We’re Training People for the Wrong Future

Late on a Sunday, my phone lit up with a text from my former lab mentor at Stanford. The message was a question, not a hello. There is a ton of training, he wrote, at universities and companies, all of it trying to teach people how to use artificial intelligence so they can stay relevant. Whenever I attend these events, they feel backwards-facing. Here are the tools that exist right now. A few months later, they are out of date. How am I supposed to keep up?

He was asking the right question, but I did not have an answer.

When change is slow, being a year late costs you a year. When change accelerates, every leap becomes a new year of falling behind. You never catch up before the next one arrives, and the inequality starts to deepen exponentially.

Institutions have always lagged technological change. What is new is that the lag is becoming existential: by the time they adapt, the world they are preparing people for may already be gone. Universities approve courses on a model six months before it is obsolete; companies train workers on interfaces that have already been replaced; the skill we teach ages faster than the diploma we issue.

This is not a curriculum problem, it is an actual category error. We are training people as human capital, bundles of skills to be accumulated and sold, when the era that exchange belongs to is ending. The work of today is to train people not for sharper skills but for a different relationship with their own working lives. Call that human agency.

Most AI training programs in the country are solving last era’s problem, and until we change what we are training people for, no curriculum will catch up.

Pew Research found last year that 52 percent of employed American adults are worried about AI’s impact on the workplace. Thirty-two percent expect fewer opportunities for themselves, while only six percent expect more. Underneath those numbers is a twentieth-century career map handed to a twenty-first-century labor market that no longer matches the territory. The advice we keep giving, choose the AI-proof major, learn the new tool, retrain into the next role, comes from a world where the tools and the work stayed put long enough for the advice to mean something. Now, that world is gone.

I have been mentoring a sixteen-year-old for about a year over video calls. Last spring she announced she wanted to be a journalist. By summer she had switched to computer science because a guidance counselor told her it was AI-proof. Three months later she switched again, to nursing, because someone on TikTok said healthcare was the only sector machines could not fully replace.

The last time we spoke, I asked her what she actually wanted, and after a long pause she said, I don’t know, I think I just want to pick something that won’t make my parents worried.

Gallup and Lumina found that 42 percent of bachelor’s-degree students have considered changing majors because of AI, and that sixteen percent already have. My mentee is not in that sample — she is only sixteen — but she is already living inside the same logic. She has been advised, three times, to redesign her future around what a thirty-year-old technology forecaster could not predict.

The question driving this essay is one I sit inside every day at OpenAI, from a vantage point that does not let me look away.

The standard debate about AI and work asks whether the technology will destroy more jobs than it creates, or create more than it destroys. Both halves sit inside the same frame, which assumes the question is about the volume of jobs.

Productivity gains and job destruction can coexist any time output per worker rises faster than demand. If one lawyer with AI can do the work of five lawyers, but the world does not need five times more lawsuits, law firms will employ fewer lawyers even as legal productivity soars. AI can be a complement for some workers and a substitute for others. It can make senior partners formidable while it hollows out the apprenticeship ladder of junior associates beneath them.

That second part is what I have not heard discussed with enough alarm. Every profession has used the apprenticeship ladder to teach itself to its next generation, and junior salaries were the cost of the lesson. When ad agencies stop hiring twenty-three-year-olds to draft headlines, they end the line of people who would have become the creative directors ten years from now. The same is true at law firms whose document review is done by software, at consultancies whose first-year analysts are now AI tools, at newsrooms whose entry-level beats have been folded into content workflows. What we are deciding, underneath the headlines about job losses, is whether anyone will learn the work at all.

Professions break the moment fewer humans become sufficient.

We have lived through structural changes before. In 1900, 41 percent of the American workforce was on farms, and by 2000, it was roughly 2 percent. In every previous transition, technology automated some tasks, workers slid into the adjacent ones, education caught up over a generation, and new sectors eventually absorbed the displaced.

The shape of this transition is different. The adjacent task is also being automated, the new skill depreciates before retraining pays off, and the institution doing the retraining moves slower than the technology changing the target.

You do not save scribes by teaching them to write faster after the printing press.

This year’s Stanford Human-Centered AI Index, a project I worked on as Erik Brynjolfsson’s teaching assistant, found this. On most contested policy questions, experts and the public face the same direction. They may disagree about magnitude, but they agree about which way the wind is blowing. But on AI, that pattern breaks. Seventy-three percent of AI experts expect a positive impact on how people do their jobs, while twenty-three percent of the public expects the same. The gap reaches fifty points, with similar splits on the economy and on healthcare. The people building the technology and the people on its receiving end have stopped arguing about it. They are looking past each other. A society survives a hard transition only when the people designing the future and the people living inside it are describing the same world.

A serious essay has to concede the strongest form of the opposing case, and the optimists may be right in the long run. Institutions could catch up, new sectors could emerge to absorb displaced labor, and the public may eventually see what the experts see. None of it helps the people standing on the disappearing rungs right now. The eventual is what people who can afford to wait point to when they are asking other people to wait.

For a hundred and fifty years, the human capital model held: individuals accumulated marketable skills and sold them as labor. When intelligence at human level becomes a service, copied at near-zero marginal cost and available on demand, the firm no longer needs the human as its primary cognitive input.

Skill is no longer the scarce thing, direction is. A person with a clear mission, a sense of taste, real judgment, and the willingness to take responsibility can now coordinate computation, capital, networks, and tools at a scale that previously required an entire company.

What this calls for is a different self-conception. The person who used to rent skills to a firm now has to author the work itself.

The word agency has drifted. People hear it and think Silicon Valley grit, self-actualization, or the personality trait of accomplishing your dreams regardless of circumstance. The definition I am offering is a more practical one: the capacity to deliver an outcome in the world by choosing an aim, directing the available tools toward it, and remaining responsible for the result.

This is the place the argument is most fragile as agency is unevenly distributed for two specific reasons.

First, scarcity disciplines imagination. A family that needs stability does not optimize for curiosity, it optimizes against disaster, and agency requires a margin of safety that allows for wrong turns.

Second, AI is removing the ladders where agency was historically built. Junior jobs supplied exposure, apprenticeship, small-stakes decisions, and a path from execution to judgment. Without that path, agency stops happening even to the people with the most luck.

The single mother working two jobs does not have a lunch break in which to redesign herself as a one-person enterprise. To frame the move from human capital to human agency as a matter of mindset is to mistake the privileges some people have for the wisdom they think they earned. I have caught myself doing it.

Agency cannot be a luxury good available only to people whose lives already had the margin for it. Institutions exist precisely to make capacities like this widely available. The era of human capital, for all its limits, asked very little of the person inside it. The era of human agency, if it is going to be real for more than a sliver of the workforce, needs to come with the material conditions for ordinary courage. It has to become a public good.

1. Education must teach taste, judgment, and end-to-end authorship.

The capacity for agency starts with the things schools have stopped teaching: taste, judgment, and the ability to drive a project to completion. Today, education trains students to complete assigned tasks and are rarely asked to own the end-to-end pipeline of something they conceived themselves.

The reform is to push students to drive independent projects through to completion, whether artistic, research, or business, using AI to execute from idea to final product. Teach judgment alongside production. If AI can produce fluent answers instantly, education has to become obsessed with what only judgment can do: evaluating outputs, comparing alternatives, knowing when the metric lies.

This is where the inward question lives. A student trained to drive her own projects practices asking what she would want to be doing if no one were paying her for it.

2. Firms must transmit experience and intuition through redesigned apprenticeship.

Firms have historically produced senior workers through informal apprenticeship. AI is removing their incentive to keep doing it, because the entry-level tasks that used to teach intuition can now be done more cheaply by software.

This is a free-rider problem. Every firm benefits from having experienced workers available to hire, yet no firm wants to pay to develop them. The fix is to deliberately re-create the learning function of junior work, through protected junior rotations, simulation environments where novices make consequential decisions with supervision, and structured apprenticeship roles that survive the AI cost-savings argument.

Medical residency is the clearest American proof this can work. Medicare funds graduate medical education at roughly $16 billion a year, supporting 150,000 residents in multi-year, paid, real-stakes training. We have only done it for one profession.

3. Government must realign the incentives and fund the material conditions.

This is a classic incentive-misalignment problem, and one of the clearest places for government to act. France’s alternance program offers one useful model: graduate students train part-time at companies, embedded in real teams while firms benefit from cheap student labor. In a world where AI provides equivalent or cheaper labor, that benefit disappears, and public investment has to restore the incentive.

The funding asymmetry tells the policy story. Pell Grants are funded at roughly $30 billion a year, and federal apprenticeship support runs at under $1 billion. We have decided which kind of training matters, and we decided it for the world we used to live in.

Government also has to provide the material conditions under which ordinary people can practice agency: stipends, transition income, access to tools, mentorship, time to experiment, the right to a second chance. You cannot train people for agency while pricing experimentation as a luxury good.

Schools cultivate the capacity, firms preserve the structures where it is transmitted, and government funds the conditions for both. Agency has to become a public good.

Sometimes I think about my sixteen-year-old mentee, and what I would have wanted someone to tell me at her age. The question I kept being asked, what should I become so the market would need me, was the wrong one. The right one was closer to this: what would I want to be doing if I never got paid for it. No one should be made to hedge their life inside a changing world before they have begun to live it.

She is going to have to ask that question for herself, and so am I. The institutions that were built to ask it for us were built for a century that has already ended.

I’m just so unbelievably excited to be alive in this era.

With Love,

Houda