at 50, so nobody should be frightened off one. But every lead shrinks, and right now the first jobs are wobbling. The better question isn’t which degree. It’s where a young person gets to do real work, with real stakes, before anyone will hire them to.

---

Jurisdiction: the New York Fed, Deming & Noray and Stanford data are American. Hanushek and colleagues studied men in 11 European countries, surveyed in 1994–98. The New Zealand figures are from Seek and Stats NZ; nobody has yet measured AI’s effect on New Zealand graduates.

Sourcing note: the New York Fed’s figures for single majors rest on small samples, so treat them as a warning, not a ranking. The Stanford study is descriptive: its authors don’t claim AI caused the gap.

---

Sources: Federal Reserve Bank of New York, “The Labor Market for Recent College Graduates,” February 2026 · Deming & Noray, “Earnings Dynamics, Changing Job Skills, and STEM Careers,” Quarterly Journal of Economics, 2020 · Hanushek, Schwerdt, Woessmann & Zhang, “General Education, Vocational Education, and Labor-Market Outcomes over the Lifecycle,” Journal of Human Resources, 2017 · Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab, updated August 2026 · Stats NZ, Labour market statistics, June 2026 quarter.