Socrates famously claimed to know nothing, and spent his life asking questions rather than providing answers. He was tried, convicted, and executed for it. ChatGPT claims to know nearly everything, and provides answers to virtually any question in seconds. It has been used by hundreds of millions of people. These facts are related. The question is how.
There is a scene in Plato's Meno that I have thought about more in the past two years than in the entire preceding decade.
Socrates has been arguing with Meno about whether virtue can be taught, and the argument has reached an impasse: they cannot agree on what virtue is, and without a definition they cannot determine whether something can be taught or not. To illustrate a point about the nature of knowledge, Socrates calls over one of Meno's enslaved attendants — a boy with no education in mathematics — and proceeds to lead him, through questions alone, to discover and correctly state the Pythagorean relationship for the diagonal of a square.
Socrates does not tell the boy anything. He asks questions. The boy makes mistakes, realises they are mistakes, corrects them, makes further mistakes, and eventually — through this process of being questioned, being wrong, and having to think — arrives at a correct answer. Socrates then makes his argument: the boy could not have learned this from Socrates, because Socrates taught him nothing. The knowledge, he claims, was already in the boy, and the questioning merely drew it out.
What struck me, returning to this scene recently, was not the theory of recollection that Socrates is trying to establish — that theory has not fared well over the centuries. What struck me was the structure of the interaction, and how completely it inverts the structure of every AI assistant I have used.
What Socrates Was Actually Doing
The method Socrates uses — called the elenchus, from the Greek for examination or cross-examination — is not a teaching method in any ordinary sense. It does not transmit information from teacher to student. It does not explain, demonstrate, or instruct. What it does is something more interesting and considerably more difficult: it makes the interlocutor's own thinking visible to the interlocutor, including the parts that are wrong.
The first move of the elenchus is always to ask for a definition. Not "what do you know about virtue?" but "what is virtue?" — what is its essential nature, the thing that makes all virtuous things virtuous? This question seems answerable. The interlocutor is usually confident they can answer it. They provide a definition. Socrates then asks questions about that definition, probing its implications, testing whether it holds in various cases, drawing out consequences the interlocutor had not anticipated. The definition fails. Another is offered. It fails too. By the end of a Socratic dialogue, the interlocutor knows considerably less than they did at the beginning — or rather, they know that they know considerably less than they believed.
This state — aporia, the Greek term for puzzlement or impasse — is not a failure of the dialogue. It is its purpose. Socrates describes himself, repeatedly, as a midwife: someone who helps others give birth to their own ideas rather than implanting ideas from outside. The midwife does not supply the child; the child was already there. What the midwife does is make the birth possible — which sometimes means the birth of an understanding that what was previously believed was wrong.
The interlocutor who leaves a Socratic dialogue in aporia is, Socrates argues, better off than they were before it: they have been disabused of false certainty, and false certainty is worse than acknowledged ignorance, because it closes off the investigation that acknowledged ignorance makes possible. The person who knows they do not know is, at least, in a position to find out. The person who believes they know, incorrectly, is not.
The Phaedrus and the Danger of Writing
In the Phaedrus, Plato records a myth that Socrates tells about the invention of writing. An Egyptian god named Theuth presents writing to the king of Egypt as a gift — a remedy (pharmakon) for forgetfulness, an aid to memory, a technology that will make people wiser and better-informed. The king rejects the gift. Writing, he argues, will do the opposite of what Theuth promises: it will produce not memory but its simulacrum, not wisdom but its appearance. People who have access to written records will believe they know things they have merely read; they will have the confidence of knowledge without its substance.
The king's argument is one of the strangest passages in the Platonic corpus, because it is an argument against writing made in writing — and Plato clearly knew this, because the dialogue is self-consciously structured around the irony. Socrates, who wrote nothing himself and was suspicious of the written word precisely for the reasons the king states, is being recorded and transmitted by Plato's text. The medium that Socrates distrusted is the medium through which his distrust has been preserved.
But the king's argument — and Socrates' endorsement of it — is not simply a technophobic reaction to a new medium. It is a specific claim about the relationship between information access and understanding. Writing, Socrates argues through the king, produces the appearance of knowledge without its substance. The reader who has consulted a written source believes they have acquired knowledge; what they have actually acquired is access to a record that can be consulted but cannot be interrogated. Writing cannot answer questions; it can only repeat what it says. It provides information without the understanding of information.
The Greek word Plato uses — pharmakon — means both remedy and poison, and the ambivalence is deliberate. Writing is a remedy for the fallibility of memory and a poison for the depth of understanding. It extends access while attenuating the cognitive engagement that genuine understanding requires. Both the remedy and the poison are real.
The Perfect Pharmakon
I want to apply this framework to ChatGPT, and to the category of large language models generally, with as little anachronism as possible. Socrates did not anticipate AI, and it would be absurd to pretend that the arguments in the Phaedrus are about something Plato could not have imagined. What is not absurd is to notice that the structure of the argument maps onto the current situation with unsettling precision.
If writing is a pharmakon — a remedy for forgetting that is simultaneously a poison for understanding — then a large language model is the pharmakon carried to its logical extreme. It is not merely a record that can be consulted; it is a record that responds to questions. It is not merely a text that says what it says; it is a text that can paraphrase, summarise, explain, and elaborate on demand. The limitation that Socrates identified in writing — that it can only repeat itself, cannot be interrogated, gives the same answer to every question — has been removed. The new pharmakon answers.
And this is precisely the problem that Socrates would have identified. The limitation he found in writing was not an accident; it was, in a specific sense, protective. A written text that could only repeat itself at least made the interrogation visible as something the reader had to supply. The reader who wanted to understand a text had to bring the questions; the text could not bring them. The gap between the text and the understanding of it was the gap that the reader had to cross, and crossing it was the process by which understanding — not information, but understanding — developed.
A language model that generates intelligent-sounding responses to every question closes this gap. Or rather, it papers over it with something that resembles the result of crossing it. The user who receives a well-articulated explanation from ChatGPT experiences something very similar to the experience of having understood something — the resolution of uncertainty, the sense of the pieces coming together, the feeling of comprehension. What they may not have experienced is the cognitive work that the feeling of comprehension is supposed to track. The explanation arrived complete, from outside. The understanding it produced — if it produced any — is a question that the experience of receiving the explanation does not settle.
Confidence Without Examination
The specific cognitive danger that Socrates identified in writing — the confidence of knowledge without its substance — is amplified by large language models in a way that writing itself was not.
A written text has properties that constrain the confidence it can generate. It is static; it cannot adjust to the reader's level of understanding or respond to the reader's specific confusions. It requires the reader to actively engage with it, to bring their own questions, to notice where they are confused and where they are following. These requirements do not guarantee understanding, but they create conditions in which the absence of understanding is at least detectable: the reader who does not understand a text will typically know they do not understand it, because the text makes demands they cannot meet.
A conversational AI minimises these constraints in ways that are, from the perspective of learning, problematic. It adjusts to the user's level — or appears to. It responds to specific confusions — or appears to. It provides exactly the information the user seemed to need, framed in exactly the way the user can follow. The experience of interacting with it is the experience of being understood and helped, which is one of the most satisfying educational experiences there is.
But the satisfaction of being understood is not the same as the satisfaction of understanding. The experience of receiving help calibrated to your confusion is not the same as the experience of resolving your confusion through your own effort. And the difference, as I have argued elsewhere in this context, is the difference between a performance enabled by help and a capacity developed through effort. The model can produce the former. The latter requires something the model cannot supply: the struggle that produces it.
I noticed this in myself during a period when I was using language models heavily for work that involved domains I was learning. I was getting things done — the outputs were competent, the questions got answered, the research moved forward. And then someone asked me, without my laptop, without access to anything, what I understood about the topic I had been working in for three months. The answer I gave was considerably thinner than the work I had produced. I had been using the model as a cognitive prosthetic — extending my reach into territory I had not actually entered — and the territory was still foreign when the prosthetic was absent.
What the Elenchus Required
What Socrates' method demanded of its interlocutors was something that no conversational AI currently demands, and that most are designed not to demand: the articulation and examination of your own beliefs before receiving any input from outside.
The elenchus begins with you. It begins with your definition, your claim, your position. And it proceeds by examining what you have said — not what Socrates says you should think, but what the implications of your own position are, where your own position contradicts itself, what your own claim commits you to. The examination is of the interlocutor's thinking, not of Socrates' thinking. The interlocutor is not a passive recipient of information; they are an active participant in the examination of their own mind.
This requires something painful and productive: the exposure of what you actually believe, including the parts that are confused, contradictory, or poorly supported. The Socratic interlocutor who ends in aporia has not been taught that they were wrong; they have been shown, through the examination of their own beliefs, that what they believed was more confused than they realised. This is a different kind of learning from the reception of information, and it produces a different kind of understanding.
A chatbot that answers questions does not examine the questioner. It accepts the question as given and responds to it. The beliefs embedded in the question — the assumptions, the confusions, the things taken for granted that might not be grantable — pass through unexamined. The user's thinking is not the subject of the interaction; their question is. And the question, received and answered, leaves the thinking that produced it intact.
The Technology Socrates Would Have Wanted
It is worth asking, seriously, what a Socratic technology would look like — what a tool designed on Socratic rather than Aristotelian principles would do. Not what it would produce, but what it would require of the person using it.
It would not answer questions. Or rather, it would not answer questions until the person asking had articulated what they thought the answer was, and why they thought it, and what they would need to believe for that answer to be correct. It would examine the question before responding to it — ask what the question assumes, what it is really asking, whether the question as formulated is the question the user actually wants answered.
It would make thinking visible before extending it. It would refuse to be a substitute for the thinking that the user is capable of doing, while offering to be a partner in the thinking the user is not yet capable of doing alone. It would be useful in proportion to the cognitive effort it required of the person using it — which is precisely the inverse of the design principle that governs most AI assistants, which are useful in proportion to the cognitive effort they eliminate.
This kind of tool would be less immediately satisfying than the tools that currently exist. It would frustrate users who want answers rather than questions. It would be harder to use, slower, and less efficient at producing the outputs that measure of productivity typically tracks. It would also — if the argument of this piece is right — be considerably more educational. Not because it is better designed, but because it is designed for a different purpose: not performance, but development.
What Socrates Would Actually Have Said
Socrates, I think, would not have been surprised by the popularity of ChatGPT. He lived in a world where the Sophists — teachers who, for a fee, would teach anyone to argue convincingly on any side of any question — were more popular than he was. The Sophists provided what people wanted: persuasive rhetoric, practical skills, the appearance of wisdom. Socrates provided what people did not want: uncomfortable questions, exposed ignorance, the discovery that what they thought they knew they did not know.
He would have recognized, in a large language model, the most sophisticated Sophist yet built: an entity of extraordinary rhetorical capacity, able to produce convincing arguments on any topic, calibrated to the level of the interlocutor, responsive to every question, never embarrassed, never uncertain, never at a loss. He would have recognised it as the pharmakon in its purest form — a remedy for ignorance that is simultaneously the most effective poison for genuine understanding yet devised.
And he would have asked it a question it cannot answer: not a question about facts, or about code, or about the implications of a text, but the question that the elenchus always eventually reaches: What do you actually know? Not what can you say — what do you know? What, if examined, would survive examination? What, if challenged, could you defend not by producing more text but by showing the understanding that the text is supposed to represent?
The model would produce an answer. It always does. And the answer would be, as Socrates would have immediately recognised, exactly the problem.
A Closing Provocation
The strange thing about the Meno — the thing that makes it more than a historical curiosity — is that it ends with an honest admission of failure. Socrates and Meno have not solved the problem they set out to solve. They have not established what virtue is, whether it can be taught, or how it is acquired. What they have established is that the question is more difficult than either of them assumed, and that the confident answers they brought to the dialogue were wrong.
The dialogue is valuable precisely because it ends in aporia — in genuine, productive uncertainty rather than false resolution. The reader who finishes the Meno knowing less than they knew when they started has, if they have followed honestly, learned something that a confident answer would not have produced: the specific shape of the problem, the specific confusions that need to be resolved, the specific thinking that needs to be done.
This is not what most people want from a tool. It is what most people need from an education. The gap between those two things is not a gap that technology has created; it is a gap that Socrates was worried about before writing existed. Technology has simply made it wider, and made it easier to stand in the gap without noticing that you are not on either side.
Aethel exists in the space that Socrates occupied and ChatGPT does not: the space of questions rather than answers, of examination rather than provision, of helping you think rather than thinking for you. This is not a more comfortable space to occupy. Socrates could tell you about that. But it is the space in which the thing that education is actually for has any possibility of occurring.