What Is Reasoning?

A cousin who drives a BMW reminds you to call Fred. A sound in the next room makes you turn your head. Each is a transition in thought, but neither one is yet reasoning. The transitions are causal — one mental event triggers another — but the second event wasn't arrived at because the first served as a reason for it.

Cues that trigger mental processing are a baseline, but are not the type of cognition that is considered "reasoning".

Adler, in the introduction to the Cambridge anthology on the subject, sharpens it into a working definition^[1]:

Reasoning starts when one thought doesn’t just trigger the next, but supports it.

Cue-triggered transitions — one thought producing the next, with no justification in play — make up most of mental life.

Reasoning is more specific: it happens when one thought gives a reason for the next, rather than simply causing it.

What Makes It Count

Four features have to hold together for a transition to count as reasoning.

First: it operates on beliefs aimed at truth.

Belief has a direction of fit — it aims to match the world, not the other way around^[2][3]. You can choose a more comfortable seat, but you can't simply choose to believe the more convenient claim. The contradiction in asserting "it's raining, but I don't believe it" — what philosophers call Moore's paradox — follows from this: asserting something expresses a belief, and belief answers to evidence rather than preference.

That is why a conclusion cannot be justified simply because it is useful.

It may be comforting, profitable, or socially convenient, but that does not make it true, and it does not give us a real reason to believe it.

Second: it involves representing the connection itself.

This is where reasoning separates from more basic forms of cognition. Consider a simple chain: "The roads are icy, so we should leave early." A person reasoning through this isn't merely producing the conclusion after hearing the premise. They are holding the relationship between premise and conclusion as something to be evaluated — asking whether this reason is sufficient, whether it actually supports that conclusion.

Cognitive scientists call this metarepresentation^[4][5]: the ability to represent a representation. To reason from a premise to a conclusion is to hold "this follows from that" as a structured object in its own right — one that can itself be questioned, weighed, or defended.

This same capacity helps us understand what other people believe, recognize what someone means beyond their literal words, and check whether our own thinking is well supported.

Reasoning is part of a broader set of human abilities that depend on this reflective layer.

Third: the conclusion has normative grip.

Once you decide the evidence is strong enough, you cannot simply ignore the conclusion. Refusing to accept what the reasoning supports is not just a preference; it is a kind of inconsistency^[6].

A well-supported conclusion has a force that mere plausibility does not.

Fourth: the reasoner is answerable.

A conclusion reached by reasoning is one you can be asked to defend. You can be asked for the reason. You can also be wrong about it. This accountability — not just that you arrived at a conclusion, but that you stand behind it and can be evaluated on how you got there — is part of what makes it reasoning rather than association.

A system that produces outputs that look like conclusions, but lacks belief, an understanding of the inference, any sense of obligation to the result, and anyone responsible for it, is doing something different.

It may still be useful, but it is not reasoning.

The Three Forms

Reasoning takes different forms, and those differences matter.

The first distinction is by aim.

Theoretical reasoning aims at settling what is true — at belief about the world. Practical reasoning aims at settling what to do — at intention and action. Practical reasoning bottoms out in desires and ends (Williams^[1]). Theoretical reasoning bottoms out in evidence.

The second distinction, within theoretical reasoning, is by inferential form.

Deduction moves from general premises to specific conclusions and is monotonic: adding more premises cannot turn a valid argument invalid. It is the only form that produces certainty, and it pays for that certainty by telling us nothing the premises didn't already entail (Stalnaker^[1]).

Induction moves from observed instances to general claims and is nonmonotonic: new evidence can defeat the conclusion. Hume's problem is permanent — every induction assumes a uniformity of nature that induction itself cannot establish without circularity^[7].

Abduction, or inference to the best explanation, starts with the evidence and looks for the hypothesis that best explains it. We judge that hypothesis by how simple it is, how well it fits with what we already know, and how much it explains^[8][9]. This is the kind of reasoning we use every day in diagnosis, history, debugging, and scientific discovery. It is also one of the hardest forms to formalize, because there is no fixed rule for what counts as the “best” explanation.

Real expert reasoning interleaves all three.

A physician abducts a candidate diagnosis from a patient's presentation, deduces which tests should confirm or rule it out, and induces from prior cases.

Knowing which kind of reasoning to use in a given situation is not itself one of the three forms. It is part of what expertise means.

Why Reasoning Exists

The evidence on individual reasoning is not especially flattering.

Subjects solve the abstract Wason selection task — a basic problem in logic — at rates of 10 to 25 percent. Asked to judge which is more probable, "Linda is a bank teller" or "Linda is a bank teller and an active feminist," the great majority pick the conjunction — a direct violation of the most elementary rule of probability^[10]. Beliefs persist long after their evidential basis has been discredited. The pattern is robust across education levels, monetary incentives, and expert populations.

If reasoning were designed to make individuals more accurate, it looks designed badly.

Mercier & Sperber propose a different account^[11][12][13]. Reasoning is not a general-purpose inference engine. The fast, automatic judgments that fill most of mental life — perception, intuition, pattern recognition — run underneath reasoning and don't require it. Reasoning proper is a narrower mechanism, and it evolved to serve a different job: the production and evaluation of reasons in communication.

The problem it solves is trust calibration. When someone makes a claim, you can't directly inspect the world they're describing. What you can inspect is their reasons — the arguments they offer.

A capacity for making arguments and evaluating the arguments of others helps a communicating species learn from one another without becoming defenseless against manipulation. On this view, reasoning functions as an epistemic immune system.

This reframing helps the evidence make sense.

People who fail the Wason task on their own solve it about 80 percent of the time when they can argue with others^[14]. Groups often outperform even their best individual member when they are able to engage in real adversarial exchange^[15]. Confirmation bias—the tendency to seek evidence that supports your own position—then looks less like a flaw and more like one side of the system working as intended.

You gather support for your view; others test it.

What looks like bias in isolation becomes part of the structure of collaborative inquiry.

Neuroscience points to the same conclusion. Prado et al. found that argumentative reasoning preferentially recruits the medial prefrontal cortex and the broader theory-of-mind network — the same brain substrate used to represent what other minds are doing^[5]. Abstract syllogisms in controlled lab conditions engage executive control instead.

Reasoning, in its natural setting—the evaluation of arguments made by real people in real situations—is fundamentally social.

What This Map Is For

Reasoning is a reflective, socially evolved practice governed by norms.

It works with beliefs aimed at truth. It requires understanding the connection between premises and conclusion, not just producing an answer. And its conclusions carry force: they are judgments, not preferences, and they come with accountability.

With this structure in place, we can ask the questions that guide the rest of this series more precisely, such as:

• What are large language models doing when they produce text that looks like reasoning?
• As machine-generated text becomes more fluent, what forms of cognitive labor remain uniquely human?
• Where is the structural limit of automation, and why?

In the rest of this section of the book, we will build on the ideas developed in this article to begin answering these questions and others.

---

References

1. Adler, J. E., & Rips, L. J. (Eds.). (2008). Reasoning: Studies of Human Inference and Its Foundations. Cambridge University Press.
2. Anscombe, G. E. M. (1957). Intention. Blackwell.
3. Searle, J. R. (1983). Intentionality: An Essay in the Philosophy of Mind. Cambridge University Press.
4. Sperber, D. (Ed.). (2000). Metarepresentations: A Multidisciplinary Perspective. Oxford University Press.
5. Prado, J., Léone, J., Epinat-Duclos, J., Trouche, E., & Mercier, H. (2020). The neural bases of argumentative reasoning. Brain and Language, 208, 104827.
6. Grice, H. P. (2001). Aspects of Reason. Oxford University Press.
7. Hume, D. (1748/1977). An Enquiry Concerning Human Understanding. Hackett.
8. Harman, G. (1965). The inference to the best explanation. Philosophical Review, 74(1), 88–95.
9. Lipton, P. (2004). Inference to the Best Explanation (2nd ed.). Routledge.
10. Tversky, A., & Kahneman, D. (1983). Extensional versus intuitive reasoning: The conjunction fallacy in probability judgment. Psychological Review, 90(4), 293–315.
11. Mercier, H., & Sperber, D. (2011). Why do humans reason? Arguments for an argumentative theory. Behavioral and Brain Sciences, 34(2), 57–74.
12. Mercier, H., & Sperber, D. (2017). The Enigma of Reason. Harvard University Press.
13. Mercier, H. (2025). Reasoning and argumentation. MIT Press Open Encyclopedia of Cognitive Science.
14. Moshman, D., & Geil, M. (1998). Collaborative reasoning: Evidence for collective rationality. Thinking and Reasoning, 4(3), 231–248.
15. Laughlin, P. R. (2011). Group Problem Solving. Princeton University Press.