The New Class Struggle
The comfortable story goes something like this. Automation eats jobs from the bottom up. First the farmhand, then the factory worker, then the cashier, and eventually, in some distant and frankly theoretical future, the people with degrees. The more education you have, the higher you sit on the cliff, and the longer it takes the water to reach you. Knowledge workers believed this for about seventy years, which is roughly how long it takes to build a mortgage, a pension, and a personality around it.
The story has a second, quieter clause, and it is the one that matters. It says that knowledge workers were paid for being clever. They were not. They were paid for being slow to manufacture. A senior tax attorney, a cardiologist, a software architect who has watched four platform migrations fail - these people commanded a premium not because intelligence was rare, but because their particular configuration of intelligence took twenty years to assemble, could not be copied, could not be seized, and expired when they did. The premium was never on the thinking. It was on the waiting.
This distinction sounds pedantic until you notice that generative AI did not make anyone less intelligent. It made a large portion of accumulated professional competence copyable. And the moment something becomes copyable, its price stops asking how hard it was to make and starts asking how hard it is to make again.
The Question
So here is the question the knowledge class has been avoiding at dinner parties. If the power of intellectual workers came from a delay - the decades it takes to turn a graduate into someone worth eight hundred dollars an hour - then what happens to that class when the delay collapses from twenty years to twenty seconds? Who inherits the power that used to sit inside their heads? And is the fight that follows a new kind of class struggle, or just the old one with better vocabulary?
The Explanation
The short version. For most of the twentieth century there were two kinds of leverage in an economy. You could own things, or you could withhold your labor. Capital had the first, labor the second, and the argument between them is what we politely call modern history. Knowledge workers were the strange third party at the table. They did not own the factory, but they owned something the factory could not replicate - expertise that lived in a skull, accumulated slowly, and walked out the door every evening. Economists gave this a flattering name, human capital, and the people who had it gave it an even more flattering one, merit.
Generative AI is, at bottom, a machine for making that capital alienable. It compresses the written record of professional work into a set of weights that can be copied for the cost of electricity. That does not eliminate the expert. It eliminates the scarcity of the median expert, which is where most of the premium was hiding. The value does not vanish. It migrates - to whoever owns the compressing machine, to whatever remains stubbornly uncopyable (accountability, presence, trust, a signature that can be sued), and to a political fight over who owns the means of cognition. That fight is the new class struggle. The rest of this piece follows one person through it, and pauses periodically to explain the machinery grinding away underneath her.
Two Plots
Let's call her Helen, because every firm has one. Helen is fifty-three, a tax partner at a mid-sized firm, and she has spent twenty-six years learning where the bodies are buried in a tax code longer than the collected works of most religions. Her value proposition is disarmingly simple. Clients bring her a plan, she looks at it for six minutes, and she says "don't do that". The six minutes are billed as a tenth of an hour, which is how the legal profession measures time, presumably because a full hour of Helen would be unaffordable to anyone who is not already a small country. Helen's time is the product. The meter is the business model. Nobody has ever asked her what she does with the other fifty-four minutes.
What Helen was selling, in the language economists use when they want to sound like engineers, is a rent on inelastic supply. Gary Becker's human capital theory, formalized in the 1960s, treats expertise as an investment with a long production time and no secondary market. You cannot buy a senior tax attorney. You can only grow one, and the growing takes two decades regardless of how much you are willing to pay. When supply cannot respond to price, price responds to demand alone, and whoever holds the fixed stock collects the difference. That difference is the skill premium, and from roughly 1980 onward it widened dramatically. Economists named the cause skill-biased technological change, which is a very polite way of saying that computers made educated people more valuable and everyone else less so. The knowledge class did not experience this as luck. It experienced it as having been right about everything.
In 2023, Helen's firm bought a subscription to an AI research tool, mostly so the managing partner could say the word "innovation" at a conference. The early results were a delight. A lawyer in New York had just been sanctioned for filing a brief full of cases the machine had simply invented, and Helen dined out on that story for a year. Then, sometime in 2025, the jokes stopped landing. The tool began producing first-draft memos that were not embarrassing. Then memos that were good. Then memos that were better than the ones her third-year associates produced, in eleven minutes rather than eleven hours, and without the associate's tendency to cry in the stairwell. The partners were thrilled. Helen was thrilled. Nobody noticed that the escalator behind them had quietly stopped moving.
Here is what the tool was doing, technically. A large language model is, among other things, a lossy compression of the written record of human expertise. Pretraining on trillions of tokens captures the statistical structure of how competent people write, reason, and argue in a given domain, and the part it captures best is the middle of the distribution - the standard memo, the typical analysis, the reasoning that appears ten thousand times in the corpus with minor variations. That middle is precisely where most professional labor lives. The tail, the novel argument nobody has made before, compresses badly because there is little to compress. What changed between 2023 and 2026 was not the basic idea but the plumbing around it. Longer context windows meant a model could read an entire case file rather than a paragraph. Retrieval grounded outputs in real documents instead of plausible fictions. Tool use and agentic scaffolding let the model check its own citations, run the numbers, and iterate on its own drafts. A team of academics who ran experiments on consultants at BCG described the result as a jagged frontier - inside it, the machine is superhuman across a startling range of tasks, and outside it, it fails in ways that look confident and stupid. The trouble for Helen is that the frontier moves, and it moves in one direction.
Meet Theo. Theo is twenty-five, has a law degree, a hundred and ninety thousand dollars in debt, and a LinkedIn profile that says "passionate about tax". He did everything right, which is historically the most dangerous thing a young person can do in the eighteen months before a technological shift. Helen's firm hired half as many associates as it did the year before, and the ones it hired are not doing the work Theo trained for, because the work Theo trained for is now a button. The profession's oldest bargain - you spend five years doing tedious research at a loss to the firm, and in exchange you slowly become Helen - has been quietly voided. The tedium was not a hazing ritual. It was the factory where Helens were made.
The empirical literature here is unusually consistent and unusually alarming for people like Helen. A 2023 study by Brynjolfsson, Li and Raymond tracked thousands of customer support agents given access to a generative assistant. Productivity rose about 14 percent on average, but the gain was concentrated among the least experienced workers, who improved by roughly a third, while the most experienced improved barely at all. Noy and Zhang found the same pattern in professional writing tasks - quality went up and the gap between weak and strong writers went down. The machine does not lift everyone equally. It lifts the novice toward the expert, which is another way of saying it compresses the very distribution the skill premium was pricing. And it creates a paradox nobody has solved. If entry-level cognitive work is where competence is manufactured, and that work is automated, the pipeline that produces senior experts shuts down. Either the models improve fast enough that senior humans are no longer needed, in which case Theo's problem is permanent, or they plateau at the jagged frontier, in which case somewhere around 2040 the world discovers it forgot to make any more Helens, and the few remaining ones become absurdly, comically expensive. Both futures are bad for Theo. Only one of them is bad for Helen.
The reckoning arrives, as most reckonings do, in the form of an invoice. A client asks Helen why the fee for a restructuring memo is the same as last year when, as everyone in the room knows, the memo took eleven minutes. Helen discovers, at fifty-three, that she has been selling time and that the market has finally noticed how little of it she uses. So she does what the consultants recommend and moves to "value pricing", a phrase that means charging for outcomes now that inputs have become embarrassing. And she discovers what her value actually is. It is not the analysis. The analysis is free. It is her signature at the bottom. She is the person who can be sued. Helen has become, professionally speaking, a liability sponge with a law degree, and the remarkable thing is that the market still pays handsomely for this, because the machine, for all its brilliance, cannot be disbarred.
This is where the class analysis earns its keep. In the classical account, class is defined by your relationship to the means of production - you own them, or you are hired by people who do. In 1977 Barbara and John Ehrenreich argued that the twentieth century had produced a third position, the professional-managerial class, whose distinguishing asset was inalienable human capital. You could not confiscate a doctor's training. You could not put a lawyer's judgment on a balance sheet and depreciate it. That inalienability was the entire source of the class's independence from both capital and labor. Generative AI makes the asset alienable. It extracts the statistical shadow of expertise from the texts experts wrote, frequently without asking, and encodes it into weights that someone owns. The cost structure of this new asset matters enormously. A frontier training run costs hundreds of millions to billions of dollars, while serving one more answer costs a fraction of a cent. High fixed cost plus near-zero marginal cost is the textbook recipe for natural oligopoly, and rent in such markets flows to whoever owns the fixed asset. The knowledge worker's inalienable capital has become somebody else's balance-sheet item. Meanwhile the human is repositioned as what the anthropologist Madeleine Elish called a moral crumple zone - the component in an automated system designed to absorb blame on impact. Helen's signature is not a residual privilege. It is a structural feature. Somebody has to be the crumple zone, and it turns out the market will pay the crumple zone rather well, right up until liability law is rewritten.
Meanwhile Helen's house needs rewiring, and the electrician quotes a number that makes her feel, for the first time in her adult life, poor. She watches him work and notices something the knowledge class spent forty years steering its children away from noticing. He cannot be downloaded. There is no corpus of a trillion tokens describing how to fish a wire through a wall built in 1961 by someone who hated the future. The plumber is fine. The paralegal is not. Helen's father was an electrician, she spent her twenties gently embarrassed by this, and she is now considering whether the embarrassment was, in the technical sense, a misallocation of capital.
The electrician's immunity has a name, and it is not "hard work". Hans Moravec observed in the 1980s that the things humans find hard - abstract reasoning, calculus, tax law - are computationally easy, while the things toddlers find easy - walking, grasping, reading a room - are computationally brutal. Language models feasted on text because text is abundant, cheap, and already digitized. There is no equivalent corpus of physical dexterity, and collecting one means building robots that fall over expensively. Layer on William Baumol's cost disease, which predicts that sectors resisting productivity gains rise in relative price as everything else gets cheaper, and you get the strange inversion of 2026 - the trades holding the premium the professions just lost. Two caveats belong here. Robotics is improving, and the immunity is a window, not a fortress. And the professions retain one moat the electrician lacks, which is the license. Competence can be copied. Permission cannot. The open question is how long a society tolerates paying rent on permission once everyone can see that the competence is free.
The Loom
Helen's firm gets a second invoice, and this one is addressed to it. The AI subscription, bought in 2023 as a conference anecdote, has become the thing the firm runs on, and at renewal the price per seat goes up by a number the vendor describes as "aligned with the value delivered". The managing partner asks IT whether they could switch. IT explains that every template, every workflow, every carefully tuned prompt library the associates built over eighteen months is written for this model's particular habits, and that the model it replaced has been retired and can no longer be bought at any price. The firm does not own the machine that produces its work product. It rents it, by the month and by the token, from a company whose founders nobody in the partnership can name, on terms that can change with thirty days' notice. Helen spent twenty-six years billing clients in tenths of an hour. The meter is still running. It is simply pointed at her now.
Marx was, for all his faults as a forecaster, precise about what a class is. It is not a level of income or a kind of manners. It is your relationship to the means of production - whether you own the machinery, or own nothing but your capacity to operate it and must therefore sell that capacity to whoever does. The knowledge class believed it had wriggled out of this arrangement because its means of production sat inside its own skull. You cannot foreclose on a cerebellum. Marx, oddly, saw the loophole closing. In an unfinished notebook from 1858, the so-called Fragment on Machines in the Grundrisse, he speculated that as industry advanced, society's accumulated knowledge - he used the English phrase general intellect - would be absorbed into fixed capital, into the machinery itself, and the worker would step to the side of the process as its watchman and regulator rather than its chief actor. He assumed this would happen to weavers and puddlers. It took a hundred and seventy years and happened to tax attorneys instead. A frontier model is the general intellect rendered as a depreciating asset. It is dead labor in the most literal sense Marx ever intended - the compressed residue of millions of hours of professional work by people who were not asked and are not being paid - and it sits on the balance sheet of someone who is not any of them.
How many someones. Fewer than ten organizations on earth can train a frontier model, and the number is falling rather than rising, because each generation costs several times the last. Those organizations buy their compute overwhelmingly from one chip designer, whose chips are fabricated by one foundry on one island, and run it in data centers owned by three cloud companies. Every layer of the stack is more concentrated than the one above it. This is not a conspiracy. It is the ordinary arithmetic of an industry with enormous fixed costs and marginal costs near zero, the same arithmetic that turned railways, telephones, and electricity into monopolies or regulated utilities within a generation of their invention. The difference is what is being centralized. In 1900 the means of industrial production were scattered across tens of thousands of firms. In 2026 the means of cognitive production are owned by a group of companies that could hold their annual meeting in a mid-sized restaurant, and whose stated ambition is to need a smaller table.
And they do not sell it. Nobody buys a frontier model the way Helen's father bought his van. You buy access, metered by the token, priced by the landlord, on a lease the landlord can rewrite. There is a name for this arrangement and it is older than Marx. In the East Midlands in 1811, the framework knitters who made stockings did not own their frames. They rented them from the hosiers, paid frame rent every week whether or not there was work that week, and the hosiers then bought the finished stockings at a price the hosiers set. When the knitters started breaking frames, the newspapers said they hated machines. They did not hate machines. They hated paying rent on the machine to the man who also set the price of what the machine made. Two centuries on, the knowledge worker subscribes to the means of cognitive production, per seat, per month, and the thing being rented was assembled from the knowledge worker's own output. It is a landlord who built the house from bricks lifted out of the tenant's garden and now charges for the view.
The obvious objection is that the rent is cheap and getting cheaper. This is true. The price of a token has collapsed faster than almost any input in economic history, and most of the landlords are still losing money, if you count the cost of building the thing being rented. Anyone who has watched a platform company for the last twenty years will recognize the shape. Standard Oil also lowered the price of kerosene. Rent in a natural oligopoly does not show up first in the price. It shows up in the dependence - in the prompt library that only works on one model, the retired version you can no longer buy, the renewal that arrives once the alternative has quietly stopped existing. The subsidy is not generosity. It is the down payment on a tenancy.
What this does to the class map is simple and unflattering. The Ehrenreichs' third class, the one whose independence rested on capital that could not be taken from it, has had its capital taken from it, copied, and rented back. Whoever owns nothing but a capacity to work and must pay for access to the machine in order to work at all has a name in the classical vocabulary, and it is not "professional". Theo lives here already. He cannot afford not to subscribe, his entire edge consists of knowing how to steer a machine he does not own, and the machine's next version may render that edge as obsolete as the research skills it replaced. He is a tenant farmer of cognition. The word for the process that put him there is proletarianization, and nobody at the dinner party uses it, because it is very hard to say while holding a glass of wine.
There is a counter-movement, and it matters, which is why the last of the four futures below is about plumbing rather than pay. Open-weight models are the cooperatively owned loom - a version of the general intellect that anyone can download, run, and refuse to return. As long as they stay within a year or so of the frontier, the landlords cannot charge whatever they like, and the question of who owns the means of cognitive production stays open. If they fall behind, it closes.
Four Futures
Nobody knows which of these arrives, and the interesting part is that the choice is largely political rather than technical. That is, after all, what makes it a class struggle rather than a forecast.
1. The Guild
Helen wins. The professions retreat behind licensing and liability, and expertise becomes something the machine produces and a credentialed human ratifies. Doctors sign, lawyers sign, engineers stamp, and the stamp is where the money lives. Helen's day consists of reviewing forty machine-generated memos and affixing her name to the ones she trusts, a job that is well paid and hollow in a way she does not discuss.
The mechanics are old. Guilds have always tried to convert technological threats into regulatory moats, and the strategy works right up until it does not. The stability of this scenario depends on how long the public will pay a permission premium on top of a competence that costs nothing, and history suggests the answer is "a while, and then very suddenly not". AI-native firms in permissive jurisdictions, insurers willing to price machine output directly, and a single well-publicized case in which the human signature added nothing but cost are all cracks in the wall. Guilds do not lose arguments. They lose relevance, and then they lose the argument.
2. The Barbell
The middle of the expertise distribution vanishes. At one end, a small elite of people with real taste, real judgment, and the ability to direct fleets of agents earn more than any knowledge worker in history, because their leverage is now measured in compute rather than in associates. At the other end, everyone who used to be the middle drifts toward verification work, checking machine output for accuracy at an hourly rate that would have made a 2015 paralegal weep. Theo is here. He reviews outputs. He is a human in the loop, and the loop does not remember his name.
This is Sherwin Rosen's economics of superstars applied to cognition. When a technology lets one person's judgment scale across millions of outputs, the returns concentrate on the very best, and the second-best becomes redundant rather than slightly less valuable. Mary Gray and Siddharth Suri documented the low end years ago under the name ghost work - the invisible human labor that keeps automated systems looking automated. This has happened before. "Computer" was a job title until the 1950s, held mostly by women with mathematics degrees, and the machines that replaced them kept the name as a courtesy. In this future knowledge work becomes what manufacturing became in the 1980s, a source of a few very good jobs, many very bad ones, and a great deal of nostalgia. The class struggle here is a rerun. The knowledge class simply discovers it was labor all along, with a nicer contract.
3. The Flood
Expertise becomes cheap, abundant, and everywhere. Helen earns a third of what she did, and her old clients are joined by millions of people who could never afford a tax attorney and now carry one in their pocket. Nurse practitioners do what doctors did. Paralegals do what lawyers did. Theo runs a small practice serving people who used to face the tax code alone. He is not rich. He is useful, which the previous generation of his profession would have considered a demotion.
The economist David Autor has argued this is the more likely outcome, and the more hopeful one - that AI's real power is to extend expert-level capability to workers without expert-level credentials, rebuilding a middle class out of augmented semi-professionals. The mechanism is the Jevons paradox, in which making a resource cheaper increases total consumption of it rather than shrinking the labor around it. Cheap coal meant more coal burned. Cheap expertise means more expertise consumed, more disputes resolved, more diagnoses made, more contracts actually read by someone. The premium dies and the profession grows. This scenario requires two things the current system resists - liability frameworks that let augmented workers act, and models reliable enough that letting them act does not end in the stairwell. It also requires the knowledge class to accept a smaller slice of a much larger pie, which is, historically, the thing it has been least good at.
4. The Fight Over the Pipe
Helen and Theo discover they are on the same side. The struggle stops being about hourly rates and becomes about ownership - of the weights, the compute, and above all the data the machines were trained on, which was, when anyone bothers to check, written by people like them. Professional associations start bargaining like unions. Unions start bargaining over training data instead of overtime. Someone proposes that a model trained on a century of legal writing owes something to the profession that wrote it, and for the first time in her life Helen finds herself agreeing with a picket line.
There is precedent. The 2023 Hollywood writers' strike was the first major labor action in which the central demand was not wages but control over how a machine could use the workers' output, and the writers won provisions that would have sounded like science fiction two years earlier. The economics that make this fight inevitable are the same ones that make the oligopoly likely. High fixed costs and near-zero marginal costs concentrate ownership unless something pushes back - open-weight models that commoditize the capital, public compute that treats inference like a utility, data licensing regimes that give training corpora a price. Each of these moves rent from the owners of the machine toward the people whose expertise it compressed. None of them happens by default. This is the scenario in which the phrase "class struggle" is not a metaphor.
The knowledge class never owned the means of production. It rented out a bottleneck in the human learning curve and called the rent merit, because that sounded better at dinner. The bottleneck is gone. The machine that replaced it is owned by a handful of companies and rented back, by the token, to the people whose work it was built from. What remains is a question about who owns the pipe, and that question will not be settled by a benchmark.
Helen, for the record, has started learning to weld. Theo found her the course. It was generated, it was excellent, and it costs nine dollars a month.