Most "AI-proof majors" lists rank majors by name, which is the wrong unit of analysis. The strongest research available right now points to a specific mechanism underneath that -- and it explains why two majors that look equally "exposed" to AI on paper can have completely different employment outcomes.

The real research: it's not "AI-exposed" vs. "not," it's "automated" vs. "augmented"

A November 2025 study from Stanford and NBER economists (Brynjolfsson, Chandar, and Chen), using monthly ADP payroll data covering millions of workers, found that early-career workers (ages 22-25) in AI-exposed occupations saw a 16% relative employment decline, while employment for experienced workers in those same occupations stayed stable. Overall employment kept growing through the period -- but employment growth for 22-to-25-year-olds specifically stalled: workers that age saw a 6% decline in the most AI-exposed occupations, against a 6-9% increase for older workers in those identical jobs over the same stretch.

The more useful finding is why. The researchers used actual Claude usage data to sort AI use in each occupation into "automating" (AI substitutes for the task) versus "augmenting" (AI complements a person doing the task) -- and found entry-level employment declined specifically in occupations where AI use skews automating, while it grew in occupations where AI use skews augmenting. Same underlying exposure to AI, opposite outcomes, depending entirely on whether the work is being replaced or assisted. The paper explicitly cites nursing aides as an example of a less-exposed occupation with stable, growing employment, in direct contrast to software developers and customer service representatives.

What tends to land on the "augmenting" side of that line

This isn't a brand-new idea -- it maps onto research that predates generative AI by nearly a decade. Harvard economist David Deming's widely cited analysis found that jobs combining high cognitive skill with high social skill have shown the strongest employment and wage growth in the U.S. labor market since the 1980s, specifically because interpersonal, socially-demanding work is difficult to automate. That finding lines up directly with the automate-vs-augment split above: work built around sustained human interaction and judgment tends to land on the "augment" side, while narrowly codifiable, routine cognitive work tends to land on the "automate" side.

Separately, Anthropic's own Economic Index -- which tracks real Claude usage patterns across occupations -- finds that physical-world work sits at the bottom of AI exposure overall, and specifically notes that teachers show up as less affected than their raw task overlap with AI would predict. That's a useful data point on its own: looking only at which tasks a job description shares with what AI can do misses the in-person, relational, judgment-heavy parts of many jobs that don't show up cleanly in a task list.

Where this shows up concretely: nursing as the clearest real-world case

Nursing is the sharpest current example of the pattern actually playing out. According to AACN's own 2025-2026 national survey (998 schools, an 89.5% response rate), BSN enrollment rose 7.6% -- nearly 19,830 more students -- marking a third consecutive year of growth, bringing total BSN enrollment to 283,303 students. At the same time, nursing schools turned away a record 93,176 qualified applicants -- meaning the binding constraint right now is program capacity, not weak demand or shaky confidence in the career's future. That combination lines up exactly with the mechanism above: nursing pairs hands-on physical care, license-gated clinical judgment, and constant interpersonal and emotional work -- precisely the traits the research ties to AI augmenting the job rather than replacing it, not some blanket immunity because "a person is involved."

The pattern extends to other licensure- and physical-presence-gated fields

The same structural logic applies more broadly to skilled trades, physical therapy, and similar fields that legally require a licensed human, involve real hands-on physical work, and demand high-stakes judgment under real-world variability that doesn't reduce cleanly to a routine, codifiable task. It's worth being precise here rather than sweeping: this isn't "avoid anything technical." It's "the risk sits with roles built almost entirely around the narrow, routine, easily-codified tasks AI is best at substituting for" -- a structural distinction that can vary within a single major, not just between majors. For a close look at exactly this dynamic playing out inside one major that gets treated as a monolith when it isn't, see our piece on whether computer science is still worth it, which covers how the same degree can lead to very different outcomes depending on which part of the field a graduate actually goes into.

What this means for you

  • Judge a major less by its name and more by whether the actual day-to-day work it leads to is dominated by routine, codifiable tasks or by physical presence, licensure-gated judgment, and sustained human interaction. The research above ties those specific traits, not the major label itself, to whether AI ends up augmenting or replacing that work.
  • Don't take "safe major" lists at face value. The strongest available evidence is about a mechanism -- automate versus augment -- not a fixed list of major names, and that mechanism can show up differently even within the same broad field.
  • If you're choosing a technical major, plan to build the judgment, systems-thinking, and communication layer around the technical skill, not just the technical skill alone -- exactly the shift already reshaping entry-level hiring in computer science specifically.
  • Watch program capacity, not just job-market demand, in fields like nursing. Record numbers of qualified applicants are currently being turned away, so strong demand for the career doesn't automatically mean a seat is guaranteed -- that's a separate bottleneck worth researching for your specific target programs.

Sources