In part one I argued that "AI replaces white-collar work, therefore the economy collapses" doesn't survive contact with the circular flow — that the real outcome is turbulent repricing and a distributional fight, not a demand void. But there's a second-order question hiding under the first one, and it's the one I actually find harder.
If cognition becomes something you can buy by the token, how does a society retrain itself — not just individuals picking up new skills, but the institutions that produce skills, the laws that decide who's allowed to act, and the identities people build around what they do? That's the tectonic shift. The job-loss debate is the surface. The retraining of civilization is the substance.
Our schools are still industrial-era infrastructure
The system we use to produce workers predates almost everything about the economy it now feeds. It was built to turn out literate, procedurally compliant, specialized people who would be useful in roughly the same role for thirty or forty years. Front-load the education between ages five and twenty-five, stamp a credential on the output, and send it into a career whose skills stay valuable for a working lifetime.
That bargain rested on one assumption: that the half-life of a skill is measured in decades. AI breaks exactly that assumption. When the useful life of a specific competence is three to five years, you can't fix the system by editing the curriculum. You have to change what the system is for — from training people for roles to training them for adaptability, and from a one-time front-loaded event to something continuous across a working life. We have almost no institutional machinery for the second thing. The university, the licensing board, the professional guild — all of them assume the role you trained for will still exist when you retire.
The transition is not a switch
It's tempting to model automation as binary: human job, then replaced, then gone. That's almost never how it goes. The actual sequence is messier and slower: a human gets an AI copilot and becomes augmented; the role gets redefined around what the human still adds; parts of it get regulated; parts get automated outright; and eventually the new division of labor gets institutionally recognized — licenses, liability, billing codes catch up last of all.
That messiness is a feature, not a bug, because it's what gives people and institutions time to move. The danger isn't the destination; it's the speed, and the speed is set as much by law and politics as by capability.
Profession by profession
The abstract version is easy to wave at. The useful version is specific, because the shape of the shift differs by how much of the job is cognition, how much is physical, and how heavily it's regulated.
The teacher. AI absorbs content delivery, personalized tutoring, and grading — the parts we've long pretended a human had to do but never did especially well at scale. What's left is the part that was always the actual job: motivation, social development, group facilitation, knowing which kid is quietly drowning. The role drifts toward cognitive coach and community leader. The irony is that automation pushes teaching back toward what good teaching always was.
The programmer and engineer. Already in motion. AI writes the boilerplate, refactors, proposes architectures. The human moves toward system design, spec-writing, validation, integration, risk ownership. Fewer engineers, each with far more leverage, doing more systems thinking and less typing. The seat count shrinks; the seat that remains matters more.
The lawyer. Here the binding constraint isn't capability — it's regulation. AI can already draft, research, and predict case outcomes. But a licensed human signs, represents, and carries the liability. So the near-term human role is strategy, high-stakes arbitration, and relationship management, with the machine doing the production underneath. Legal recognition of AI in structured matters will lag the technology by years, because the entire profession is a liability structure wearing a robe.
The accountant. Routine compliance is the first thing to collapse — it's the most rules-bound, highest-volume, lowest-judgment work, which is exactly what models eat. The human survives in audit sign-off, regulatory responsibility, and advisory. Expect heavy compression in headcount and a shift toward the parts where someone has to be accountable.
The doctor. The most complex case. AI can diagnose, recommend, monitor, and triage — much of it already at or above the median clinician. But only a licensed physician can prescribe, bill insurance, and carry the liability. So the human becomes final authority, risk-carrier, and the person who actually talks to the patient, very likely paired with a mid-level provider plus AI handling the volume. A longer-term shift toward AI as a co-licensed entity is conceivable, but that's a decades-scale governance change, not a product release.
The banker and investor. The quantitative core automates heavily. The human persists where trust, large negotiations, political navigation, and genuine capital-allocation judgment live. Fewer seats, higher stakes per seat.
The artist. AI floods the low and middle tiers of content. Human art moves toward what a model can't counterfeit — identity, authenticity, the story behind the work. Scarcity migrates from production to provenance. The skill of making the image matters less; the fact that you made it, and why, matters more.
The entrepreneur. This one changes least, which is worth sitting with. Entrepreneurship is resource coordination under uncertainty, risk-bearing, political navigation, and capital orchestration — exactly the bundle that doesn't reduce to a well-specified task. AI massively augments founders; it doesn't remove the need for one. The high-leverage founder gets more powerful, not less. If you're looking for where human agency concentrates as cognition gets cheap, follow it here.
The pattern across all of them: the routine, high-volume, rules-bound cognition goes first; what survives is judgment, liability, relationship, and physical presence. "Up the value chain" turns out to mean toward the things that were never really cognition in the first place.
The legal and governmental bottleneck
This is the part technologists systematically underweight, and it may be the single biggest governor on the pace of everything above.
Right now, an AI cannot sign, cannot be sued, cannot carry malpractice insurance, cannot hold a license. That's not an oversight waiting to be patched — it's load-bearing. Enormous parts of the economy run on liability anchors: a specific human or corporation that is responsible when something goes wrong, who can be insured, fined, disbarred, or jailed. Until we answer "who is legally responsible for an AI's decision," full replacement simply cannot happen in regulated sectors, no matter how good the model gets.
The likely bridges are visible already: a human-in-the-loop requirement that keeps a person on the hook; corporate AI liability insurance; certification regimes for AI systems the way we certify medical devices; eventually, maybe, a "licensed AI entity" framework. Each of these is a way of re-creating an anchor for responsibility when the actor is a model. And each will arrive slowly, because liability law moves at the speed of litigation and legislation, not inference.
This regulatory lag is usually framed as friction. I think it's better understood as a shock absorber. It's a large part of why the transition will be survivable: even where AI can technically do the job, society can choose — for reasons of stability, trust, and the distribution of power — not to fully let it. The pace is political, not purely technological. That's not a bug in the system. In the near term it may be the thing that keeps the transition from becoming the rupture part one warned about.
The identity problem nobody budgets for
The subtlest cost isn't economic at all. We've trained people to become their roles. "I am a lawyer." "I am a programmer." The job title isn't just income; it's identity, status, the answer to who-are-you at a dinner party. When the category erodes — not vanishes, but stops meaning what it meant — the loss people feel is not only the paycheck. It's the self.
Continuous re-skilling solves the economic side: you can always learn the next thing. It does nothing for the identity side. A society that asks people to shed and re-form their professional self every few years is asking for something it has never asked before, and it has no rituals, institutions, or stories ready for it. Moving from a role-based identity to a capability-based one — "I am someone who can figure out hard things" rather than "I am an X" — sounds like a slogan, but it's a genuine cultural project, and it's the part I'd watch most closely for where the strain actually shows.
What becomes scarce
Step back and the whole thing resolves into a single question: if cognition gets commoditized, what becomes scarce?
Historically, scarcity never disappears — it relocates. When muscle got cheap, attention and skill became dear. If intelligence gets cheap, the new scarce goods are the ones intelligence can't manufacture: sustained attention, authentic trust, physical presence, the ability to coordinate other people, ownership of the capital stack, and — quietly the largest of all — meaning. The economy will reorganize around those the way it once reorganized around the things factories couldn't make.
So my answer to the second-order question is the same shape as the first. Society retrains itself the way it always has under a productivity shock — painfully, unevenly, slower than the technology and faster than the institutions would like — but it does retrain. The failure mode isn't that the machine got too good. It's that we let the institutions calcify, let the surplus concentrate, and let people's sense of themselves go un-tended while the ground moved. Those are choices. They're the ones worth arguing about — far more than whether the model can pass the bar.