Here is the uncomfortable part, up front: the single number the entire student-loan-advice industry uses to certify a loan as "safe" is, by construction, blind to the risk students ask us about most. Run the standard affordability test across 22,642 federally tracked bachelor's programs and it hands its cleanest bill of health to engineering, computer science, and nursing — three of the fields whose day-to-day work overlaps most heavily with what large language models now do. That is not a paradox. It is proof that the test measures one thing and stays silent on another.
The rule in question is simple, and it is good advice as far as it goes: don't borrow more in total student debt than you expect to earn in your first year of work. Keep the ratio at or below 1.0 and you can usually clear the balance in ten years on the standard plan, spending roughly a tenth of your gross income on payments. The heuristic is widely attributed to student-aid analyst Mark Kantrowitz, and it has anchored "how much should I borrow" guidance for a decade. (Kantrowitz, via PBS NewsHour)
So we computed that ratio for every major we track, using U.S. Department of Education College Scorecard figures: median federal debt at completion divided by median first-year earnings. Across the 178 majors with enough programs to read reliably, the typical debt-to-earnings ratio is 0.56 — comfortably inside the safe zone — and only 2.4% of majors carry a median above 1.0. On the affordability question alone, most of higher education passes.
The safest loans sit in the most AI-exposed fields
Sort majors by that ratio and a pattern jumps out. The "safest to borrow for" end of the list is almost entirely engineering, computing, and licensed health fields — degrees where a $22,000-ish typical debt sits against a first-year wage in the mid-to-high $70,000s.
| Safest to borrow for | Debt-to-earnings | Median debt | First-year earnings | AI task-exposure |
|---|---|---|---|---|
| Systems Engineering | 0.24 | $20,500 | $79,942 | Very High |
| Computer Engineering | 0.29 | $23,000 | $79,276 | Very High |
| Electrical & Electronics Eng. | 0.30 | $23,000 | $77,732 | Very High |
| Registered Nursing | 0.30 | $22,500 | $74,861 | High |
| Computer Science | 0.31 | $22,250 | $71,764 | Very High |
Now flip the list. The most over-leveraged majors — the handful where median debt actually meets or exceeds first-year pay — are dominated by the performing and studio arts, fields whose task profiles are among the least exposed to automation.
| Most over-leveraged | Debt-to-earnings | Median debt | First-year earnings | AI task-exposure |
|---|---|---|---|---|
| Dance | 1.13 | $24,726 | $21,408 | Low |
| Drama & Theatre Arts | 1.11 | $23,727 | $20,600 | Moderate |
| Graphic Communications | 1.02 | $26,958 | $24,941 | High |
| Fine & Studio Arts | 0.90 | $23,344 | $24,486 | Moderate |
It is tempting to read those two tables as an inverse law — "AI-safe majors are debt traps, AI-risky majors are debt bargains." Resist that. The full-dataset correlation between the debt ratio and our AI task-exposure score is barely there: a Pearson coefficient of −0.19, which means AI exposure explains under 4% of the variation in borrowing ratios. The honest description is not "inverse." It is independent. Knowing a major's debt-to-earnings ratio tells you essentially nothing — good or bad — about how exposed its work is to AI. The tables above aren't a rule; they're a vivid demonstration of that independence.
Why the ratio can't see the risk
This is the whole point, and it is structural, not a data quirk. The debt-to-earnings ratio answers exactly one question: can I repay this loan at today's entry wage? It takes the first-year salary as a fixed input. But the thing students are actually worried about — will this field's earning power hold up as AI absorbs more of the work? — is a question about whether that input stays put. A ratio built on today's wage cannot, even in principle, price the durability of tomorrow's. It is the right test for the wrong horizon: a ten-year loan against a forty-year career.
It matters, too, what "AI exposure" actually measures. The academic indices we lean on — Eloundou and colleagues' "GPTs are GPTs" and the Felten-Raj-Seamans AI Occupational Exposure index — score how much of an occupation's tasks overlap with what AI can do. They are deliberately agnostic about the consequence. High exposure can mean substitution (the tool does the task instead of you) or complementarity (the tool makes you faster, and your wage rises). A computer scientist and a paralegal can both score "very high" and land on opposite sides of that line. Exposure is a flag to investigate, not a verdict.
Which is precisely why you can't fix the affordability ratio by folding AI into it. You might think: if AI is a risk to earnings, just discount the salary and recompute. We can do that — our model applies a deliberately conservative haircut, trimming first-year earnings by up to 30% of a field's exposure score as a stress test. It moves the safe-borrow engineers only modestly (computer engineering slides from 0.29 to 0.37, computer science from 0.31 to 0.39) and they stay well inside the safe zone. But notice what that exercise requires: a number nobody can credibly supply. Is the right discount 30%? 5%? Negative, because AI raises the wage? Baking a made-up coefficient into the ratio would trade an honest blind spot for false precision. The uncertainty isn't a reason to merge the two questions — it's the reason to keep them apart.
Judge debt on two questions, not one
The fix isn't a better single number. It's a second question. Before you sign, run both:
- Affordability (the old question): Is your total expected debt at or below a realistic first-year salary for the specific field and where you'll work — not the sticker average? On this axis, most majors clear the bar, and the arts are the real exceptions to watch.
- Durability (the new question): How exposed is the day-to-day work to AI, and is that exposure more likely to augment the role or replace it? A low borrowing ratio buys you nothing here; you have to look it up separately.
Both reads are on the site. Our major pages carry the AI task-exposure score alongside earnings — start with a field you're weighing, like computer science, and read the exposure line as "investigate," not "avoid." For the earnings-versus-exposure trade-off in depth, we mapped where pay and AI risk pull apart in Are High-Paying Majors Safe From AI?, and we broke down the two employment risks most rankings ignore in The Two Risks Every Bachelor's Major Carries. Every figure here traces to source data; the mechanics are in our methodology.
Frequently asked
Does this mean I shouldn't get a computer science or engineering degree?
No — and that's the trap the "AI-proof major" lists fall into. High task-exposure isn't displacement. In software and engineering, AI has so far behaved more like a productivity multiplier than a replacement, and the borrowing math on those degrees is about as safe as it gets. The point isn't to flee exposed fields; it's to go in with eyes open about a risk the debt ratio hides.
So should I pick a high-debt, low-AI major like fine arts instead?
That inverts the mistake. A degree that fails the affordability test and depends on a thin entry-level wage isn't "safe" just because its work is hard to automate. Low AI exposure doesn't repay a loan. Both questions have to clear.
How do I estimate a realistic first-year salary for my field?
Use the field-specific median, not a national college-grad average, and adjust for where you'll actually work. Our major pages report first-year earnings from College Scorecard for exactly this reason — a "safe" borrowing ceiling for one field can be twice another's.
The borrowing rule isn't broken. It's just answering last decade's question. Keep using it to size the loan — then ask the second question out loud, in writing, before you sign: not only "can I repay this at the starting wage," but "what happens to that wage." The ratio will never tell you. You have to.