You Can't Offload What You Never Carried: Who Becomes a Person?

I should confess something before this essay begins. I wrote it with an AI assistant open in the next window, and several times, deep in the argument, I caught myself wanting to ask it to simply finish the job. Summarize the papers for me. Smooth the transitions. Read the difficult passages on my behalf. Each time I had to make a small, deliberate choice to keep reading, to keep struggling with the sentences myself.

I am not telling you this to claim virtue. I am telling you because if I, a person whose entire life has been built out of exactly this kind of effort, who is literally in the middle of writing an argument against cognitive offloading, still feel the pull, constantly, then the pull is not a character flaw. It is the design.

And the question of what that design is doing to us is the question of this essay.

Not the question everyone is asking. The one underneath it.

The loop everyone argues about

By now the debate has a familiar shape. AI is coming for the jobs. The economy runs on a loop: firms pay labor, labor earns wages, wages become consumption, consumption becomes revenue, revenue pays labor again. If the machines do the work, who earns? And if nobody earns, who buys? You can't fire your customers and expect them to keep shopping.

The economists have a ready answer, and it deserves to be stated fairly, because its proponents are not fools. David Autor's version is the strongest: machines replace tasks, not jobs. Jobs are bundles of tasks, and when the routine tasks are automated, the value of the remaining human tasks rises. ATMs spread, and bank tellers did not vanish, they became relationship bankers, because branches got cheaper to run and customers still wanted a person. We have heard this panic before, the argument goes, about the loom, the tractor, the computer. Every time, new work appeared. The lump-of-labor fallacy, it is called, and the people who name it have two centuries of evidence behind them.

I have read this argument carefully, and I am not going to pretend it is weak. It may even be right about jobs. But notice what the entire debate, doomers and optimists alike, takes for granted: that a job is a way of moving money. The fight is over who keeps the income. Almost nobody is asking the other question.

A job delivers two goods, not one. The first is income. The second is something we barely have a word for.

The second good

Call it formation: the slow, effortful process by which sustained engagement with difficulty turns a person into someone. Skill is part of it, but only part. Work also gives structure to days, a reason to get up, a story you can tell about yourself at dinner, a place in other people's eyes. It is how most of us find out what we are good at, what we can endure, who we are.

Is this second good real, and is it really separate from the money? The bleakest natural experiment of our time says yes. Anne Case and Angus Deaton documented what happened when work left the American rust belt: deaths of despair, suicide, alcohol, opioids, rose among working-class Americans even as they fell across the rest of the rich world. Read their mechanism closely. It is not "people got poorer." The paycheck was partially replaced by transfers; what was not replaced was the role, the stable employment, the dignity of providing, the standing in the community that came with it. Follow-up research found that deaths of despair track the decline of religious participation at least as closely as economic decline. When the church emptied and the union hall closed, the job became the last load-bearing source of meaning. Then the job went too.

We built institutions to protect the first good: minimum wages, pensions, unemployment insurance. They were not gifts of the market; they were forced into existence, in the shadow of strikes and riots, by people who understood that systems collapse when people cannot live. For the second good, the formation that work performs on the person, we have built nothing. No institution, no policy, no vocabulary, no metric. It does not appear on anyone's dashboard.

Here is the optimistic story you have heard a hundred times, and it is time to stop telling it: once AI does the drudgery, humans will be free for creative work. Most people are not poets. We know this because we ran the experiment, retirement, pensions, aristocracies of leisure, and freedom from necessity did not produce a flowering of creativity. It produced a small self-directing minority and a great deal of drift. And the creative escape hatch is closing anyway: the poetry is being automated too, trained on all the poetry ever written. The claim that AI will free us for higher things assumes a latent population of artists waiting to be released. The evidence says that population is mostly imaginary.

So what is formation, exactly? Let me show you instead of defining it, because I happen to have spent my whole life as an unwitting experiment in it.

What formation is

Nobody who knew me at school would have predicted any of this. I was not the student about whom anyone said: he will do a PhD in AI, he will get a research offer from Oxford. (I got one, eventually. I never even listed it on my CV, by then it was just one more thing the process had produced).

From high school onward I wanted, in succession, to be a classical musician, then a conservatory professor, then a chemist, then a computer scientist, then an iPhone app developer, then a low-level systems programmer, then, I actually sat the admission exam, an intensive care clinician, then definitely not a clinician at all, then a data scientist in medicine, then in finance. For seven years I studied clarinet at a conservatory, a full degree running in parallel with high school and into university. Through my master's I served as a volunteer paramedic. Each of these was a genuine commitment, not a dabble; each ending was a real discovery, bought with time. I did not become a musician partly because I learned what a musician's life actually costs. I learned I did not want to be a clinician by getting close enough to feel it. The paths I abandoned formed me as much as the one I finally kept.

This is what the efficiency mind misses about a life like that. It looks like waste, all those dead ends, all that doubling of effort. But the knowing and the becoming were the same event. No advisor, human or artificial, could have handed me that knowledge as information: you are not a clinician. The sentence is cheap. The sentence arriving after years of attempting is a fact about who I am, deposited in me, permanently.

And here is the deeper structure, the one that matters for everything that follows: formation is path-dependent. Each stage of the struggle produced the person capable of the next stage. The version of me who received that Oxford offer did not exist at sixteen, and could not have been predicted at sixteen, least of all by me. The later stages of a formation require a person who does not yet exist. That is why it cannot be compressed, or parallelized, or shortcut, not because the technology isn't good enough, but because compression is the destruction. The struggle is not the price of the outcome. The struggle is the product.

This is not a metaphor, by the way. When London taxi drivers spend years acquiring the Knowledge, twenty-five thousand streets, their posterior hippocampi physically enlarge, and the enlargement scales with years on the job. Difficulty, sustained, deposits neural architecture. Struggle is the gym, and the gym builds an organ.

The machine that eats the gym

Now hold that picture against what we are building, and you can see the three distinct ways the machine eats the gym.

The first is removal, friction deleted for people already inside the pipeline. We no longer have to speculate about this; it has been measured. In a randomized trial of nearly a thousand high school students, those who practiced mathematics with a standard GPT-4 assistant solved dramatically more problems during practice. Then the researchers took the AI away and tested the students on their own. The AI group scored worse than the students who had never used AI at all. They had been carried to the summit and arrived unbuilt. A forklift can lift you to the top of the mountain; it cannot give you the legs.

The second is preemption, and it is worse. Removal takes friction from someone who has a self to defend; preemption means the first iteration never runs. A sixteen-year-old automating his homework is not offloading, you can't offload what you never carried. He is not delegating a process he owns; he is skipping the process by which he would have become someone who owns anything. The first generation raised inside the frictionless pipeline is not a lazier version of us. It is a different kind of outcome.

The third is the subtlest, and I feel it myself: resignation. When I was young, the world held still enough to aim at. If you found something you loved, you could spend entire days dreaming of becoming it, and the dreaming pulled the effort, the destination was still there when you arrived. That stability is gone. A student today cannot dream of becoming the radiologist, because the radiologist is being automated in the time it takes to become one. Why invest a decade in a self the market is depreciating in a product cycle?

The rational answer, increasingly, is: don't. Investment in formation becomes irrational at exactly the moment formation matters most. And we can already see the signal in the data: enrollment in computer science, the safest bet of the last twenty years, has fallen for the first time in two decades, while the narrow classes labeled "AI" fill up. Students are not fleeing the future. They are fleeing the deprecated destinations, herding toward whatever still looks alive, on ever-shorter horizons. I am not immune to any of this. Ask me what I confidently expect my work to look like in five years and I will give you an honest answer: I don't know, and neither does anyone. The difference is that I had my decades of dreaming first.

The demand is bottom-up

It would be comforting to blame the vendors, the evil owners of the future, force-feeding us convenience. I build these systems for a living, and I have to tell you: that is not what the market looks like from the inside. The demand is bottom-up.

Watch what happened when the first AI agents reached the professionals. The early adopters, brilliant, educated people, formed in the pre-AI world, did not test the new tools the way a surgeon tests a robot. A surgeon who finds the robot still needs him to finish the incision does not storm out calling the machine useless. Yet that is precisely what happened, at scale: knowledge workers tried to offload their entire jobs onto systems that were obviously not ready, hit the limits, and concluded the technology was "shit." Read that again. The complaint of the most privileged workers in history was that the machine was not yet good enough to take their job. They were not afraid of being replaced. They were disappointed they couldn't be.

Forty years of being told that labor is a cost center, and the professional class finally believed it, about themselves. The vendors are not imposing deskilling on an unwilling population. They are fulfilling orders. Please, take this from me.

The obvious rejoinder writes itself, and it should be written: every generation mourns the friction it lost. Socrates warned that writing would rot the memory, and he was right, it did, and we survived it. Calculators were going to destroy numeracy; Google was going to destroy knowledge itself. And the honest social-science picture of the smartphone generation is more contested than the headlines suggest: the best specification-curve analyses find the average effect of screen time on wellbeing is small, and researchers warn that the catastrophe narrative outruns the data. So why is this time different?

Because of where the gradient points. Every previous technology removed friction from the means and left the ends human. The calculator did arithmetic so the mathematician could do mathematics. The spell-checker fixed the typing so the writer could write. Each time, there was a higher layer of judgment to migrate to, and the migration compounded into livelihoods. AI removes friction from the ends themselves, the thinking, the drafting, the judging. The adaptive story assumes there is always a higher layer. The open question of this era is whether there is a layer above judgment. Maybe there is, taste, responsibility, care. But "we always found one before" is an argument from a sample in which a higher layer provably existed. The burden of proof has shifted, and it sits with anyone who assumes the ladder goes one rung higher.

Selfhood inequality

Put the pieces together and a new stratification comes into view. Not income inequality, we have institutions and vocabulary for that, but selfhood inequality.

On one side, the formed: people who built a self before the forklift arrived and now use the machine as a power tool. On the other, the unformed: people for whom the machine is not a tool but a substitute for ever having become anyone. Between them lies not a wall but a slope, because the formed are not safe either, we face a standing invitation to decay, renewed at every prompt box, as my own experience writing this essay attests. The difference between the formed and the unformed is not immunity. It is the ability to notice the pull and choose the effort anyway. Noticing is what formation buys.

What makes this inequality more fundamental than the income kind is that it cannot be redistributed. You can tax income and transfer it. You cannot tax a formed self and transfer it. There is no UBI for personhood. The policy imagination of our time, reskilling programs, universal basic income, jobs of the future, addresses exclusively the first good. For the second, we have zero institutions, zero metrics, and mostly zero awareness that there is anything to measure. We have instrumented everything about this transition except whether the young still believe in a future worth training for.

And here I want to be honest about the complication, because it does not weaken the argument, it locates it. Work became the last sanctioned source of purpose only because we let the alternatives rot first. The congregation, the union hall, the extended family, the civic association: their collapse predates AI by decades. There is nothing sacred about employment as the vessel of meaning; a medieval peasant had a third of the year as feast days and would find our arrangement pitiable. But both things are true at once: we dismantled every other source of meaning, and AI is now removing the last one. The first fact is an indictment of us. The second is an emergency.

Put it on the dashboard

So, argument, not lament. What would it mean to take the second good seriously?

The problem, I have come to believe, is not the technology's capability. It is which use of the technology the incentive structure prices. The same model is a master or a forklift depending on what the product chooses to remove. In that trial of a thousand students, there was a third group: students given a guardrailed version of the same AI, one that offered hints instead of answers, that insisted they do the thinking. Those students gained the practice benefit without the collapse in unaided performance. Same technology. Same students. Opposite formation outcomes. The difference was not in the model. The difference was a design decision.

Design decisions can be governed. We know how to do this, because we have done it, with something equally intangible. Nobody could measure privacy either, once. So we did not try to measure it; we forced a documented process. The data protection impact assessment: before you deploy, you answer in writing what data you take, why, at what risk, with what mitigation, and someone signs. The EU AI Act now does the same for fundamental rights: since this August, deployers of high-risk systems must conduct a fundamental rights impact assessment, a structured document with named triggers, impact levels, mitigations, and signatures. Soft goods, it turns out, can be governed, not by metrics, but by making their destruction a documented, accountable decision.

I am proposing the same instrument for the second good. A Formation Impact Assessment, completed and signed before an AI system is deployed into a workflow where humans still grow. Its questions are not exotic. What friction does this system remove? For whom was that friction formative, is this the junior's ladder or the senior's shovel? Which of the elements of formation does it touch, the resistance itself, the duration over which difficulty compounds, the real-world stakes, the witness of others, the story the person will get to tell? And where does that formation come from now? If the honest answer is nowhere, then de-formation is the design, and it should be signed as such, owned, like every other risk, by a named person, rather than dissolving into the general atmosphere of innovation.

I build AI systems for regulated enterprises. I have owned the security certifications, sat through the audits, filled in the assessments, defended the architecture in front of the people whose job is to say no. I know exactly how this instrument would land, because I have landed its siblings. And I am saying, as a practitioner and not a critic: this is the question I now insist on asking before I ship. Not whether the system works. What the system unbuilds.

We spent the last century learning, at enormous cost, that income is not a natural product of markets, it is a governance achievement, maintained or lost by choice. We are now discovering that the same is true of the person. Formation was never guaranteed either; we just never had to notice, because the friction came free with the world. It doesn't anymore.

The gym is being demolished to build the elevator. We should at least make someone sign for it.