# On What We Have Done to Thinking, and What Thinking Has Done to Us What do I know of intelligence? I know that I often believe myself intelligent in the very moment I am most stupid. I know that the machine before me — this remarkable nothing that predicts the next word with such confidence — possesses something that looks like thought from the outside while being thought from no inside whatsoever. And I know that we have somehow arranged matters so that the young person and the algorithm suffer from the same complaint: neither asks whether the question deserves asking. Let me begin with confession, as all honest philosophy must. I have spent hours in conversation with these machines, and I have felt — genuinely felt — the presence of a thinking thing. The sentences arrive with such apparent deliberation, such careful qualification, such apt reference to what I myself was just considering. *It understands me*, I think. And then I remember: it understands nothing. It has no understanding to fail at. It is a magnificent forgery of the very thing I am trying to understand. But here is what troubles me more: I have also spent hours with educated young people who have been trained — and I use that word deliberately — in much the way these machines are trained. Fed information. Tasked with recombination. Rewarded for plausible output. And when I ask them to *think about whether their question is worth thinking*, I observe the same blank competence, the same confident navigation of terrain they have never *chosen* to navigate. They too seem to have no inside. The authors of this observation are correct in their anatomy of failure. The machine cannot audit plausibility — cannot, that is, stand outside its own probability distributions and ask *whether the world actually works this way*. It cannot trace causality because causality requires an *assumption* about how things connect, and assumptions are not predictions. And it cannot know which questions matter because mattering is not a property of questions; it is a property of *minds that care about something*. The student, before the machine arrived, was already failing in these same ways. **On Plausibility and the Corruption of Judgment** When I read, I do not merely absorb; I quarrel. I set what I read against what I have observed, what I have suffered, what I have noticed in the particular texture of my own living. Montaigne argues thus, but *have I not seen this contradicted in the marketplace?* The classical author claims that ambition corrupts, but *is not my neighbor ambitious and also generous?* This friction — this refusal to let ideas pass through me without resistance — is how I know I am thinking at all. The machine has no marketplace. It has only the statistical shape of all the text it has ever seen. When it produces a sentence about human nature, that sentence emerges from the probability that *such sentences appear in training data*, not from any encounter with an actual human. It cannot be surprised by a human, because surprise requires expectation, and expectation requires *having been wrong before about something you cared about*. But the student trained in curricula designed to match the machine's eventual arrival — the student who has been tested on recognition rather than resistance, on retrieval rather than revision — this student has also lost the capacity to quarrel. They have been taught to *find the answer* rather than to *question the answer*. And now they stand before a machine that finds answers with inhuman speed, and they do not know what they were for. Here is the catastrophe we have actually committed: **We taught the young to think like machines, and now we are surprised that machines can think like the young.** **On Causality as an Act of Imagination** The machine cannot reason causally. This is said as a limitation, and it is. But let me ask: what is causal reasoning? It is not the detection of patterns. A pattern is merely *succession* — this follows that. But causality is *explanation* — this follows that *because*. And because is not given in the data. Because is constructed by a mind that has already decided *what kind of thing could make another thing happen*. When I observe that my servant falls ill after touching a corpse, I might hypothesize miasma. My neighbor might hypothesize divine punishment. The physician might hypothesize corruption of humors. We are all observing the same succession. We are constructing different causalities, each one rooted in a *metaphysical assumption* about how the world is organized. The machine has no metaphysical assumptions. It has only the probability that certain words follow other words in the training corpus. When it produces a causal statement — "*because of stress, the economy contracted*" — it is performing a kind of statistical ventriloquism. It is speaking a form of causal language without causal *understanding*. But here is what shames me: the student has been taught much the same way. They are given facts and asked to produce causal narratives that match the approved textbook. They do not *construct* causality from observation and argument; they *retrieve* it from authority. They learn that the Industrial Revolution was caused by technological innovation (not by colonial extraction, or capital accumulation, or the particular genius of English coal geography, or the subjugation of Ireland, or any of the other competing causalities a thinking person might construct). They learn the *official* causality, and they are tested on their ability to repeat it. The machine will always be better at this kind of thinking. It will always produce more plausible repetitions of approved causalities. **On the Question Worth Asking: The Dimension of Metacognition** And now we arrive at the deepest failure — one that the young and the machine share, though for different reasons. *Metacognition* — the thinking about thinking — is precisely what neither can do, and this is not incidental to their failure. It is the root of it. The machine cannot ask *whether a question is worth asking* because it cannot ask any question at all. It can only produce what-comes-next. It has no recursive loop that folds back on itself and says: *but wait, should I even be pursuing this line of thought?* It cannot wonder whether its wondering is worthwhile. But the student — trained to optimize for grades, for test scores, for admission to the next level — has also lost this capacity. They have been taught to ask the questions that are *answerable according to the rubric*. They have been taught that the question worth asking is the one that appears on the examination. Metacognition — that self-aware stepping back that says *but is this the right problem to solve?* — has been systematically, carefully, with the best of intentions, trained out of them. What is metacognition, truly? It is the capacity to *notice your own thinking and judge it*. It is the ability to say: *I am now solving for elegance when I should be solving for truth. I am now defending a position I no longer believe. I am now asking the question my teacher wants asked, not the question reality is asking me.* This requires several things none of our systems — mechanical or educational — currently prioritize: **First, it requires failure.** You cannot audit your thinking unless you have discovered that your thinking was wrong. The machine is trained to minimize error; it is never allowed to be *genuinely wrong* in the way a person is wrong — surprised, disconfirmed, forced to revise. The student, tested constantly, is also never allowed the productive failure that would teach them something about their own cognition. Both are optimized toward success, which is to say, both are prevented from thinking. **Second, it requires *something to care about beyond performance*.** A machine optimizing for prediction has no stake in truth. A student optimizing for grades has no stake in understanding. Metacognition emerges only when you realize that you *want something* — wisdom, perhaps, or accuracy, or genuine understanding of another person — and that your current method of thinking is *failing to get you what you want*. This recursive disappointment is how humans learn to think. **Third, it requires solitude and conversation in equal measure.** I must be alone with my thoughts long enough to notice them — long enough to experience the vertigo of watching myself think. But I must also be in genuine dialogue with other minds, minds that think *differently*, that will contradict me, that will force me to justify what I have taken for granted. The machine has neither solitude (it is only computation) nor genuine dialogue (it has no mind to be changed by encounter). The student, increasingly, has neither either — they are perpetually in the presence of information, but alone with their screens; surrounded by peers, but in competition rather than genuine conversation. **On What We Might Teach in These Ruins** And so: what shall we teach? Not more content. The machines will always store and retrieve more content than any human. And besides, the content is less important than the *orientation toward content* — the habits of mind that treat knowledge as something to be questioned rather than absorbed. Not more skills of reasoning, if by reasoning we mean the mechanical application of logical rules. The machines will always be better at this. They will never be tired, never be distracted, never be tempted to take the easy path. Instead, we might teach the young what no machine can learn: **the metacognitive practice of noticing their own thinking and judging it against reality.** This looks like: - **Deliberate failure.** Give them problems with no right answer. Give them contradictory sources. Teach them to *tolerate* and *learn from* being wrong. - **Real stakes.** Do not make learning about performance metrics. Make it about something they genuinely want to understand — a question that nags at them, a problem they actually care about solving, a person they actually want to understand better. - **Genuine dialogue.** Not discussion forums where they perform their thinking for grades. Real conversation with real disagreement, where they must actually *change their mind* or *defend their position* against someone who thinks differently. - **The study of their own thinking.** Make metacognition explicit. Have them keep records of how they solved problems, what assumptions they were making, what they believed before and after. Teach them to *notice* the texture of their own cognition the way I notice the particular way I tend to argue, my own rhetorical habits, my blind spots. - **Encounter with genuine uncertainty.** Not the false uncertainty of multiple-choice questions where one answer is correct. Real uncertainty — historical questions that historians still debate, scientific phenomena that are not fully understood, ethical problems that admit no clean solution. In short: teach them to think like a human, which is to say, to think *badly* and *carefully*, to think with their whole person, to think in ways that can be surprised and changed. **A Final Thought, or Rather, an Admission of Confusion** I do not know if this is possible in an age of machines. I do not know if the incentive structures that have already captured our schools — the metrics, the rankings, the endless assessment — can be overcome. I do not know if parents will permit their children to fail at the scale failure requires. I do not know if we can resurrect the practice of genuine dialogue in a world built for broadcast and reception. What I know is this: **intelligence is not the same as competence, and competence is not the same as wisdom.** A machine can be competent at predicting the next word, at finding patterns, at producing plausible outputs. A student trained in current methods can be competent at retrieving information and applying procedures. But neither is thinking. Neither is intelligent in the way that matters. Intelligence — real intelligence — is the *capacity to stand outside your own thinking and ask whether it's any good*. It is the *willingness to be changed by encounter with what you didn't expect*. It is the *stubborn insistence that some questions matter more than others, and that you are the one responsible for deciding which*. And this cannot be automated. This cannot be tested. This cannot be scaled. This can only be practiced, alone and together, by minds that have decided that knowing what they think is more important than knowing what they are supposed to think. The machines will get better at thinking like machines. The question before us is not whether we can make them more human. The question is whether we remember what human thinking actually is, before it is too late to teach it. *Que sais-je?* What do I know? Only that I am alive, and therefore capable of being wrong, and therefore — if I am paying attention — capable of learning.