https://humanistreview.ai/issue-1/acemoglu-ai-replace-workers/
In the essay: “Will AI replace workers? Not if we build it right”, MIT professor Daron Acemoglu argues that AI is chasing the wrong goal. Instead of replacing workers, it should make them better at their jobs.
Daron Acemoglu argues that AI is chasing the wrong goal. Instead of replacing workers, it should make them better at their jobs.
Acemoglu writes a compelling post on how we might be training AI to make humans better at their jobs. And he is right that what we can do to help humans be better at their jobs is a prime requirement going into this huge change.
This relates to my longstanding theory that it is humans using AI agents well who will take jobs and not AI agents alone. Frankly there is not much point of having AI agents without humans being here to benefit from the technology.
So the agreement is real, and it is worth naming plainly before anything is asked of it. A Nobel laureate economist has stood up in public and said that the industry’s dominant ambition - machines that mimic and then replace the human - is the wrong ambition. He has said that the two intelligences are different in kind, and that when two things are different, the natural way to proceed is to complement one with the other. He has said that on the current trajectory we may squander the technology entirely.
There may be some extremist science fiction enthusiasts who see a planet run by AI agents with humans being either subservient to AI or sent into extinction. Acemoglu is not one of them, and neither am I. The disagreement this essay takes up lives somewhere else - not in the destination, but in the road he draws toward it, and in what that road quietly assumes has already been built.
I. Human-AI Partnership, With A Missing Partner
Here is Acemoglu on the partnership, in the paragraph that comes closest to my own position:
“It would have been a colossal mistake for Jobs to try to replace Wozniak’s engineering skills or Wozniak to ignore Jobs’s genius in design and marketing. They succeeded because they collaborated, combining their different skills in order to complement each other. This is what humans and AI can and should do as well.”
Take the sentence; it is correct as far as it goes. Two different intelligences, each incomplete alone, each making the other more than it was. This is my own division of labor, written by an economist: humans bring the meaning, the machine brings the compute.
Now the question: what does the Jobs-and-Wozniak story actually require?
It requires two partners who are both good at their half. Wozniak had taught himself engineering since grade school; Jobs had spent years sharpening a judgment no one could have delivered to him as a curriculum. The partnership worked because each man arrived already formed — and the forming was the hard part, the decades part.
Acemoglu’s paragraph describes the collaboration and skips the formation entirely. It is a blueprint for a meeting between two educated partners, and the essay never says how the human one gets educated.
He includes important background in this essay. "Writing about J. C. R. Licklider, who imagined man-computer symbiosis in 1960, Acemoglu - an economist - quotes him:"
“The hope is that, in not too many years, human brains and computing machines will be coupled together very tightly, and that the resulting partnership will think as no human brain has ever thought and process data in a way not approached by the information-handling machines we know today.”
From Licklider we take the vision, and it is a real one - the partnership was worth imagining in 1960 and it is worth building now. What we refuse is the assumption riding silently inside it: Licklider imagined the coupling; he did not ask who trains the human half of it. Sixty-five years later, the same question is still unasked - this time by the most cited economist writing on AI and labor.
II. The Answer Key Remains In Lockup
Acemoglu’s own example of pro-worker AI is education, so his essay answers my next question before I ask it - and the answer is the problem. Here is the passage:
“Rather than seeking to eliminate teachers, AI tools could identify which part of curricular material different students are having problems with, and then make recommendations for reorganizing the classroom into smaller groups where the original teacher, perhaps with the support of additional educators, can target the weaknesses and strengths of each subgroup.”
Ask it now: what is this picture, really?
It is the factory classroom with better instrumentation. The rows are rearranged, the diagnosis is faster, the teacher’s aim is truer — and every one of those improvements operates inside a structure in which the material exists somewhere upstream, already organized, already answered, and the student’s job is still to receive it. Nothing in the passage describes a student learning to think: to hold a question open, to test an answer against evidence, to discover that the answer key itself can be wrong. The tool tells the teacher where the student is weak. It never asks whether the student has practiced being strong.
This is the seam between his program and mine, and it is exact. There is a word in circulation — democratization — that Acemoglu’s program quietly relies on, and the word has two meanings that are not the same. One meaning: more people get the delivered product, delivered more efficiently. The other: more people get the practice — the doing of the thing, thinking itself, with all the friction and failure that the doing carries. His pro-worker AI delivers curriculum to smaller groups at lower cost. That is access to the product. The practice of thinking is a different access entirely, and no version of it can be delivered, because the delivery is precisely what it replaces.
III. The Worker and the Citizen
Acemoglu’s case for pro-worker AI is that it protects people — and he is more honest than his industry about the stakes. Here he is on what the automation path does:
“...pro-worker AI leverages and amplifies human capabilities,” while automation “displaces workers, leads to job losses and creates inequalities (and via these processes, endangers not just societal cohesion but a democratic future).”
Ask: a democratic future is endangered — by what, exactly, and defended by whom?
His answer, throughout the essay, is economic: inequality, displacement, concentration. Those are real. But a democracy is not endangered by inequality in the way a bridge is endangered by a crack — it is endangered when the people in it can no longer hold the verdicts. A citizen who has spent twelve years in school being handed answers and trained to select them on demand is a worker who can be complemented, upgraded, deployed — and still never once asked to judge. The machine can complement that person’s labor all day and it will not make them a citizen, because citizenship is not a skill shortage. It is a practice, and the practice has a name older than the republic that depends on it.
Acemoglu’s unit of concern is the worker. Mine is the citizen. The worker asks: will I keep my job? The citizen asks: who keeps the verdict? A society can answer the first question well and lose everything on the second, and the loss would not show up in his productivity statistics for a generation — by which time the verdict will have been offloaded so long that no one remembers holding it.
There is an American particular here that his framework has no shelf for. The founders built the republic’s checks and balances with extraordinary care — and enshrined none of the education floor that makes those checks operable. They argued for it; they did not build it in.1
Two and a half centuries later, the omission is the load-bearing absence: the machinery of self-government runs on educated judgment, and the judgment was never made the structure’s foundation. Pro-worker AI - or any technology - cannot repair an omission that was constitutional before it was technological.
IV. Pulling an Industrial Lever, Forgetting Human Meaning
Acemoglu closes his essay by turning from policy to perspective, and the turn is the most interesting thing in it:
“The most important tool for steering AI onto a better path is not a silver bullet policy: there is none. It is a change of perspective.”
And at the very end:
“Only human ingenuity, open-mindedness and our ability to simulate different worlds and act on them can put AI - and our entire democratic system - onto a better path.”
Ask the closing question, because the essay cannot: where does a human being get ingenuity, open-mindedness, and the ability to simulate different worlds - and act on them?
The essay spends thousands of words on the steering mechanism - symmetric taxation, demonstration subsidies, antitrust, a digital-ad tax, data markets in which experts own and sell their expertise. These are serious proposals from a serious economist, and several of them are worth enacting. But look at what they all are: adjustments to the incentives of institutions. Every one of them assumes a population already capable of the “change of perspective” he calls the true requirement — and the population’s capacity is exactly the thing no tax code touches. His lever is policy. The floor under every lever, including his, is education - and specifically not the education we have.
The factory-style, Prussian derived, sort of education we are accustomed to will simply not prepare humans to deal with Superior Intelligence. It was not designed to. It was designed in the nineteenth century to produce disciplined, interchangeable workers for an industrial state - and it has been producing them, with decreasing market value, ever since. The students it graduated are now the adults being told that a new intelligence has arrived, and that their options are to be complemented by it or replaced by it. Nobody in that arrangement was ever taught to think, so nobody in it can meet the new partner as an equal. The partnership Acemoglu wants - the Jobs-and-Wozniak desk - has one side permanently unstaffed, and his policy tools are all aimed at the furniture.
It is doubtful however that different training of agents will help make humans better at much of anything. It is humans themselves who either will or will not step up to engage in a worthwhile relationship with AI.
There is only one foundational thing which will help humans step up to use AI responsibly and articulately - and Acemoglu mentions it only as curriculum: delivery, diagnosis, the answer key, never the practice of thinking. That one thing is education - and specifically a certain type of education.
Socratic education has long been out of favor in modern society, but it remains the sole method of training brain power to educate a humanity capable of rising to these technological challenges.
Socratic education, which trains human beings how to think, combined with use of excellent subject matter, would prepare a population of trained brain power, fully capable of interacting positively with Superior Intelligence.
V. How the Human Meaning was Lost — the Historical Fracture
It is worth pausing on how the practice of thinking fell out of favor, because the fall was not an accident of neglect. It was a sequence, and every step of the sequence was chosen.
Socrates practiced a dynamic, destabilizing dialectic interrogation that rejected static dogma, leading to his execution by a state wary of independent citizens. Afterward, his methods were codified and systematized into rigid taxonomies by Aristotle and later institutionalized, culminating in modern standardized schooling and reinforcement-trained AI systems designed to neutralize divergence.
Read that sequence alongside Acemoglu’s four causes of the automation fixation - the business models, the concentration, the coordination failure, the AGI ideology descending from Turing’s imitation game. His fourth cause is the oldest: the dream of the machine that thinks like us. But there is a fifth cause his list does not reach, and it runs in the other direction: the long, patient institutional project of training humans not to think. The same century that standardized the classroom built the assembly line, and the two designs shared an author’s logic - reduce the human contribution to a repeatable, testable, supervisable output. The machine that mimics us was never the opposite of that project. It is its completion. An education that trains answer-selection produces the perfect complementarity partner for an answer engine: two answer-selectors, facing each other, neither able to ask a question that could change either of them.
And the psychology of the arrangement deserves its own accounting, because the vulnerability is not merely institutional. Modern conditioning fosters an outer-directed reliance on external validation, leaving minds vulnerable to cognitive ease. Socratic education introduces deliberate epistemic friction, training raw brain power to inhabit cognitive tension, navigate confusion, and maintain neurological sovereignty when engaging with complex systems.
Friction is the operative word. Everything in the modern interface is engineered to remove it - the autocomplete, the summary, the confident first answer. Acemoglu worries that AI will make workers worse by replacing them. The quieter worry is that it will make minds worse by pleasing them: a population that never sits inside a question it cannot resolve in one exchange, being handed resolution in every exchange. The answer engine and the factory school converge on the same product - a mind optimized for comfort with supplied answers. The Socratic method is the standing refusal of that product, and it is uncomfortable by design, because the discomfort is where the thinking happens.
VI. Two Pathways Toward Meeting the Machine
So what does the educated half of the desk actually look like, on the human side? Not a curriculum. Not a certification. Two ways of meeting the machine, both of them hard work, both available now, to anyone, at no cost but attention:
The Dissolution - inhabiting a conceptual loop with the model without yielding to its automated normalization - staying with the question until the convenient answer dissolves and the real one stands visible.
The Topological Transection - using the machine’s orthodox pushback as a sharpening stone - slicing cleanly through the settled answer to reveal the structure it was smoothing over.
Neither treats the machine as an authority. Both treat it as a calculation tool that returns exactly the thinking brought to it. Acemoglu’s design-and-engineering complementarity presumes two educated partners; this is what the educated half looks like on the human side of the desk.
Notice that neither practice requires a subsidy, a demonstration program, or an act of Congress. That is not an argument against his policy tools - several are worth having. It is an argument about sequence and about human meaning: the policies can only amplify a population that can already do this. Build the practice first, and every lever he describes gets a hand that can pull it.
VII. When the Man Who Trains the Machine Walks Out
This essay asks who trains the human half of the partnership. This week the question was answered from the other half of the desk - by a man who spent three years training the machine half, and then walked out.
Jacob Coxon is twenty-eight years old. He spent three years in pre-training research at OpenAI and Anthropic - the rooms where models are made capable. On September 8 he resigned in public, and his resignation statement ricocheted around the internet: more than ninety million views in under a day. Monday night he sat across from Jon Stewart and gave the accounting. Two things moved him from curiosi…
