What should I tell my children to study?

Nobody can tell you which subjects are safe, and the organisation best placed to try has published its own score. The United States Bureau of Labor Statistics went back and marked its ten-year occupational forecasts against what actually happened. It got the direction of change right 78 per cent of the time. It got the question people actually care about, which occupations would grow faster than the economy, right 57 per cent of the time. Its largest errors came from a shock it could not foresee.

That is the honest ceiling on this advice. Anyone telling your child which field is future-proof is making a claim the only scored forecaster in the world cannot support about a decade that did not contain a general-purpose technology.

The one place forecasting has been marked

Almost nobody scores their predictions. BLS does, publicly, and the 2006 to 2016 evaluation is the most useful document in this whole argument.

Across 840 detailed occupations, direction of change was right 78 per cent of the time, which is respectable. Relative growth, whether an occupation would outpace the economy, came in at 57 per cent. And the magnitude was badly out: projected average occupational growth of 10.4 per cent against an actual 3.6 per cent, because the projections assume full employment and the decade contained a financial crisis.

Hold those two facts together. The forecast failed hardest exactly where a discontinuity arrived, and the current question is what a discontinuity does to work. There is no scored record at all for a technological shock of this kind, which means the confident answers now circulating have less behind them than the 57 per cent.

Subject choice matters enormously, in the direction nobody discusses

The financial spread by subject is very large, and it dwarfs the spread between going and not going.

The Institute for Fiscal Studies linked school, university and tax records for an entire English GCSE cohort and projected earnings to 67. The average net lifetime return to a degree is around 100,000 pounds. Medicine and economics exceed 400,000. Creative arts, philosophy and languages show low or negative average returns, and roughly 20 per cent of women and 30 per cent of men are projected to see a negative return overall.

So the decision does carry money. What it does not carry is the safety everyone is actually asking about, and the IFS is explicit: it declines to model structural change including AI, because the historical data cannot see it.

The advice most parents give is aimed at the wrong level

The standard instruction is a category: do STEM, avoid the arts. Georgetown's analysis of 152 majors shows why that is the wrong unit.

Median prime-age earnings run from about 58,000 dollars in education and public service to 98,000 in STEM. But within STEM alone the range is 64,000 to 146,000, and several humanities majors beat the STEM 25th percentile. The variation inside a field is wider than the gap between fields.

Which means advice given at the level of the category is close to noise. It sorts on the smaller difference and ignores the larger one.

What is actually happening to graduates right now

Two current numbers, and both need care.

The New York Fed's series shows recent-graduate unemployment around 5.6 per cent as at the second quarter of 2026, with underemployment around 42 per cent, the highest since 2020. Underemployment is the bigger and less discussed number: working, but in a job that did not require the degree.

Separately, the Stanford Digital Economy Lab, using payroll records covering more than four and a half million workers, finds employment of 22 to 25 year olds in the most AI-exposed occupations running about 19 per cent below where it would be had it tracked less-exposed peers. The gap concentrates where AI substitutes rather than complements, and in codified rather than tacit knowledge. Experienced workers show no equivalent gap.

The authors say plainly that these are descriptive patterns rather than causal estimates, and that they cannot yet say AI is the cause. Take the signal and refuse the headline.

The complication in the reassuring answer

The comfortable response to all this is that transferable skills will carry you, so study anything and stay adaptable. The evidence complicates that more than it supports it.

Gathmann and Schoenberg, using German administrative data, found that task-specific human capital accounts for up to 52 per cent of wage growth. People move between occupations with similar task profiles, and the distance of those moves shrinks with experience. Skill is portable, but it travels along task lines rather than being general.

So the position is narrower than either camp wants. Specific capability does the work. It just is not tied to the job title people think it is tied to.

What to actually tell them

Five things that survive the evidence, and none of them is a subject.

• Refuse the safe-field frame, and say why. The best forecaster on record hits 57 per cent on relative growth. A parent asserting more than that is guessing with more confidence than the data allows.
• Ask what tasks the subject builds, not what job it leads to. Task-specific capital is what transfers. The occupation is the packaging.
• Weight the spread inside the field. Choosing physics over history matters less than what they do within either, and the Georgetown ranges say so directly.
• Prefer fields where the practice is hard to skip. The Stanford pattern concentrates in codified knowledge that a model can reproduce. Work that is learned by doing, in contact with people or physical reality, is exposed differently.
• Treat interest as a real input, not a soft one. Task-specific capital accumulates through years of effortful practice, and nobody sustains that in a subject chosen defensively.

And one thing to stop saying. Learn to use AI is not career advice. It describes a capability being deliberately engineered to require less skill every quarter. See why "learn to prompt" is weak career advice.

What this page will not do

It will not rank subjects by safety. No source found in preparing it supports doing so, and the one organisation that has scored its own attempt got the useful question barely better than a coin toss over a decade with no comparable technological shock in it.

If someone publishes a scored forecasting record for the generative AI transition, this page changes.

Related SuperSkills research

On the occupation question, which jobs are safest from AI and will AI replace my job. On the entry-level evidence, will AI replace entry-level jobs and the missing rungs. On children and AI directly, should children use AI. On what appreciates, staying valuable in the age of AI. On coding specifically, should I still learn to code.

Key research and primary sources

• United States Bureau of Labor Statistics. Occupational Projections Evaluation, 2006 to 2016.
• Waltmann, B. (2026). New Estimates of the Impact of Undergraduate Degrees on Lifetime Earnings. Institute for Fiscal Studies.
• Georgetown CEW (2025). The Major Payoff.
• Federal Reserve Bank of New York (2026). The Labor Market for Recent College Graduates.
• Brynjolfsson, E., Chandar, B. and Chen, R. (2026). Canaries in the Coal Mine? Stanford Digital Economy Lab.
• Gathmann, C. and Schoenberg, U. (2010). How General Is Human Capital? Journal of Labor Economics, 28(1).

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. This page declines to rank subjects by safety, and the reason is stated rather than implied. The Stanford employment figures are descriptive rather than causal, which their authors say and this page repeats. Not careers advice, and not a substitute for knowing the person choosing. Reviewed quarterly.