# Teaching How to Think: the forgotten imperative of the age of artificial intelligence

## I. The paradox of the moment

There is a quiet irony running through our offices and our classrooms. Never have we had at our disposal machines so capable of producing text, code, analyses and arguments, and never has it been so urgent to know how to think without them. Large language models write reports in seconds, summarise scientific literature, generate marketing strategies and debug software. They do, with disconcerting fluency, what for decades we considered the core of intellectual work. And yet the economic and human value of rigorous thinking has not diminished. It has increased.

The reason is simple to state and difficult to internalise: when the production of answers becomes cheap, what becomes expensive is the quality of the questions, the judgement about the answers and the capacity to decide what to do with them. A language model always returns something plausible. Distinguishing the plausible from the true, the adequate from the brilliant, the convenient from the correct, that remains human work. And it is precisely that work that our schools rarely teach and our companies rarely train.

This essay defends a thesis and proposes a method. The thesis: teaching how to think must become the explicit objective, and not the accidental by-product, of education and professional training. The method: there already exists a solid body of research and practice, from cognitive psychology to management, that tells us how to do it. We do not need to invent. We need to decide.

## II. The evidence of the problem: cognitive debt

Let us start with the data, because the alarm is not rhetorical.

In 2025, a team at the MIT Media Lab led by Nataliya Kosmyna published a study that became known by its provocative title, "Your Brain on ChatGPT". The researchers asked three groups of participants to write essays: one group with access to ChatGPT, another with access to search engines, another with only their own head. They measured brain activity by electroencephalography over several months. The results were unsettling: the group that used the language model showed the lowest neural connectivity, the lowest capacity to recall and quote their own text, and the weakest sense of authorship over what they had written. The authors coined an expression that deserves to enter common vocabulary: cognitive debt. Like technical debt in software, cognitive debt accumulates invisibly, saving effort today and charging interest tomorrow.

In the same year, researchers from Microsoft Research and Carnegie Mellon University presented at the CHI conference a survey of hundreds of knowledge workers on the use of generative AI. The central conclusion: the greater the confidence placed in the tool, the lower the reported critical thinking effort. Intellectual work shifts from production to verification, but many users skip the verification, especially in tasks they consider routine. Michael Gerlich, in a study published in the journal Societies, also in 2025, found a significant negative correlation between the frequency of AI tool use and performance in critical thinking tests, mediated precisely by the phenomenon of cognitive offloading, the habit of delegating to the machine what we used to do mentally.

And there is an even finer piece of data, from education. Hamsa Bastani and colleagues, from the Wharton School, conducted in 2024 an experiment with almost a thousand secondary school mathematics students in Turkey. The students who studied with access to an unrestricted ChatGPT improved during practice, but performed worse in the final exams taken without the tool, compared with those who never used it. The title of the paper says it all: "Generative AI Can Harm Learning". Curiously, the students who used a version of the model designed as a tutor, which gave hints instead of answers, did not suffer that harm. Let us hold on to this detail, because it contains the key to almost everything that follows.

The picture is, therefore, this: technology does not necessarily make us more stupid, but it makes stupidity comfortable. It removes the friction that was, after all, the gymnasium where thinking exercised itself. The answer cannot be to ban the tools, which would be as useless as banning the calculator. The answer is to deliberately teach what used to be learned through accidental friction.

## III. What thinking well actually means

Before the method, the definition. Thinking well is not having a high IQ nor accumulating encyclopaedic knowledge. It is, as John Dewey wrote in "How We Think" as early as 1910, the capacity to suspend judgement, examine evidence and conduct a disciplined investigation before concluding. Daniel Kahneman, in "Thinking, Fast and Slow", gave us the modern architecture of this idea: we have a fast, intuitive and economical system, and a slow, deliberate and costly one. The first governs most of life and fails in predictable ways. Thinking well is knowing when to distrust the automatic and activate the deliberate.

But there is a third piece, less known and perhaps more important. David Perkins and his colleagues at Project Zero, at Harvard, demonstrated that the difference between good and bad thinkers rarely lies in capacity and almost always lies in disposition. People fail not because they cannot think critically, but because it does not occur to them to do so at the right moment. Perkins calls it sensitivity to occasion: recognising that this situation, now, calls for scepticism, or calls for creativity, or calls for analysis of consequences. Thinking is less a muscle and more a repertoire with an alarm system.

This is where the idea of an explicit repertoire of thinking modes comes in. Charlie Munger, Warren Buffett's partner, spoke of a "latticework of mental models", a lattice of mental models drawn from several disciplines, without which experience does not turn into wisdom. There are several formulations of this repertoire; one of the most complete organises thirteen distinct modes, from divergent to integrative thinking. It is worth walking through them, because each one answers a different question that reality puts to us.

Divergent thinking generates possibilities when the problem is open: that is how Airbnb, in difficulty, decided not to cut costs but to hire professional photographers for all listings, doubling revenue. Convergent thinking chooses when it is time to decide: NASA, among twenty-four possible landing sites on Mars, did not choose the most exciting one but the one that maximised the rover's survival. Lateral thinking changes the framing: a hospital drastically reduced complaints about waiting times not by speeding up service, but by installing mirrors in the corridors. Systems thinking sees relationships and indirect effects: the reintroduction of wolves in Yellowstone changed the behaviour of the deer, recovered the riverside vegetation and even changed the course of rivers. Associative thinking imports solutions from other domains: a surgeon reorganised the handover of responsibilities in the operating theatre after observing the choreography of a Formula 1 pit crew.

Critical thinking questions accepted premises: Barry Marshall doubted that stress caused ulcers, deliberately ingested the bacterium Helicobacter pylori to prove it, and received the Nobel Prize. Analogical thinking transfers principles: Google's PageRank was born from the analogy with academic citations, in which a paper is worth the number and quality of those who cite it. Abductive thinking builds the best explanation from incomplete data, like Sherlock Holmes deducing Afghanistan from a tan and a military posture; it is the reasoning of medical diagnosis and of business decisions under uncertainty. First-principles thinking dismantles the problem down to fundamental truths: SpaceX discovered that the raw materials of a rocket were a tiny fraction of the cost and redesigned everything from there. Inverted thinking asks for the opposite: aviation became safe not by asking how to fly better, but by systematically eliminating the causes of accidents. Second-order thinking anticipates consequences of consequences, as in the famous cobra effect of colonial India, where the bounty per dead cobra led to the breeding of cobras. Probabilistic thinking replaces certainties with continuously revised estimates, the logic with which Amazon launches thousands of experiments knowing that most will fail and that one AWS pays for all the others. And integrative thinking refuses false dichotomies: the iPhone was born from the refusal to choose between power and simplicity.

Thirteen lenses, a single underlying competence: knowing which to use, when, and having the honesty to change lens when the first does not serve. This is what schools and companies need to teach. And it can be taught. Let us see how. (The thirteen lenses are developed, with examples, exercises and references, in the [interactive reference of the thirteen lenses](/resources/ensinar-a-pensar/modelos.html) that accompanies this essay, in Portuguese.)

## IV. How to teach thinking in schools

The first piece of good news is that educational research has already answered the hardest question: thinking is best taught explicitly and infused into content, not as an isolated subject nor as an implicit hope. Three approaches have robust evidence.

The first are the thinking routines of Harvard's Visible Thinking project, described by Ron Ritchhart, Mark Church and Karin Morrison in "Making Thinking Visible". A routine is a micro-structure of three or four steps repeated until it becomes habit. In the "See, Think, Wonder" routine, faced with an image, a chart or a document, the student first describes what they see, then what they think it means, then what they would like to know. It seems trivial; it is not. The routine separates observation from interpretation, which is exactly the confusion from which almost all reasoning errors are born. In the "Claim, Support, Question" routine, any statement must come accompanied by the evidence that sustains it and by a question that could bring it down. Practised weekly over a school year, this structure transforms the culture of a class: thinking stops being an event and becomes the environment.

The second approach is Philosophy for Children, created by Matthew Lipman in the seventies. The format is a community of inquiry: students read a stimulus, generate their own questions, choose one by vote and discuss it under rules of argumentation, with the teacher as facilitator and not as oracle. The independent evaluation conducted by Stephen Gorard for the British Education Endowment Foundation, in 2015, with more than three thousand students, found measurable gains in reading and mathematics, with larger effects among disadvantaged students, beyond the expected gains in reasoning. The meta-analysis by Trickey and Topping had already shown consistent effects a decade earlier. The mechanism is clear: when a child has to justify a position before peers who disagree, critical thinking stops being an exercise and becomes a social necessity.

The third is structured argumentation. Stephen Toulmin's model, formulated in "The Uses of Argument" in 1958, gives students a grammar of reasoning: a claim, the data that support it, the warrant that links the data to the claim, and the conditions under which the claim would fail. Teaching teenagers to decompose a newspaper editorial, or a viral post, into these components is probably the cheapest vaccine there is against disinformation.

The three approaches, with guided practices and ready-to-use session plans, are developed in this series' [guide for schools](/resources/ensinar-a-pensar/escolas.html), in Portuguese.

To these three approaches, the age of AI adds a golden rule that comes directly from the Bastani study: think first, consult after. The student solves, writes or sketches before opening the language model; the AI comes in as a Socratic tutor that gives hints, counter-argues and asks for justifications, never as a dispenser of answers. The experiment by Gregory Kestin and colleagues at Harvard, published in Scientific Reports in 2025, showed that a carefully designed AI tutor, with pedagogical scaffolding, outperformed even in-person active learning in physics learning gains. The tool is the same; the pedagogical design is everything. The school that bans ChatGPT and the school that lets it do the homework commit the same error in opposite directions.

## V. How to teach thinking in companies

In organisations, the problem changes scale but not nature. A company does not think; the people inside it think, and they think better or worse depending on the rituals the organisation institutes. The best companies in the world have already understood this and turned thinking modes into procedures. Four practices deserve to be copied.

The first is writing as thinking. In 2004, Jeff Bezos banned PowerPoint from Amazon's leadership meetings and replaced it with the six-page narrative memo, read in silence at the start of each meeting. The justification, repeated in the letters to shareholders, is a cognitive theory disguised as company policy: complete sentences force the explicit statement of the causal logic that bullet points hide. Whoever cannot write the argument does not have it. In an era in which AI writes for us, this ritual gains added value: the memo written by oneself, even if later polished with the machine's help, is the test that the thinking exists before the text.

The second is the pre-mortem, proposed by Gary Klein in the Harvard Business Review in 2007. Before launching a project, the team meets and assumes it has already failed catastrophically: each person writes down, individually, the reasons for the disaster. It is inverted thinking and second-order thinking turned into a thirty-minute meeting. The research by Deborah Mitchell and colleagues on "prospective hindsight" showed that imagining an outcome as certain increases by about thirty percent the capacity to identify its causes. The pre-mortem works because it gives social licence to scepticism: in that room, being the pessimist is the task, not the betrayal.

The third is the review after action, the After Action Review developed by the US Army and adopted by companies such as Shell. Four questions, always the same: what was supposed to happen, what actually happened, why the difference, what do we do differently next time. No culprits, with a written record. It is systems and critical thinking applied to experience, and it is the difference between an organisation that has twenty years of experience and one that has one year of experience repeated twenty times.

The fourth is the decision journal, recommended by Daniel Kahneman in several interviews as the cheapest tool for improving judgement. Before each important decision, the decider records what they decided, why, what they expect to happen and with what probability. Six or twelve months later, they re-read it. The journal fights hindsight bias, that mechanism by which we rewrite the past so as to have always been right, and it is the only way to train probabilistic thinking with real feedback: without a record of the estimates, no calibration is possible.

The four rituals, with fillable tools for the pre-mortem, the AAR and the decision journal, are in this series' [guide for organisations](/resources/ensinar-a-pensar/empresas.html), in Portuguese.

To these practices, the company of the AI age must add an explicit policy of use: the machine as sparring partner, not as oracle. Concretely, three simple rules. First, human draft before the prompt, in the decisions that matter. Second, always ask the AI for the counter-argument: "attack this proposal", "what assumptions am I making", "what would our best competitor say". Third, verification with a name: every AI output that circulates internally has a responsible human who validated it. These rules cost minutes and buy exactly what the Microsoft and MIT studies show to be evaporating: the moment of friction in which thinking happens.

## VI. A methodology in five steps

Synthesising the research and the practice, here is a methodology that any school or company can adopt in a semester.

First, make the repertoire explicit. Give names to the thinking modes, the thirteen described above or another equivalent taxonomy, and teach them with memorable examples. What has no name cannot be summoned. A team that knows the expression "second-order thinking" starts asking "and then what?" in meetings; one that does not know it, does not ask.

Second, institute routines, not events. An annual two-day training on critical thinking is theatre. A ten-minute routine practised weekly, "See, Think, Wonder" at school, pre-mortem at the company, is culture. Perkins's research on dispositions is clear: what is missing is not capacity, it is the habit of activation, and habits are built by repetition in context.

Third, write before consulting. In any cognitively relevant task, the first draft, the first estimate, the first hypothesis are human and are recorded. Only then does the machine come in. This sequence, validated by the contrast between the groups in the Bastani study, preserves the generative effort that learning requires, what Robert Bjork calls desirable difficulties.

Fourth, use AI in Socratic mode. Configure the assistants, in schools and in companies, to ask before answering, give hints before solutions, demand justification before validation. The difference between the AI that harms and the AI that teaches, the Wharton and Harvard studies show, is not in the model; it is in the design of the interaction.

Fifth, close the loop with feedback. Decision journals, after-action reviews, re-reading of predictions. Thinking only improves when confronted with its results. An organisation that decides a lot and reviews little is training confidence, not judgement.

The five steps, with a persistent checklist to track implementation, are in this series' [methodology guide](/resources/ensinar-a-pensar/metodologia.html), in Portuguese.

## VII. Conclusion: the advantage the machine does not replicate

Let us return to the initial paradox. Language models are here to stay and will keep improving. Fighting them is as sensible as fighting Gutenberg's press. But the history of technology teaches a consistent lesson: every tool that automates a competence increases the premium on the competence immediately above it. The calculator did not end mathematics; it ended mental calculation as a profession and made mathematical reasoning more valuable. The press did not end memory; it freed it for analysis. Language models will not end thinking; they will end writing as a proxy for thinking, and will make thinking proper, the formulation of problems, the judgement of evidence, the anticipation of consequences, the synthesis of opposites, the scarcest and best-paid asset in the economy.

The choice that schools and companies face is not, therefore, between adopting or rejecting AI. It is between forming people who think with the machine and people who let the machine think for them. The first path requires what this essay has described: an explicit repertoire of thinking modes, routines that turn them into habit, the rule of writing before consulting, tools designed to ask instead of answering, and feedback loops that confront judgement with reality. None of this is expensive. All of this is validated. All that is missing is the decision to do it.

The quality of our life, individual and collective, has always depended on the quality of our thinking. The difference is that, for the first time, we have a technology capable of sparing us the effort of thinking, and nothing reveals better the character of a person, a school or a company than what it does when effort becomes optional.

---

## References

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2024). *Generative AI Can Harm Learning*. The Wharton School Research Paper.

Bezos, J. (2004 and following). Amazon letters to shareholders; internal policy replacing PowerPoint with six-page narrative memos.

Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In *Metacognition: Knowing about Knowing*. MIT Press. (Concept of "desirable difficulties".)

Dewey, J. (1910). *How We Think*. D. C. Heath.

Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. *Societies*, 15(1), 6.

Gorard, S., Siddiqui, N. and See, B. H. (2015). *Philosophy for Children: Evaluation Report*. Education Endowment Foundation.

Kahneman, D. (2011). *Thinking, Fast and Slow*. Farrar, Straus and Giroux.

Kestin, G., Miller, K. et al. (2025). AI tutoring outperforms in-class active learning. *Scientific Reports*.

Klein, G. (2007). Performing a Project Premortem. *Harvard Business Review*, September 2007.

Kosmyna, N. et al. (2025). *Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task*. MIT Media Lab (preprint, arXiv).

Lee, H. P., Sarkar, A. et al. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. *Proceedings of CHI 2025*. Microsoft Research and Carnegie Mellon University.

Lipman, M. (2003). *Thinking in Education* (2nd ed.). Cambridge University Press.

Mitchell, D. J., Russo, J. E. and Pennington, N. (1989). Back to the future: Temporal perspective in the explanation of events. *Journal of Behavioral Decision Making*, 2(1).

Munger, C. (2005). *Poor Charlie's Almanack*. (Concept of the "latticework of mental models".)

Perkins, D., Jay, E. and Tishman, S. (1993). Beyond abilities: A dispositional theory of thinking. *Merrill-Palmer Quarterly*, 39(1).

Ritchhart, R., Church, M. and Morrison, K. (2011). *Making Thinking Visible*. Jossey-Bass. (Project Zero, Harvard University.)

Toulmin, S. (1958). *The Uses of Argument*. Cambridge University Press.

Trickey, S. and Topping, K. J. (2004). Philosophy for Children: A systematic review. *Research Papers in Education*, 19(3).

*The thirteen thinking modes (divergent, convergent, lateral, systems, associative, critical, analogical, abductive, first principles, inverted, second order, probabilistic, integrative) are developed in the interactive series [Pensar Melhor](/resources/ensinar-a-pensar/index.html) (in Portuguese), which accompanies this essay.*