Pivotum · Fall 2026 edition
Computer science graduates in the United States now face a higher unemployment rate than fine arts graduates.
That is Federal Reserve data, not a provocation.
We came across it trying to answer a question at our own kitchen table. We build AI systems for a living, and we have teenagers — which turns out to be an uncomfortable combination. You spend the day deploying the technology and the evening failing to answer your own child's questions about it.
So we went looking for a straight answer about which careers actually hold up. We could not find one. Careers advice has not caught up, the reporting is either breathless or dismissive, and the people who understand it best are inside these professions rather than writing guidance for families. In the end we decided to build the answer ourselves, which is what this is.
The Federal Reserve number matters because it describes a labor market that has quietly inverted the single most common piece of advice given to teenagers for the past twenty years. Learn to code. Do something with computers. Anything but art.
That advice is still being given. School careers offices give it. Well-meaning uncles give it at Christmas. It was reasonable when it was formed, and it is now, for a large number of the young people receiving it, actively bad.
But the part that should hold your attention is worse than "some degrees are riskier than others."
Your child can choose exactly the right profession and still end up on the wrong side of it.
A nurse working at a bedside scores 2.8 out of 10 on our exposure scale. A nurse doing utilization review — same license, same training, same person — scores 5.4. A trial lawyer scores 4.5; a document-review associate at the same firm scores 7.9. A senior software architect, 5.4; the graduate developer sitting next to them, 8.1.
Same qualification. Often the same building. Roughly double the exposure.
And this is why the question most parents are asking has stopped being useful. Is this degree safe? has no good answer any more, because professions are no longer moving as a unit. They are pulling apart from the inside, and which side of that split your child lands on now matters more than which profession they pick.
So we measured it. Twenty-eight careers, 158 distinct tracks inside them, every one scored the same way — a paramedic and a paralegal put through the identical process — plus reconstructed scores for 2023 and 2025 so we could see how fast the ground is moving. The method sits at the bottom of this piece and it is deliberately dull.
Seven things came out of it. Most of them contradict advice that is still being given in schools today.
Start with the reassuring half, because it is real. Almost nothing on our list is disappearing.
Legal employment in the United States hit a ten-year high in early 2026 — in the same period that first-pass document review was largely automated. Nursing faces a projected shortfall of more than 250,000 registered nurses by 2028. Accounting is losing people to retirement faster than it can replace them: applications to pharmacy school fell more than 60% in a decade, and the accounting profession has spent two years dismantling its own biggest barrier to entry because it cannot recruit fast enough.
This is why the public argument keeps talking past itself. One side points at automated tasks and predicts collapse. The other points at employment figures and calls it hype. Both are looking at real data and reaching opposite conclusions, because both are looking at the wrong level.
The movement is not between professions. It is inside them.
Every profession we scored contains protected work and exposed work, and the distance between the two is widening. A creative director sits at 5.6 while a production designer sits at 8.4. Bookkeeping is at 7.9; the CPA signing the audit is at 5.2. A court interpreter scores 4.3; a document translator using the same two languages scores 8.8 — the widest split we found anywhere.
These are not different careers. They are the same career at different altitudes, and the ladder between them is getting harder to climb — which is finding four.
For a family, that changes what you are choosing. Not a profession, but a position inside one. And most degree programs are still being sold as though the first thing were the only decision that mattered.
If you want one heuristic, this is it. Across every category we scored, the same work becomes roughly twice as exposed when it moves behind a screen.
| Same qualification | In person | On a screen |
|---|---|---|
| Nursing | 2.8 | 5.4 |
| Teaching | 3.6 | 6.8 |
| Medicine | 2.9 | 4.9 |
| Engineering | 4.0 | 6.3 |
| Physical therapy | 2.5 | 6.1 |
| Skilled trades | 2.5 | 5.0 |
| Construction | 3.3 | 7.4 |
| Transport | 3.4 | 7.3 |
We did not go looking for this pattern. It emerged from applying the same six factors to every track, and then held everywhere we tested it.
The reason is not mysterious. Work that requires a body in a room with another body — hands on a patient, presence in a classroom, judgment on an unfamiliar site — resists automation for reasons that have nothing to do with how clever the model is. Robotics lags software by decades, and it lags furthest in unpredictable environments.
The trades row is the one worth sitting with. A service electrician diagnosing a fault in a sixty-year-old building scores 2.5, while the same skills on a production line or in modular construction score 5.0. So the protection was never working with your hands. It was working with your hands somewhere unpredictable, which is a much narrower claim and a much more useful one.
This is the finding that most reliably surprises people, and it appeared independently in three different fields.
Ask most parents which is safer — a bioinformatician or a lab technician, a precision-agriculture specialist or a livestock farmer, a logistics planner or a delivery driver — and they will pick the technical one every time. The scores go the other way.
| The "modern" version | Score | The hands-on version | Score |
|---|---|---|---|
| Bioinformatics | 7.1 | Wet lab research | 5.2 |
| Precision agriculture | 6.2 | Field farming | 4.0 |
| Logistics planning | 7.3 | Local delivery driving | 3.4 |
| Revenue management | 6.6 | Kitchen and culinary | 3.2 |
The mechanism is the same in every case. Exposure follows the shape of the work, not its subject or its status. Does it take structured inputs? Produce text, code or analysis? Require no physical presence? Carry no licensed accountability?
Computational biology answers yes to all four — so it is exposed for exactly the same reason data analysis is exposed, because it is data analysis. The biology is the subject. The shape of the work is computation.
This has a blunt implication for anyone reasoning about their own field. Being close to a technology is not the same as being protected from it. A data analyst is not protected by understanding AI, any more than a copywriter is protected by understanding language.
This is the finding we consider most important, and the one that gets the least attention.
AI has not, in the main, replaced professionals. It has absorbed the tasks that junior professionals used to do. Document review. Boilerplate code. First-draft analysis. Grunt research. Content production. Reconciliation.
Those tasks were not just work. They were the apprenticeship. They were how a person built the judgment that eventually made them senior.
The numbers describe the squeeze precisely. Stanford's Digital Economy Lab, working from ADP payroll records, found entry-level software engineering postings fell 67% between 2023 and 2024. The share of tech jobs open to people with three years' experience or less went from 43% in 2018 to 28% in 2024. In law, only 38% of 2025 associate hires came directly from graduating classes, down from 46% two years earlier, while lateral hiring of experienced lawyers rose. Firms are buying judgment rather than building it. Professional services are now saying it out loud: in 2026 PwC's US CEO confirmed the firm had hired fewer consultants and accountants than two years before and was trimming its US campus-hiring targets, even as its new-business pipeline hit a record. PwC's own AI Jobs Barometer, drawn from more than a billion job postings, found roles demanding more human expertise growing twice as fast as the rest, with wages there rising 42% faster since 2021 — the routine tier contracting and the judgment tier commanding a premium inside a single firm.
So a profession can be simultaneously healthy and closed. Legal employment at a ten-year high, and the graduate route to it narrowing. That is not a contradiction; it is the defining pattern of this moment.
For anyone under twenty-five this matters more than any individual score on our list. A safe profession you cannot enter is not a safe bet, and the question to put to any path is not only whether the job will exist but whether you can still get in and get good before the bottom rungs go.
Licensing is the most durable protection we score. It moves at the speed of legislation rather than technology, and no advance in model capability repeals a scope-of-practice rule.
But we found that licenses vary enormously in how much they actually protect, and the difference is not the strength of the license. It is what the license reserves.
Consider three licensed professions:
Nursing's license reserves clinical practice — physical assessment, hands-on care, observation. Those acts are physically irreducible, so the license protects work that could not have been automated anyway. The protection compounds. Bedside nursing scores 2.8.
Law's bar card reserves the courtroom appearance and the signed advice — but not document review. Nothing in any statute says a human must read forty thousand documents. So when that became automatable, nothing stopped it. Junior law scores 7.9.
Pharmacy's license reserves dispensing and verification — which is precisely the structured, rules-based, high-volume task a machine performs well. The protection and the exposure land on the same activity. A doctorate-level clinical profession scores 4.9, worse than nursing, dentistry, veterinary medicine or physical therapy.
The same logic sorts entry routes, which is where it matters most for a family. Where a license mandates supervised experience hours, the on-ramp cannot be automated away. A nursing student cannot qualify without clinical hours. A doctor cannot exist without residency. An electrician needs an apprenticeship; a CPA, supervised hours; an architect, logged experience.
And there is one protection stronger than any license, which we found in an unexpected place. Statutory social work carries delegated state authority — the legal power to remove a child or restrict an adult's liberty. That is not a restriction on who may do the work. It is a coercive power vested in a named individual who can be questioned in court and struck off. There is no accountable subject for a system to be. It is the most durable protection in the entire index, and almost nobody thinks of social work in those terms.
So when you assess any career's regulatory protection, do not ask whether a license exists. Ask what it reserves, whether that act is one a machine can do, and whether it reserves the training as well as the practice.
Follow the logic above and it produces a result we did not want.
The single most protected route into work that we scored is not a degree. It is an apprenticeship.
It is inconvenient, it cuts against the assumptions of most people likely to read this, and we checked it more carefully than anything else in the index. It held.
Where entry is purely market-driven — software, marketing, design, business analysis, journalism — nothing prevents employers from simply hiring fewer juniors. And that is exactly what they have done. Where a license mandates the training, they cannot.
The skilled trades have both: paid apprenticeships, legally embedded in licensing, chronically undersubscribed while an aging workforce retires. A service electrician scores 2.5 and moved just 0.4 points in three years — the slowest movement in our entire index, at a time when entry-level software moved 1.8.
There is a detail here that we found genuinely striking. Microsoft's president has publicly identified the electrician shortage as the single biggest constraint on data centre expansion — ahead of chip supply, land or permitting. The technology squeezing white-collar entry work is simultaneously creating a shortage of electricians. That is not irony; it is the clearest possible illustration of where protection actually sits.
Construction management deserves the same attention and gets even less. It scores 3.3 at the site end — better than engineering, better than architecture, better than a business degree — with a BLS-projected 9% growth, roughly 46,800 annual openings and a median wage above $106,000. And it has something almost nothing else does: four independent routes in, including coming up through a trade. A career with one entry route is fragile. A career with four is resilient even when one narrows.
The paths with no tuition debt and a strengthening on-ramp are the ones most families never seriously consider.
But licensing is not a law of nature. It is a political artifact, fought over constantly, usually by people with a commercial interest in loosening it. Scope-of-practice battles run in state legislatures every year. Telehealth restrictions that had stood for decades were suspended in weeks in 2020 and never fully returned.
So the strongest protection in this index is also the one that could change fastest, and it would not change through a technical breakthrough. It would change through a committee vote. Anyone relying on this protection should watch statehouses as closely as they watch labs.
Every argument in this piece is a forecast, which is a weak thing to build a decision on. So it is worth knowing that two of the careers we scored already went through this, decades ago, and the results are in.
Dentistry automated its own supply chain. For most of the twentieth century, crowns and bridges were made by hand by skilled technicians — a genuine craft learned over years. Then CAD/CAM arrived: digital scanning, computer design, milling machines. The work moved from a bench to a screen.
The profession was completely fine. Dentists still practise, patients still get crowns, demand did not fall. But a whole tier of skilled, respected craft work contracted permanently, and the people in it largely moved out rather than up. Dental lab technicians now report median pay around $58,000, and the tier still scores 5.5 while chairside dentistry scores 2.7.
Agriculture went further and took longer. USDA's own data: self-employed and family farmworkers fell from 7.6 million in 1950 to 2.06 million by 2000 — down 73% — while total farm output nearly tripled. Hired farmworkers halved over the same period.
Three-quarters of a workforce gone. Output up threefold. And almost nobody outside farming experienced it as a crisis, because it happened gradually and food kept arriving.
Those two cases give us three findings more solid than any projection:
Automation removes a tier, not a profession. Which tier you are in matters more than which profession you choose.
The survivors got bigger, not safer. In agriculture, 6% of farms now produce more than three-quarters of all output. Scale absorbed the gains, not the individuals.
And the displacement was invisible from outside. Nobody outside dentistry noticed the lab tier contract. That is precisely what is now happening inside law, accounting and design — the profession looks healthy, and one layer of it has quietly stopped being a career.
There is a counterweight to all of this, and honesty requires giving it properly.
For nine years, driving has been the example everyone reached for — the job AI would eliminate first. Here is where that stands in mid-2026.
There are still no commercially scaled, fully driverless freight operations on US highways. Aurora, Kodiak and Plus run driver-out pilots on specific Texas and Arizona corridors, accounting for fewer than 50 trucks combined — against roughly 4 million Class 8 trucks operating nationally. Even optimistic industry forecasts put commercial autonomy at 5–10% of long-haul lanes by 2032.
That is about one truck in eighty thousand, nine years in.
Meanwhile, the back office of that same industry was quietly rebuilt. Dispatch, route planning and freight brokerage have been substantially automated, with industry analysis suggesting those functions can be offloaded at 30–50% overhead reduction. Logistics planning scores 7.3 against local delivery driving at 3.4.
The person behind the wheel is better protected than the person planning their route.
We include this because it is the most confidently repeated prediction about any career we score, and it has not happened. It should make you appropriately sceptical of confident predictions — including ours.
Every protection we score has a shelf life, and treating them as equivalent is a mistake.
Regulatory and licensing protection is the most durable — subject to section 5's qualification about what it actually reserves. It moves at the speed of legislation and public trust, not technology.
Human trust and accountability is nearly as durable. People want a human present when they are frightened, and institutions want a human to hold responsible when something goes wrong. Both are socially rooted and erode slowly, if at all.
Judgment under novelty is real but eroding — with one exception worth naming. In most fields, novelty comes from complexity, and complexity can be learned. In cybersecurity it comes from an adversary who is deliberately manufacturing it, and who changes whenever a defensive system learns the pattern. There is no equilibrium where that work becomes routine. Adversarial novelty resists automation better than complex novelty does — which is why threat hunting holds 5.0 despite having no license, no physical requirement and a fully digital workflow.
Physical protection is the strongest today and the most fragile over a long horizon. It is the one thing physical AI directly targets, and a seventeen-year-old entering a hands-on profession will practice for forty years.
Agriculture is where that last assumption has been tested longest, and it held least well: three-quarters of a workforce displaced by machinery over fifty years. Anyone relying on physical protection should weight that history heavily.
And the mechanism to watch is not robots getting better at unpredictable environments. It is industries moving work into predictable ones — prefabrication in construction, controlled-environment growing in agriculture, standardised high-volume kitchens in hospitality. Automation performs well in controlled settings. Moving the work indoors raises exposure without any breakthrough at all.
If you are choosing a degree: the choice of program now matters more than the choice of subject, and how much more depends entirely on the field. Where employers have pulled back from supplying entry-level experience — software, law, marketing, design, business analysis — the degree has to supply it instead, and most degrees are not set up to. Where the on-ramp is still intact because a license requires it, this matters far less. In the first world, mandatory placements, real industry projects, employer pipelines and honest first-destination outcome data stop being nice-to-haves and become the qualification itself.
If you are advising a teenager: widen before you narrow. The safest category we scored was not technology or finance. It was work that helps people — nursing, medicine, dentistry, veterinary medicine, allied health, psychology, social work, teaching. Eight of those nine fields have a track scoring below 4.0. Very few families are told this.
If you are already mid-career: the useful question is which side of the split you are on. Work that is executional, screen-based, and pattern-following is where the pressure is. Work that carries accountability, requires presence, or handles the cases nobody wrote a procedure for is where it isn't. And be careful about promotions: in construction, transport and hospitality, the conventional next step moves toward the exposed end. Site superintendent 3.3, preconstruction 7.4. Delivery driver 3.4, logistics planner 7.3. Those look like advancement.
And for everyone: the most valuable emerging skill is not operating these tools. It is recognizing when they are confidently wrong, which requires knowing your own field well enough to catch a plausible-sounding error. Courts have now sanctioned lawyers in roughly 1,600 documented proceedings for filing AI-generated citations to cases that do not exist — every one of them a person who signed something they had not read. That is domain expertise doing a new job, not a new skill replacing it, and it means the fundamentals matter more now rather than less.
One more thing, which applies to all four groups. Whatever you conclude from this, put a date on it.
This is our first edition, so we cannot yet tell you which of our own judgments aged badly. We will be able to in the spring. But the backcast — the reconstructed 2023 and 2025 scores — already contains something worth flagging. Every heavily exposed track we scored moved less between 2025 and 2026 than it did in the two years before. Junior law rose 1.2 points and then 0.7. Entry-level software, 1.1 and then 0.7. Production design, 1.1 and then 0.6. Content marketing, the same.
Meanwhile the two safest clinical tracks in the whole index did the opposite: bedside nursing moved 0.3 and then 0.4; patient-facing medicine 0.2 and then 0.3. Both accelerated slightly.
The numbers are small and we would not build a thesis on them yet. The shape is still worth noticing. It suggests the easy displacement has largely happened at the exposed end, and that what movement is left is slower — and starting to reach into work that had previously not moved at all.
These scores measure exposure to what AI can already do. They do not measure how much of it your future employer has actually installed.
Those are very different things, and the gap between them is often years. Capability arrives first at large, well-capitalized organizations with the budget and technical staff to deploy it. It arrives much later at small practices, rural providers, public-sector employers, regional firms and anywhere operating on thin margins.
So the same job title can carry quite different day-to-day exposure depending on where you do it. A nurse at a large teaching hospital running ambient documentation is already working in the world these numbers describe. A nurse at a small rural facility may not encounter any of it for years. Both are nurses. Only one is living in the present tense of this index.
We score the frontier on purpose, because that is what makes it a leading indicator. But if you are making a decision on this basis, the lag is real and worth understanding correctly. It is a buffer, not a shelter. It buys time. It does not change direction, and it is not something you can plan a thirty-year career around.
If you've read this far and want to know why you should believe a word of it, this is that section.
We are not a research institute. We are practitioners who build these systems for a living, and we have applied one consistent framework to twenty-eight careers and 158 tracks inside them, which as far as we can tell nobody else has done.
What we did. Six factors, fixed weights, applied identically to every track. Two of them measure exposure — how much of the work is automatable, and how badly the entry route has eroded. Four measure protection — physical presence, human trust and accountability, regulatory reservation, and judgment under novelty. The protection factors carry more weight in total than the exposure factors, which is deliberate: it means protection can outweigh exposure. Teaching rates 7.7 on automatability — higher than nursing or medicine — and still scores 3.6, because everything else runs the other way.
Where the numbers come from. The AI analysis is our own judgment, applied consistently and published in full. Everything else — pay, turnover, debt, growth projections, workforce data — is sourced to primary bodies wherever it exists: the Bureau of Labor Statistics, the AMA, the ABA, the AVMA, the AICPA, USDA, the Federal Reserve Bank of New York, Georgetown's Center on Education and the Workforce. Where we have relied on industry commentary rather than a professional body or a statistical agency, we say so in the profile.
We publish the whole audit trail. Every factor rating for every track, the 0–10 anchors, the arithmetic connecting factors to scores, a sensitivity analysis showing which disagreements would actually change a band, and the specific indicators that would move each score. A score nobody can check is an opinion with a decimal point.
We correct ourselves in public. Writing this index changed several things we thought we knew. We had written that physician burnout was "high and rising" — it is high and falling, three consecutive annual declines, and the popular narrative was a decade out of date. We had written that half of nurse turnover happens in the first two years; the data says retention in those years is actually strong, because residency programs hold people, and turnover spikes at around thirty months when the support comes off. That changes the advice entirely.
Our framework cannot see demand. A task can be highly automatable and highly in demand at once. Data science is the clearest case: we rate the entry tier among the most exposed work we measure, while BLS projects the occupation growing 33.5% — the fourth-fastest in the economy — and entry-level salaries have risen roughly $20,000 in a year. Prices do not usually climb when demand collapses. We hold our score and publish the disagreement, and a reader who weighs it differently is reaching a defensible conclusion.
It cannot see oversupply. Business is the clearest case here. The Federal Reserve Bank of New York puts recent-graduate underemployment at 41.5% — against 12.8% for nursing, the lowest of any major it tracks. That gap is not about intelligence or effort. It is about whether a degree points at a job that legally requires it. For business, the thing we cannot measure probably matters more than the thing we can.
It cannot see industry economics. Journalism has contracted roughly 60% since 2008 for reasons that finished before generative AI existed. Our score describes the automation; the economics did the damage.
And it cannot see whether the work is worth doing. Pay, hours, debt, satisfaction and meaning sit entirely outside these numbers. Physical therapy scores 2.5 and carries debt-to-income ratios ranging from 0.82:1 to 4.42:1 depending on which program you attend — a difference of over $300,000 for the identical license. Social work has the most durable protection in the index and turnover roughly six times the occupational average. A career can be automation-proof and still be a difficult decision.
We would rather tell you that than let a reassuring number do work it has not earned.
Twenty-eight careers. 158 tracks. Six factors, fixed weights, applied identically to every one. Scores run 1–10, where 10 is most at risk. 2023 and 2025 figures are retrospective reconstructions using the current methodology — our best assessment of what each factor would have rated at the time, not archived measurements.
Every career has a free page with its scores, the reasoning, and the sources. The full profiles go considerably further: factor-by-factor analysis, the honest downsides, what to ask an admissions office, discussion questions to work through with your child, and a technical appendix showing exactly how each number was reached.
We re-score everything every six months, and we publish what moved — including where we were wrong.
Computer science is free in full, so you can see what the paid ones look like before spending anything.
Pivotum. We build AI for a living, and we’re figuring out our own teenagers’ futures alongside yours.