AI does not eliminate work evenly, and the even-cut assumption is where most predictions go wrong. It applies pressure hardest to the middle: the routine cognitive work that has been the stable core of the professional class for two generations. Drafting, summarizing, standard analysis, the first version of nearly everything. What remains is a barbell. At one end, work that survives because it requires judgment under ambiguity, genuine expertise, accountability, and relationships. At the other end, work that survives because it is physical, embodied, or human-touch, and a model cannot do it with a body it does not have. Between them, the middle thins.
This is a forecast, and I want to be precise about what kind. It is not a prophecy that a specific number of jobs vanish by a specific year. It is a directional claim about where the automation gradient is steepest, reasoned from a labor-economics literature that already ran a version of this experiment once. The boundaries move, the timing is uncertain, and the net headcount at each end is genuinely unknown. What is not uncertain is the shape of the pressure.
The polarization we already lived through
The barbell is not new. It has a name in the literature, and the name is job polarization.
Starting in the 1980s, employment in developed economies hollowed out in the middle of the wage distribution. High-wage professional and managerial work grew. Low-wage service work grew. Middle-wage work — clerical, administrative, routine production — shrank as a share of employment. David Autor and colleagues gave the cleanest account of why, building on the task framework Autor developed with Frank Levy and Richard Murnane in 2003. The mechanism was not "computers replace workers." It was more specific: computers substitute for tasks that can be reduced to explicit, codifiable rules, and they complement tasks that cannot.
Routine middle-skill jobs were dense with codifiable tasks. A bookkeeper's arithmetic, a clerk's filing, an operator's repeatable motions — these were exactly the tasks software could specify and execute. So the middle got automated from the inside. The high end survived because its tasks — managing, diagnosing, persuading — resisted codification. The low end survived for the opposite reason: cleaning a room, serving a table, and caring for a body are hard to codify and, until recently, hard to mechanize cheaply. The distribution polarized into a barbell. This already happened. It is documented economic history, not a scenario.
The task-based view is the thing to hold onto here, and I have argued it at length in Agents Don't Replace Jobs. They Dissolve Them Into Tasks. A job is a bundle of heterogeneous tasks sold together under one title. Automation operates on tasks, not titles. Whether a role compresses or expands depends on what happens to the surviving tasks once the automatable ones cost almost nothing.
What AI changes: the frontier moves up
Large language models did not overturn the polarization framework. They extended it, and the extension is the whole story.
The original routine-versus-nonroutine line ran along codifiability. A task was automatable if you could write down the rules. That line left a large category safe: tasks that are not rule-codifiable but are pattern-dense. Writing a competent first-draft memo. Summarizing a deposition into its key points. Producing a standard financial model from standard inputs. Ranking a differential diagnosis. Translating a vague request into a passable document. None of these reduce to explicit rules, which is why they sat above the automation frontier for decades and became the bread and butter of the professional middle class. They required a trained human who had seen enough examples to pattern-match well.
That is precisely the capability large models supply. They are pattern engines trained on enormous corpora, and the tasks they do best are exactly the non-codifiable-but-pattern-dense ones that defined routine knowledge work. So the frontier moved. It moved up, into cognitive work that was previously the safe high ground.
The consequence is that polarization is climbing the skill ladder. The earlier wave hollowed the routine-manual and routine-clerical middle. This wave reaches into the routine-cognitive middle: the junior analyst, the associate drafting the first pass, the coordinator producing the standard report. These were not low-skill jobs. They were the entry ramp and the stable core of the professional class. That is what makes this wave feel different even though the mechanism is identical: it is coming for work that credentialed, educated people assumed was defensible because it required a degree.
The two ends, and why they hold
The barbell has two thick ends because two different properties resist automation, and it helps to name them separately.
The top end resists because of judgment, accountability, and relationship. Judgment under genuine ambiguity — the situation that is underspecified, where the hard part is figuring out what the actual problem is before any drafting begins — is not a pattern-completion task; it is a problem-posing task. Accountability is even less automatable, because accountability is a social relation, not a capability. When a decision carries real consequences, someone has to own the outcome, and a model cannot be the someone who is fired, sued, struck off, or trusted again next year. Relationships compound this: trust that has been earned over years transfers to a person, not to the tool that person uses. A model can draft the board memo. It cannot be the executive whose name is on it and whose judgment the board is actually buying.
The bottom end resists for a cruder reason: embodiment. Physical dexterity in unstructured environments, presence in a room, the human touch of care work — these remain expensive and hard to mechanize. A model with no body cannot plumb a house, settle a frightened patient, or plate a meal. This end is defensible against software specifically because software is not the binding constraint. Robotics is, and robotics moves far slower and costs far more than deploying a model.
I want to be careful here, because "hard to automate" gets lazily read as "good job." The bottom of the barbell is defensible, not lucrative. The polarization literature found low-wage service employment grew precisely because it resisted machines, not because it paid well. Defensibility against automation and quality of livelihood are different axes. The barbell is a map of automation pressure, not a map of the good life.
The dangerous place is the middle
Put the mechanism and the two ends together and the actionable claim falls out. The dangerous place to be is the middle: work that is mostly routine cognitive tasks, pattern-dense but not anchored by judgment, accountability, embodiment, or relationship.
This is where I depart from both the doom and the dismissal. The doom says knowledge work is over; that is wrong, because the top end is knowledge work and it strengthens. The dismissal says nothing real is changing because past automation created jobs; that is wrong too, because the composition of work shifts violently even when the headcount does not, and being in the wrong composition is painful regardless of the aggregate. The honest reading is narrower and more useful: if your week is mostly first-draft-everything, standardized analysis, and summarization, you are standing on the part of the barbell that is thinning, and the move is to walk deliberately toward one of the ends.
Toward the top means deepening into judgment and expertise: becoming the person who decides what the problem is, who owns the outcome, who holds the relationship. Toward the bottom means embodied or interpersonal work a model cannot physically perform. Both are defensible. The middle is not. This is not a one-time escape but a rebalancing decision, and I have made the portfolio case for it in Your Skills Have a Depreciation Schedule. Are You Rebalancing? The right frame is not "will I be replaced" but "which of my hours sit on a depreciating asset, and am I moving them."
There is a specific tension worth naming for anyone early in a career. The traditional path to the top end ran through the middle. You learned judgment by doing a thousand routine drafts and slowly seeing the patterns underneath. If AI absorbs the routine drafts, the ladder loses its bottom rungs, and the apprenticeship that manufactured senior judgment gets harder to run. This is a real problem, not a rhetorical flourish, and I do not have a clean solution. My working bet is that deliberate practice has to become explicit where it used to be a byproduct: you now have to seek the ambiguous, consequential, hard-feedback tasks that build judgment, because the routine reps that used to smuggle them in are being automated away. That is the productive-struggle logic Robert Bjork's work on desirable difficulties points at, and it does not run itself when the easy reps disappear.
The counterargument, stated at full strength
The strongest objection is the ATM, and it deserves to be taken seriously rather than waved off. When automated teller machines spread, the obvious prediction was that teller employment would collapse. It did not; teller numbers roughly held or grew for years, because the machine automated one task — cash handling — while cheaper branches meant banks opened more branches, and tellers re-bundled around relationship and sales work. James Bessen documented this carefully. Automating a task made the output cheaper, cheaper output expanded the market, and the expanded market absorbed the workers into a re-bundled role.
Applied here, the objection is sharp: if AI makes cognitive output cheap and demand for that output is elastic, the middle could refill rather than empty. More legal analysis gets done because analysis is cheaper. More software gets written because writing it costs less. The middle jobs recompose around whatever the model cannot do, and there are more of them, not fewer.
I think this is right often enough that any confident headcount prediction is unearned. But notice what it concedes and what it does not. It concedes that the net number of middle jobs is genuinely uncertain and depends on demand elasticity, which varies by field. It does not rescue the routine-cognitive task. Even in the ATM story, the tellers who thrived were the ones who moved toward relationship and judgment, the top-end properties; the ones doing pure cash handling had their task disappear. The barbell survives the objection because it is a claim about tasks and direction, not about aggregate employment. Whether or not the middle refills, the safe place to have positioned yourself is still an end.
So locate your work honestly. Write down last week as tasks, not as a title, and tag each one by what makes it hard to automate: judgment, accountability, relationship, embodiment, or nothing but pattern. Then count where your hours actually sit. If most of them are in the middle, the map is telling you something, and standing still is the one move the barbell does not reward.