AI and Human History: The Latest Step in a Very Long Story
Mat Siems · 7 min read · from Sapiens Plus Plus
It is easy to feel that artificial intelligence arrived from nowhere. One year the computer answered when asked the right way, and the next it was writing, planning and offering to book the flights. The speed makes it feel like a break in the story.
It is closer to the next chapter. Humans have been moving their abilities outside their bodies for a very long time: cutting, remembering, lifting, counting. Each move followed a similar pattern, and each brought a similar bill. Seeing the pattern does not make the present less serious. It does make it less strange, and it suggests which questions are worth asking.
Here is the short version, in five moves.
The edge that left the hand
Start with a stone flake small enough to sit in your palm. One face is smooth, the other scarred where earlier flakes were struck away, and the edge between them is thinner than any knife in your kitchen. The earliest tools of this kind that archaeologists agree on, the Oldowan tools from Ethiopia, are about 2.6 million years old.
Other animals use tools. Chimpanzees do, and so do crows, sea otters and capuchin monkeys. What matters here is what the flake did for its maker. A hand cannot slice hide or open a bone for marrow. A hand holding a flake can. The ability to cut had moved out of the body and into an object, where it could be made sharper, copied, passed on and left behind for someone else.
How much those early tools shaped diet and brain size is still argued over. The basic move is not in doubt, and it is the template for everything that follows.
Memory moves into clay
About five thousand years ago, in the city of Uruk in Mesopotamia, someone cupped a small pillow of damp clay in one hand and pressed marks into it with a cut reed. The oldest tablets of this kind date to around 3300 BCE. They are not poetry. They are counts: so many measures of barley, so many workers, a ration, a delivery, a debt.
Writing was invented independently more than once, in Mesopotamia, China and Mesoamerica. In each place a record could now outlive the person who made it and travel further than any voice. In Mesopotamia reading stayed a specialist skill, taught in scribal schools, so the people who could read the records had a certain power over the people who could not.
That is the second half of the pattern. When an ability leaves the body it becomes larger and longer lasting. It also becomes something that can be owned, withheld and controlled by someone else.
The same move is easy to observe today. Ask a room of adults to write down, from memory, the phone numbers of the three people they would most want to reach in an emergency, and watch the pens hover. This is a familiar scene, not a study, but most people recognise it.
Nothing has gone wrong with anybody's memory. The numbers moved out of heads and into a device that is better at keeping them, and the heads stopped bothering, exactly as they stopped memorising epics once writing arrived.
Muscle moves into engines
For most of history, work was done by muscle: human or animal, with some help from wind and water. In the early eighteenth century a new kind of machine began draining mines. It burned coal to make steam, and the steam moved a piston. Early engines wasted nearly all the heat they produced. Later changes, a separate condenser and rotary motion, made them practical for far more than pumping.
Stand in front of a preserved beam engine on an open day and the point is hard to miss. A cast iron wheel several metres across is turning, and no one is touching it.
Nobody in the hall is pushing anything.
The word horsepower is a relic of that moment. It began as a sales comparison: a way of telling a buyer how many horses the engine would replace.
Why industrialisation began in Britain is still debated. Historians argue for coal and wage levels, for institutions, and for capital drawn from empire and slavery. What is clear is that strength had become something you could buy, install and run all night, and that whoever owned the engine owned the strength.
Calculation moves into gears
The next ability to leave was arithmetic. Pascal built an adding machine in the 1640s. Leibniz followed with his stepped reckoner in the 1670s. The hard part, the advance over the abacus, was the carry: getting a machine to pass a one to the next column by itself.
By the second half of the nineteenth century, mechanical calculators were ordinary office equipment. A clerk in an insurance office would set a row of brass levers, turn a handle and read the result in a row of small windows. In 1890 the United States census was tabulated by punched-card machines designed by Hollerith, a line of business that leads to IBM.
The detail worth noticing is in the insurance office. The machine is working out a premium that will matter a great deal to somebody.
It has no idea what the numbers mean.
Historians tend to tell this story as one about management more than intelligence. Mechanical calculation was a way to organise large amounts of routine mental work, first with rooms of human computers and then with gears. Understanding stayed with the people. The procedure went into the machine.
From answering to acting
Everything since has been a version of the same move, made faster. Switches got smaller. Procedures became software. Networks joined the machines together, and a great deal of memory, navigation and attention moved into a device in the pocket.
What is new in the present decade is machines that produce language, and then machines that act. An agent is a system that senses, decides and acts in a loop, which is an old engineering pattern, now driven by a language model. A sensible setup still looks like a robotics bench: the machine shows its plan, a person reads it and approves it, and there is a stop button within reach.
The honest picture has caveats. Small errors at each step compound over a long task. Tool-using systems can be attacked by instructions hidden in the material they read. Any claim about how well agents perform depends on the test, the setting and the date. Whether bigger models alone will close the reliability gap is debated. Where legal responsibility falls when an agent acts is unsettled and varies by country.
The costs follow the old pattern too. Research on learning has long shown that effortful recall strengthens memory, and controlled studies find weaker route learning in people who follow turn-by-turn navigation. Aviation and medicine have documented skills fading when automation is trusted too readily. Other popular claims are shakier: the idea that search engines erode memory came from a 2011 finding that did not replicate well in a larger 2018 project.
What the long view is good for
A history like this is not a forecast. It does supply a few steady points.
- The pattern is old. Cutting, remembering, lifting and counting all left the body before thinking did. AI is the newest move, not the first.
- Each move enlarges the ability. An ability outside the body can be scaled, kept and shared.
- Each move changes who holds it. What can be shared can also be owned, sold and switched off by someone else.
- The practice inside the person fades. This is sometimes fine, as with phone numbers, and sometimes not, as with a pilot's manual flying.
- Responsibility does not move with the ability. The machine in the insurance office did not answer for the premium. Someone did.
The last point is the one that matters most now. We already live with extensions and rarely think about them. Spectacles appeared in northern Italy in the late thirteenth century, and anyone who wears them forgets they are there within a day or two. The useful question about any new extension is who can use it, who can withdraw it, and who answers for what it does.
That question runs through Sapiens++: From Stardust to the Intelligence We Build, my illustrated history of this pattern in a hundred chapters, from the first atoms to the machines we now build. Every chapter ends with an Evidence box that separates what is established from what is debated and what is still open. It is an original, independent book. It is £9 on Gumroad with a 30-day refund: https://matsiems.com/books/sapiens-plus-plus. The first ten chapters are free to read at matsiems.com/books/sapiens-plus-plus.
Running several projects with AI agents? The Unbothered Machine is the operating manual for a company of one.