The Wrong Question

“Which jobs are AI-proof?” is the question everyone is asking. It is also the question that guarantees you will get the future of work wrong.

TL;DR: Stop asking which jobs are “AI-proof” — it’s the wrong question. AI doesn’t replace jobs; it replaces tasks, hollowing out the codified, on-a-screen work that entry-level knowledge roles were built on (Stanford’s data already shows a 13% employment drop for young workers in the most exposed fields). But the same model that commodifies your expertise can amplify it instead — and which one happens is a deployment choice, not a property of the technology. What survives isn’t a tidy list of safe professions; it’s a set of specific human capacities: framing the problem rather than solving it, knowing where AI’s “jagged frontier” fails while it sounds certain, owning decisions someone has to be accountable for, and the taste to tell good from merely fluent. The future of work isn’t happening to us. Task by task, we’re deciding it.

Complimentary reading link

Talk to a worried parent of a high schooler, a recent / soon-to-be graduate college grad, or walk into any conference on artificial intelligence and the future of work, and you will hear the same anxious question asked a dozen ways: Which jobs are safe? The answers arrive on cue — learn a trade, lean into creativity, double down on empathy, because robots can’t lay pipe or hold a dying patient’s hand. There is comfort in the list of survivors. There is also a quiet intellectual surrender in it, because the list assumes the wrong unit of analysis and the wrong theory of how this technology actually bites.

The job is not the unit of disruption. The task is. And once you accept that, the entire conversation reorganizes itself — away from a defensive search for shelter and toward a much harder, more useful question:

what, specifically, does scarce human expertise consist of once the codifiable parts of my work are free?

That is the question the serious researchers are now asking, and their answers are far more interesting than the airport-bookstore version of this debate.

Jobs are bundles of tasks, and the bundle is coming apart

The most important idea in this field belongs to **David Autor, the MIT labor economist **whose work on trade and automation has shaped how a generation of policymakers thinks about disruption. Autor’s reframing is deceptively simple: an occupation is not a single thing you either keep or lose. It is a bundle of tasks, each requiring different expertise, and automation does not remove jobs so much as it removes tasks from the bundle. When enough of the bundle is hollowed out, the job changes shape — or disappears.

The consequence is sharper than it first appears. Autor draws a distinction that I wish every executive deploying AI would tattoo on the inside of their eyelids. There are, he argues, two kinds of tools. An automation tool eliminates expertise. A collaboration tool is a force multiplier for it. The same large language model, pointed at the same workflow, can do either — and which one it does is not a property of the technology. It is a property of the choices made by the people who deploy it.

This matters because expertise, in Autor’s framing, has economic value only when it is both useful and scarce. AI can attack the scarcity directly. If a system makes a once-rare competency trivially easy and generic, it does not augment the expert — it commodifies the expertise and collapses its market value. His own grim example is the London cab driver, who spent years memorizing every street in the city only to watch that hard-won knowledge rendered economically superfluous by a navigation app. The expertise didn’t get cheaper. It stopped being expertise.

Here is the uncomfortable part Autor does not flinch from: when a task is fully automated, the relevant expertise is no longer needed anywhere in the economy. It is not relocated. It is extinguished. And there is no economic law guaranteeing that the new work AI creates will arrive fast enough, or land on the same people, as the work it destroys. Recent evidence, he notes, suggests automation is currently outpacing new-work creation. The displaced cabbie does not become the prompt engineer.

The data has stopped being theoretical

For years, this debate ran on speculation and exposure scores — clever estimates of which occupations could be affected. That era is over. We now have receipts.

In August 2025, Stanford’s Erik Brynjolfsson, with Bharat Chandar and Ruyu Chen, published Canaries in the Coal Mine?, built on ADP payroll records covering roughly 25 million American workers. The finding is the first large-scale, near-real-time signal of generative AI reshaping employment, and it is precise enough to end a lot of hand-waving: early-career workers, ages 22 to 25, in the most AI-exposed occupations — software engineering, customer service — saw a 13 percent relative decline in employment since late 2022. Older workers in the very same roles held steady or grew.

Read that again, because the age gradient is the whole story. This is not a story about coders being replaced. It is a story about a specific kind of knowledge being replaced. Brynjolfsson’s explanation is that large language models are trained on exactly the book learning that universities sell — the codified, general, written-down knowledge that a new graduate brings to a first job. The model arrives already fluent in it. What the model does not have is the tacit, situated, hard-to-codify knowledge that accrues only through experience. So the buffer against displacement turned out not to be your credential. It was your years.

And the augment-versus-automate split showed up cleanly in the same data: where firms used AI to augment workers, employment did not fall. Where it substituted, it did. Brynjolfsson’s summary is the entire policy debate compressed into a sentence — the workers using AI to augment rather than automate their work have seen gains. The technology rearranged employment; it did not simply subtract it.

Anthropic’s Economic Index, which maps millions of real AI conversations onto the U.S. Department of Labor’s roughly 20,000 cataloged work tasks, fills in the texture. As of its early-2026 reporting, augmentation has overtaken automation as the dominant interaction pattern — people collaborating with the model more often than fully delegating to it. The wage curve is the detail worth dwelling on: both the lowest-paid and the highest-paid occupations show low AI use. Heavy use clusters in the mid-to-high-wage knowledge work in between — the programmers, the analysts, the copywriters. The floor is protected by physical dexterity the machines can’t match. The ceiling is protected by something else, which we’ll come to.

The frontier is jagged, and that is the most useful thing to know

If the Stanford data tells us that disruption is real, the most rigorous experiment we have tells us how it behaves at the desk. In 2023, a team spanning Harvard, Wharton, and MIT — Fabrizio Dell’Acqua, Ethan Mollick, Karim Lakhani, Katherine Kellogg and others — ran a preregistered field experiment with 758 Boston Consulting Group consultants, fully 7 percent of the firm’s individual-contributor workforce. They gave them real consulting work, with and without GPT-4, and measured what happened.

For tasks that sat inside the AI’s competence, the gains were enormous: roughly 12 percent more tasks completed, 25 percent faster, 40 percent higher quality. This is the number that gets quoted in keynote decks. The number that should get quoted is the other one. For tasks that sat outside the model’s competence — and this is the catch — consultants using AI were about 19 percent less likely to reach the correct answer than those working without it. The model produced fluent, confident, wrong work, and capable professionals trusted it.

The researchers named this the jagged technological frontier: a boundary where AI is brilliant at some tasks and useless at others, and where the two can sit side by side in the same workflow at seemingly identical difficulty, with no label telling you which is which. There is no user manual for the frontier. You only know where the edge is if you already have the expertise to recognize it.

That single fact reframes what human skill now means. The study identified two productive working styles — the centaur, who cleanly divides labor and hands the AI only what it does well, and the cyborg, who interleaves with it continuously. Their later work added a third and dangerous mode: the worker who simply hands the whole task over and stops thinking. Dell’Acqua had already shown the mechanism in a separate study of recruiters: given a high-quality AI, they grew lazy and made worse decisions than colleagues relying on their own judgment. The frontier punishes the credulous.

There is one more finding from that work that I think about constantly. The AI-assisted output, while higher in quality on average, was measurably more homogeneous — same-y, in the researchers’ word — across the group. When everyone has the same brilliant collaborator, everyone’s work converges. Distinctiveness becomes scarce. Scarcity, as Autor reminds us, is where value lives.

What actually resists, and why

Now we can rebuild the “safe jobs” list — not as a list, but as a set of properties, which is the only version that survives contact with a new model release.

Three clusters genuinely resist, each for a different and instructive reason.

The first is physical work in unstructured environments — electricians, plumbers, HVAC technicians, mechanics. The barrier here is Moravec’s paradox, the old observation that the things we find intellectually hard are easy to automate, while the sensorimotor competence a toddler has is fiendishly hard. You can program a robot to torque a bolt on an assembly line. You cannot program it to diagnose a leak behind forty-year-old plumbing in a building with no blueprint, in a crawlspace, in the dark, improvising against whatever it finds. The U.S. Bureau of Labor Statistics projects hundreds of thousands of annual openings in these trades through the early 2030s, and they pay well, and they do not require a four-year degree. The floor of the wage curve is concrete, literally.

The second is embodied care and high-trust human presence — nurses, therapists, physical therapists, eldercare workers, surgeons. Here the barrier is that the human presence is the product. AI can deliver the information in a diagnosis; it cannot deliver the reassurance, the read on a frightened patient, the practical bedside trick that one human passes to another. In the moments that matter most to people, the demand is specifically for a human who understands what the other human is going through.

The third is the most overlooked and, for senior professionals, the most important: accountability-bearing judgment. Physicians, judges, licensed engineers, executives. The barrier is not capability — a model may well outperform the human on the narrow analytic task. The barrier is responsibility. Some decisions cannot be delegated to a screen because someone has to be answerable for them, in law, in ethics, in front of a board or a court or a family. Liability is a moat AI cannot cross, because a model cannot be held responsible. That moat will widen, not narrow, as the technology grows more powerful and the stakes of trusting it rise.

And then there is the middle — the zone almost everyone reading this actually occupies. Radiologists, lawyers, architects, teachers, analysts, designers. These jobs are bundles where AI does specific tasks superbly while the job survives as a re-bundled role around what’s left. This is the genuinely hard place to be, not because these jobs vanish, but because the transition is wrenching and unguided. The work that disappears and the work that becomes newly valuable rarely belong to the same person on the same day.

The casualty list, where the data already shows movement, is consistent across every serious source: entry-level software engineering, customer service and call centers, copywriting and content production, graphic design, paralegal and legal research, bookkeeping, data entry, translation. The pattern is not “low-skill.” The pattern is codified, language-mediated, and performed on a screen. AI is not coming for the jobs in the physical world first. It is coming for the ones that already lived inside a document.

The human attributes that actually matter — and they are not the ones on the poster

So we arrive at the question worth asking, and I want to answer it without the usual incense. “Creativity and empathy” is not wrong, exactly. It is just too vague to act on, and vagueness is its own kind of slop. Here is what the evidence actually points to.

Problem framing over problem solving. The model is an extraordinary solver once a problem is well-specified. Specifying it — choosing which problem is worth solving, under ambiguity, with missing information and competing stakeholders — remains stubbornly human and is the single highest-leverage skill of the next decade. The person who can define the question will direct the people and the machines who answer it.

Frontier judgment. The ability to feel where the jagged edge is — to know which tasks the AI will nail and which it will fail while sounding completely certain — and to validate accordingly. This is not a soft skill. It is the hardest skill, and the BCG research is unambiguous that it requires deep domain expertise, which becomes more critical as the technology improves, not less. The cruel irony of this moment is that you need expertise to safely use the tool that threatens to commodify expertise.

Accountable judgment. The willingness to own a consequential decision and be answerable for it. This is structural, not sentimental, and it is the durable core of every senior role.

Tacit, embodied skill. The dexterity, the situational improvisation, the relational trust — the knowledge that lives in the hands and in the room, not in the training corpus.

Taste. When generation is free and output converges toward the same-y mean, the scarce thing is discernment: the judgment to know what is actually good, what is distinctive, what is worth shipping. Editorial judgment was always undervalued. It is about to be repriced.

And the discipline not to fall asleep at the wheel — to keep exercising independent thought rather than degrading into the worker who rubber-stamps the plausible and wrong.

This is a choice, not a forecast

The thread running through all of it is the one Autor insists on and the one the doomers and the boosters both keep missing: the outcome is not predetermined. The same model can eliminate expertise or amplify it. Whether your work is commodified or elevated depends far less on what the technology is capable of than on how your organization, your profession, and you choose to deploy it.

That is an inconvenient conclusion, because it removes the comfort of inevitability. You cannot retreat to a list of safe jobs and wait out the storm. But it is also the hopeful one, and the only one that respects the agency of the people living through this. The future of human work is not something that is going to happen to us. It is something we are, task by task and decision by decision, deciding. The people who understand that — who learn to direct the machine, judge its output, and stay accountable for the result — will not be the ones it replaces. They will be the ones it makes scarce, and therefore valuable.

The wrong question asks where to hide. The right question asks what only you can be trusted to decide.

References and further readings

Sources informing this piece: David Autor & Neil Thompson, “Expertise” (MIT/NBER, 2025) and “Applying AI to Rebuild Middle-Class Jobs” (NBER); Erik Brynjolfsson, Bharat Chandar & Ruyu Chen, “Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence” (Stanford Digital Economy Lab, 2025); Fabrizio Dell’Acqua, Ethan Mollick, Karim Lakhani, Katherine Kellogg et al., “Navigating the Jagged Technological Frontier” (Harvard Business School Working Paper 24–013, 2023) and “Cyborgs, Centaurs and Self-Automators” (HBS WP 26–036); the Anthropic Economic Index (2025–2026 reports); and the World Economic Forum, Future of Jobs Report 2025.