Around 370 BCE, on a riverbank outside Athens, Plato set down a story about a king who turned down a gift.
The god Theuth comes to King Thamus carrying his inventions. He has worked out number and calculation, geometry and astronomy, even the games of draughts and dice. He saves his favourite for last. Writing, he tells the king, will make the Egyptians wiser and sharpen their memory. He calls it an elixir for memory and for wisdom.
Thamus says no.
The inventor, the king tells him, is the worst judge of what he has made. Writing will do the reverse of what Theuth promises. People who learn it will stop exercising memory and lean on marks scratched outside themselves, so they will possess the form of knowledge without the substance of it. Writing gives its users only the appearance of wisdom. They will take in a great deal and seem to know much, and for the most part they will know nothing.
That is the same fear we can read about today regarding AI slop and rotting student brains. The technology has changed but the fear has not.
Thamus made an assumption that memory was a fixed quantity. Pour the remembering into the scroll and there is less left in the head, the way water leaves a glass when you tip it into a jug. Knowing was a fixed pie, and the new tool would eat the human’s slice.
Psychologists have a name for this. In 2010 the cognitive scientist Daniel Meegan ran a set of experiments and called the result zero-sum bias, the habit of thinking one side’s gain must equal the other side’s loss, even when nothing forces that trade. Meegan’s students assumed a high grade for a classmate lowered their own odds of a high grade, in a course where any number of people could earn an A.
The pie was not fixed, but they felt it was anyway.
The fear of AI runs on the same two circuits. One track is about work. If the machine does the task, the wage that paid for the task comes off your plate. The other is the one Thamus saw. If the machine does the thinking, the thinking drains out of you. Both feel like a fixed pie being divided.
But it doesn’t divide the way we’d expect, and the places where we’re right are the places worth watching.
The AI revolution(s)
Before we start let’s get and separate the hype from the facts.
The field got its name in the summer of 1956, at a workshop at Dartmouth College where a handful of researchers proposed that every feature of intelligence could be described precisely enough for a machine to copy it. They thought a summer might crack it. Then Nobel laureate and AI pioneer Herbert A. Simon predicted in 1965 that machines would do any work a person could within 20 years. Then Marvin Minsky said in 1970 that human-level intelligence was 3 to 8 years out.
However, when the British mathematician James Lighthill handed the government a report in 1973, judging that the field had failed its promises, the UK funding collapsed and the first AI winter ran through the late 1970s.
A second boom followed, built on expert systems that encoded specialists’ rules into software, but a second winter arrived around 1987 when the market for the machines that ran them caved in.
70 years of “this changes everything” on a cycle.
The current boom is real, but knowing the history is the start of thinking clearly about it.
1. AI is coming to take my job
The first AI fear has an older and more concrete history in the fear of the loom.
When the power loom spread through English textile towns in the early 1800s, it did to weaving what every machine does to a craft. Fewer hands made more cloth. The weavers who smashed the frames, the Luddites, had spent their lives at a handloom and now watched a machine do their day’s work in an hour.
The Luddites were not wrong about what was happening to them. Handloom weavers lost their livelihoods, and most never recovered them. The machine did not hand them new jobs on the way out. The transition took a generation, and the people caught in the middle paid for it with years of poverty and broken trades.
But cloth became cheaper. People who had owned one coat now bought three. Demand for fabric climbed so fast that the mills needed more hands to spin, dye, cut, ship and sell, and the towns that built those mills swelled with workers doing jobs that had not existed when the handloom weavers were young. The lump of work was not fixed. It grew, changed shape and moved.
The economist David Frederick Schloss named this the lump of labour fallacy. He had met a dock worker who used a machine to stamp out metal washers and felt guilty about it, as if by working faster he were stealing the next man’s job. The lump of labour fallacy is the idea that there is a fixed lump of work in the world, so any task a machine takes is a task subtracted from people for good.
But the lump is not fixed. As goods get cheaper to make, people buy more of them and want ones they could not afford before, and the work moves rather than vanishing. It is zero-sum bias wearing an apron.
There is another version of the same idea. Make a task cheap and people often do more of it, not less. Economists call it the Jevons paradox. Lower the cost of an activity and total demand for it climb. When cash machines arrived in the 1970s and 80s, many assumed bank tellers would be made redundant. But since running a branch became far cheaper, banks instead opened many more of them, and the total number of tellers increased. Then the smartphone arrived, put the branch in every pocket, and made the teller genuinely redundant. The paradox has a limit. Cheaper does not mean safe forever.
So where does that leave AI and knowledge work today? AI does the work faster across a widening range of tasks like drafting a first version, summarising a thousand pages, pulling one relevant case from ten thousand, and writing code. A 2026 review of the empirical evidence found little sign of economy-wide job loss or falling wages despite the fast adoption. An IMF study of Denmark, tracking 25,000 workers across 7,000 firms, found that AI tools saved about 3% of work time.
3% sounds unimpressive, and right now it is, but the loom did not transform textile towns in its first year either. What matters more than the current saving is where the saving goes. The Danish study found that part of the 3% went straight back into checking the machine’s output, which is a sign of early-stage adoption where trust is still being established. As that trust grows and the tools improve, the productivity gains in knowledge work are likely to follow the same curve the loom traced in textiles. Modest and patchy at first, then steep enough to reshape the whole field.
The Luddites smashed their frames in 1811. By 1850 Britain was producing cloth for the world.
That last detail matters, because the gains are smaller and grubbier than the headlines on either side suggest. BetterUp and Stanford researchers named a new tax on the office in 2025, workslop, the low-effort AI-generated material a colleague then spends about 2 hours deciphering and fixing.
The machine produces faster, but someone downstream pays for the speed.
2. AI will do my thinking
The second fear is the one Thamus had, and it is the new front in this revolution, because earlier machines mostly took over muscle and routine. This one reaches for the part we thought was ours, and only ours: thinking.
Writing offloaded memory, as Thamus warned. The printing press offloaded copying. And in the 1970s, when the cheap pocket calculator reached classrooms, parents and teachers had Thamus’s worry again. The debate ran for 20 years. A child who never struggled through long division by hand would never grasp what division meant and trust the answer without the judgement to know when it was absurd. A mid-1970s survey found that 72 percent of teachers and ordinary people disapproved of 7th-graders using calculators at all.
In 2025 a team at the MIT Media Lab wired up 54 people writing essays, some with ChatGPT, some with a search engine, some with nothing but their own heads, and watched their brains on EEG over 4 months. The group that leaned on the model showed the weakest neural connectivity and the thinnest sense of owning what they had produced. Asked to quote a line from the essay they had handed in minutes earlier, most of the AI group could not. The researchers called it cognitive debt and described it as thinking on credit.
Read it as a warning against leaning on the tool too early, before you have done the cognitive work yourself, and not as proof for getting “brain rot” or dumber by using AI.
Where the gut is right
The comforting advice would tell you to relax because history always works out. The honest answer says it’s worse than it first looks.
Take the work fear again. So far, no mass unemployment. But the overall picture hides people. In 2025 three economists at Stanford went through payroll records for millions of American workers and found something the average rate concealed. Employment for workers aged 22 to 25 in the jobs most exposed to AI had fallen by 13% since late 2022, while older workers in the same jobs held steady or grew. The signal is early and still argued over, but if it holds, the reading is plain. The pie is growing, but the youngest slice is being cut away.
The reason for this cut is important to know about for anyone who works with their mind. A decade ago the economist David Autor described why some jobs resist machines:
“We know more than we can tell.”
A senior partner’s judgement of a deal, a nurse’s read of a room going wrong, both run on tacit knowledge, rules we follow without being able to write them down. You could not program what you could not state, so you could not automate it. It is the moat protecting skilled human work.
Generative AI drains the moat from the other side. It does not need the rules written down. It learns the tacit pattern by swallowing millions of examples of people doing the work. The faculty Thamus thought lived only inside a human, the judgement that could not be spoken, is what the new machines are trained to imitate from the outside.
The entry-level tasks AI eats first, drafting the memo and writing the boilerplate code, are the same tasks where a junior used to build the tacit judgement that makes a senior. Cut the bottom rung the machine cannot yet climb, and the apprenticeship goes with it. The junior years were where people became the seniors the machine still cannot replace.
We are automating the ladder, not just the floor.
The reflex hides a second mistake under the first. It assumes the slices hold still long enough to count, when the lesson of every earlier wave is that they do not. Benedict Evans points out that nobody scoring jobs for “smartphone exposure” in 2005 would have put taxi drivers near the top, yet cheap location data rewrote the work and left the $1m New York medallion all but worthless. The job that changed most was the one no model thought to flag.
Worse, the machine can leave your task untouched and still end your job, by dissolving the business that paid for it. The internet never changed what made a good journalist. What it changed was who paid the journalist, killing the classified-ad monopoly that had bankrolled the newsroom. The craft survived but the salary did not.
Your work may be untouched while the product your employer sells, the one that funds your desk, is what AI turns cheap or free.
Thamus was not wrong that writing cost us something. We lost the bards who could hold the whole Iliad in memory and recite it through a night. The skill you stop practising does fade, and that part of the worry has always been true. What is new is the recent habit of using the calculator’s happy ending to wave away every fresh concern. Calculators turned out fine, the argument goes, so AI will too. But the calculator automated a narrow band of arithmetic and left the reasoning to you. A model that writes the essay, builds the argument and reaches the conclusion is not working in that narrow band.
Back to Thamus
So go back to the river, and to the king who said no.
Thamus was wrong about the catastrophe. Writing did not hollow out the human mind. It gave us Plato, who reaches us only because someone wrote him down, including this very warning against writing, preserved in writing, for 2,400 years. It gave us the law, the notebook, the proof you can check line by line, the argument that survives its author. The mind did not shrink. It built on the new floor and reached higher.
And Thamus was right about the cost. We did trade something away. He saw the trade clearly when the salesman in front of him could see only the upside. That clear sight is worth keeping.
The question AI puts is not whether the pie shrinks. It mostly does not. The question is which struggles you choose to keep. Offload the typing and keep the thinking. Let the machine find the needle and keep, for yourself, the judgement of which haystack is worth searching. Hand it the calculation and hold on to the sense of when the answer is absurd. That sense is what nobody can sell back to you once you have let it fade.
In 1997 a machine beat the world chess champion and the obituaries for human chess were filed, but hess did not die. More people play it now than at any point in its 1,500 years. The machine did not shrink the game. It grew it, and changed the kind of greatness it rewards. A grandmaster today trains against an engine that would crush any human alive, and is sharper for it, the way Thamus is more available to you now than to any Egyptian who had to keep him in memory.
Leave feedback about this