Socrates distrusted one of the greatest technologies human beings ever invented: writing. He wrote nothing himself. What we know of him comes principally through others, especially Plato. In Phaedrus, Plato has Socrates tell the story of the Egyptian god Theuth, who proudly presents writing to King Thamus as an invention that will improve wisdom and memory.
The king sees a trap. Writing, he says, will encourage people to retrieve knowledge from external marks instead of exercising memory within themselves. They will acquire the appearance of wisdom rather than wisdom itself. People will know how to find and repeat things without necessarily understanding them.
It is a wonderful argument. It was also, judged by history, spectacularly wrong. Writing did not destroy human thought. It enormously expanded it. Without writing there would have been no accumulated mathematics, modern science, libraries, universities or complex technological civilisation. Plato could preserve the argument against writing only by writing it down.
But perhaps Socrates made the wrong prediction about the right problem. His concern was what we would now call cognitive offloading: transferring work previously performed by the human mind to an external device. Writing was too primitive an external device to replace thought. Information technology, and especially artificial intelligence, may be different.
We have already seen hints of the phenomenon. Consider GPS. Most of us no longer memorise routes or construct detailed mental maps when travelling somewhere unfamiliar. We follow the little blue line. Research published in Scientific Reports found that greater lifetime GPS use was associated with poorer spatial memory when people subsequently had to navigate without GPS. In a small longitudinal component, heavier GPS use was also associated with a steeper decline in hippocampal-dependent spatial memory. The researchers cautioned against overly strong causal conclusions, but their findings suggested that habitual GPS use can reduce the mental work involved in forming cognitive maps.
The machine does the navigating, and eventually we may become worse navigators. Calculators produced a milder version of the same phenomenon. Phone contact lists mean many people no longer remember telephone numbers. Search engines reduced the need to remember facts when the facts could be retrieved in seconds.
Artificial intelligence takes the principle much further because it does not merely store information. It performs cognitive tasks. It summarises, translates, writes, calculates, searches, explains and codes. It proposes arguments and counterarguments. Increasingly, it can look at the first stages of an intellectual problem and carry the work through to a finished product.
The early evidence deserves attention. A 2025 study by researchers from Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 examples of using generative AI at work. Greater confidence in AI was associated with less critical thinking, while greater confidence in one's own ability was associated with more. The researchers also found that AI changed where human cognitive effort occurred, shifting it towards verification, integration and supervision of machine output.
That does not prove that AI is making humanity stupid. Indeed, other research examining more than 128,000 human-LLM conversations found evidence that people sometimes use AI in genuinely educational ways. Cognitive engagement appeared in nearly a third of the user turns studied, although the researchers' deeper category of constructive engagement appeared in only 4.9 per cent. AI can therefore function either as intellectual scaffolding or intellectual substitution, depending considerably upon how it is used.
That distinction may become one of the important questions of the AI age. There is a world of difference between asking an AI to challenge an argument you have constructed and asking it to construct the argument so that you never have to think about it. There is a difference between using AI to teach yourself calculus and using AI to complete the calculus assignment. The first potentially strengthens the human. The second potentially replaces him.
And this suggests an AI takeover scenario considerably less dramatic than The Terminator. The machines need not become conscious. They need not hate us. They need not manufacture killer robots or decide that humanity is an inefficient use of planetary resources. They merely need to become indispensable.
Imagine several generations growing up with systems that write their documents, navigate their journeys, calculate their finances, diagnose their machinery, produce their software, answer their technical questions and increasingly make routine professional decisions. As AI improves, retaining the old skills begins to look economically irrational. Why spend years mastering something the machine performs instantly and better?
Then the human knowledge base begins to hollow out. There may still be engineers, doctors, programmers and technicians, but progressively more of their competence consists in operating and supervising systems whose underlying operations they could not reproduce themselves. Eventually civilisation could resemble a passenger sitting in the cockpit of an aircraft flown entirely by autopilot. Everything works beautifully until the screens go black.
That is where Socrates' ancient objection suddenly becomes more interesting. Writing externalised memory, but the written information remained inert. A book cannot read itself, interpret itself or perform a task for you. To use the knowledge in a library, human beings still had to learn how to think. AI externalises something closer to thinking itself.
And technological civilisation has another vulnerability: electricity. A severe geomagnetic storm is one obvious thought experiment. The famous Carrington Event of 1859 disrupted telegraph systems across considerable distances. A comparable or larger event striking today's vastly more electrically dependent civilisation would confront a world radically different from Carrington's.
Suppose the disruption were prolonged. GPS disappears. Communications become unreliable. Cloud computing vanishes. Electronic payment systems fail. AI assistants are gone. Some computer-controlled infrastructure becomes unavailable. What remains is not simply a technological question. It is a human-capability question.
Can people navigate without GPS? Repair machinery without online manuals? Calculate without software? Farm without electronically managed supply chains? Diagnose equipment from experience? Preserve food? Maintain water systems? Operate complex institutions using paper and memory?
Our grandparents possessed enormous quantities of practical knowledge that most modern people have never needed. That was progress because technology replaced much dangerous, exhausting and inefficient labour. AI may extend the bargain from physical labour to intellectual labour.
Socrates underestimated writing because writing preserved human thought without becoming a substitute thinker. AI presents a different proposition. For the first time, the external storehouse can answer back. The danger may therefore be less that artificial intelligence wakes up one morning and decides that humanity is obsolete. It is that humanity gradually decides the same thing about its own abilities.
And if one day the lights go out, we may discover just how much of ourselves we had quietly handed over to the machines.