How To Build A Better Mental Model

A mental model is an internal representation of how something works well enough to predict what happens next, not just recall what happened last time. The idea has a specific origin: Scottish psychologist Kenneth Craik proposed in his 1943 book The Nature of Explanation that the mind builds small-scale models of external reality and manipulates them to anticipate events before they occur. Everything that follows in this piece, from why facts alone fail you to how investor Charlie Munger turned the idea into a business philosophy, traces back to that one claim: understanding is simulation, not storage.

What a mental model actually is, and why Craik's 1943 idea still matters

Craik died in a bicycle accident in 1945, only two years after publishing the book that made his reputation, and he never got to see the field he helped found. But The Nature of Explanation laid out a claim that cognitive science has spent the following eight decades confirming in increasingly precise ways: the brain does not store the world, it builds working replicas of it. Craik described thought as a kind of internal machine that could "model or parallel" external events, letting an organism try out actions in its head before committing to them in the world. A rat that has built a spatial model of a maze can find a shortcut it has never physically walked; a rat that has only memorized one correct path cannot.

That distinction, between a stored path and a working model of the space, is the entire subject of this article. Mental model — an internal, simplified representation of a system's structure or behavior, used to explain the past and predict the future, rather than a memorized fact about a single instance of that system. The definition sounds abstract until you notice how often it explains ordinary failures of judgment: the manager who can recite "communication is important" but can't diagnose why a specific team meeting fell apart, or the investor who knows "diversify your portfolio" as a slogan but can't explain why concentration risk actually destroys wealth in a downturn.

Decades after Craik, British cognitive scientist Philip Johnson-Laird extended the idea into a formal theory of reasoning itself. In his 1983 book Mental Models: Towards a Cognitive Science of Language, Inference, and Consciousness, Johnson-Laird argued that people do not reason by applying formal logical rules to abstract propositions. Instead, they construct a working mental simulation of the situation being described and check what happens inside that simulation. When you read "the cup is on the table, and the book is on the cup," you don't manipulate symbols, you build a small scene in your head and read the answer off it. Johnson-Laird's theory has become one of the more influential frameworks in cognitive science precisely because it explains something formal logic never could: why people are good at reasoning about concrete, story-like situations and bad at reasoning about abstract, symbol-heavy ones. A model you can picture is a model you can actually use.

Kenneth Craik's *The Nature of Explanation* was essentially his only major published work — he died in 1945 at age 31, two years after the book appeared, yet it is credited as the founding text of the modern mental-models concept in cognitive science.

Facts decay, mechanisms transfer

Physicist Richard Feynman was known for a test he applied to his own understanding: if he couldn't explain a concept in plain language to someone with no background in it, he considered his grasp of it incomplete, however fluently he could recite the formal version. That test points at the same distinction Craik and Johnson-Laird were describing decades earlier: a fact you've memorized and a mechanism you understand are different cognitive objects, and only one of them survives being asked a follow-up question.

Cognitive science distinguishes between rote factual recall and what's sometimes called a causal model, an internal representation of how one thing produces another. Research on the "illusion of explanatory depth," documented by researchers Leonid Rozenblit and Frank Keil in a 2002 study published in Cognitive Science, found that people confidently rate their understanding of everyday mechanisms, such as how a zipper works or how a toilet flushes, far higher before being asked to actually explain the mechanism step by step. Once asked to write out the explanation, their confidence collapses (Rozenblit & Keil, Cognitive Science, 2002). The fact "zippers interlock two rows of teeth" is easy to state and easy to mistake for understanding; the mechanism of how the slider's wedge shape forces offset teeth into alignment is harder to state but, once actually grasped, doesn't evaporate the same way.

This matters because facts without mechanisms don't transfer to new situations. Someone who memorized "antibiotics kill bacteria" cannot predict why a given antibiotic fails against a specific resistant strain. Someone who understands the mechanism, that antibiotics target specific bacterial structures or processes and resistance arises when mutation or gene transfer changes that target, can reason about a new case they've never encountered before. The fact is a single data point; the mechanism is a generator that produces correct predictions about data points you haven't seen yet.

This is also why the illusion of explanatory depth is worth taking personally rather than treating as a curiosity about other people. It doesn't just apply to zippers and toilets. It applies to anything you've absorbed secondhand and never had to reconstruct: how your own retirement account grows, how the antibiotic you were prescribed actually works, how the policy you vote on is supposed to function. The gap between "I've heard this explained" and "I could explain this" is exactly the gap the 2002 study measured, and it is almost always wider than people expect before they try.

Two well-documented biases that quietly distort the models you already have

Even a well-built mental model degrades if the inputs feeding it are systematically skewed, and psychologists Daniel Kahneman and Amos Tversky spent much of their careers documenting exactly how that happens. Two of their findings are especially relevant to mental-model building because they don't just produce wrong answers, they produce false confidence in wrong models.

The first is confirmation bias, the tendency to search for, interpret, and recall information in ways that confirm a model you already hold, while underweighting information that would challenge it. If your mental model of a market, a person, or a policy is wrong, confirmation bias makes that wrongness self-reinforcing: every ambiguous data point gets read as support, and every disconfirming data point gets explained away or forgotten. Kahneman and Tversky's broader research program on judgment under uncertainty, much of it collected in their landmark 1974 Science paper "Judgment under Uncertainty: Heuristics and Biases," established that these are not occasional lapses but structural features of how human judgment works under normal conditions, not just under stress or fatigue (Tversky & Kahneman, Science, 1974).

The second is the availability heuristic, the tendency to judge the likelihood or frequency of an event by how easily examples of it come to mind, rather than by its actual base rate. A model built on availability rather than data systematically overweights vivid, recent, or emotionally charged cases and underweights common but unremarkable ones. This is why people tend to overestimate the risk of dramatic, widely reported causes of death and underestimate the risk of common but rarely reported ones, a pattern Tversky and Kahneman documented directly in their heuristics-and-biases research program. A mental model shaped by availability isn't necessarily built on bad mechanisms, it's often built on a skewed sample of which mechanisms got noticed in the first place.

How do confirmation bias and the availability heuristic interact to make a bad model feel solid?

They compound. Availability supplies the vivid, memorable cases a model gets built from in the first place; confirmation bias then filters everything afterward to fit that initial impression. The result is a model that feels well-supported by "everything I've seen," when what actually happened is a biased sample followed by biased filtering, not an unbiased test of the model against reality.

Counterexamples are not optional extras

A mental model that only accounts for typical cases is fragile precisely where it matters most, at the edges. Understanding "how vaccines work" only through the lens of a healthy adult's immune response leaves you unable to reason about why immunocompromised patients need different vaccination strategies, or why some vaccines require boosters and others don't. The boundary cases aren't decoration on top of the core mechanism, they're often what reveals whether you understood the mechanism or just memorized its most common application.

This is also where confirmation bias does its quietest damage. A model that has never been tested against a genuine counterexample looks identical, from the inside, to a model that has survived one. The only way to tell the difference is to go looking for the case that should break your model if it's wrong, rather than waiting for one to show up on its own. Investors who built a mental model of "housing prices don't fall nationally" going into 2007 had plenty of confirming data from the previous several decades; they had almost no exposure to the counterexample that mattered, because it hadn't happened yet in their lifetime of direct observation.

How do you actually test whether you understand something, rather than just recognize it?

Try to generate a new example the original explanation didn't give you, or predict what would happen if one condition changed. If you can only recognize the original example as correct but can't extend the logic to a new, related case, you likely have a memorized fact rather than a working model of the mechanism behind it. This is essentially the same test the 2002 illusion-of-explanatory-depth study used, just applied deliberately rather than sprung on you by a researcher.

Building the model, step by step

A durable mental model typically needs four components working together, and skipping any one of them tends to produce a model that feels complete but fails the first time it's tested against something slightly unfamiliar.

1. The mechanism itself. What causes what, and through what pathway. Not "vaccines protect you," but "vaccines expose the immune system to a harmless version or fragment of a pathogen, prompting it to produce antibodies and memory cells that respond faster if the real pathogen appears later."
2. A concrete example. An abstract mechanism stays abstract until it's anchored in something specific and checkable. Johnson-Laird's own research emphasized this: people reason more reliably about a specific, imaginable scenario than about the same logical structure presented in pure symbolic form.
3. A counterexample or edge case. The situation where the simple version of the model breaks down, forcing you to specify the conditions under which the mechanism actually applies.
4. An explicit statement of the model's limits. Not "this is roughly true," but a specific description of where the mechanism stops holding and what takes over instead.

The table below maps four techniques people commonly reach for when building mental models, and where each one is genuinely suited versus where it tends to produce a false sense of understanding.

Technique	What it does	Best used for	Where it misleads
First-principles reasoning	Breaks a claim down to its most basic, verifiable components and rebuilds understanding from there, rather than reasoning by analogy to convention	Novel problems with no good precedent, or situations where "that's just how it's done" is the only justification on offer	Slow for well-understood domains; can waste time re-deriving things that are already reliably established
Analogical mapping	Transfers structure from a familiar domain (electricity as flowing water, natural selection as a filter) to an unfamiliar one	Fast initial intuition-building, teaching, and communicating a mechanism to a non-expert	Analogies break down at the edges; treating the analogy as the mechanism itself, rather than a scaffold for it, produces confident errors
Inversion	Asks what would cause failure or the opposite outcome, then works backward to avoid it, rather than asking directly what produces success	Risk assessment, avoiding catastrophic and hard-to-reverse mistakes, stress-testing an existing model	Doesn't by itself generate a positive plan; needs to be paired with a forward-looking model to be actionable
Feedback-loop mapping	Traces how outputs of a system loop back to become new inputs, distinguishing reinforcing loops from balancing ones	Systems that unfold over time — ecosystems, markets, habits, organizational dynamics	Requires more information than a snapshot explanation; easy to mis-identify the direction or strength of a loop without real data

No single technique substitutes for the others — first-principles reasoning and inversion catch what analogies smooth over, while analogical mapping and feedback-loop mapping supply the intuition that pure first-principles work can be too slow to build alone.

Munger's latticework, and why more models beat one good model

Investor Charlie Munger, Warren Buffett's longtime business partner, gave a 1994 speech at the University of Southern California titled "A Lesson on Elementary, Worldly Wisdom As It Relates to Investment Management and Business," in which he argued that no single mental model, however good, is sufficient for reasoning about complex real-world problems. His core claim was that facts only become usable knowledge when they're organized on what he called a "latticework of mental models" drawn from multiple disciplines, not stored as isolated facts. As he put it, "you've got to have models in your head, and you've got to array your experience, both vicarious and direct, on this latticework of models."

Munger's specific point was that relying on one dominant model from one discipline, whether that's pure economics, pure psychology, or pure engineering, produces the kind of error he called "man with a hammer syndrome": to someone who only has a hammer, every problem looks like a nail. A business problem viewed only through an accounting lens will miss the psychological incentives driving the numbers; a policy question viewed only through an economic-efficiency lens will miss the second-order behavioral effects that determine whether the policy actually works as designed. Multiple partial models, cross-checked against each other, catch errors that any single model, applied alone, would sail straight past.

This is where the four components from the previous section and the four techniques in the table connect back to Craik and Johnson-Laird's foundational claim. A single mechanism, a single example, and a single edge case describe one model. A latticework is what you get when you hold several such models of the same domain simultaneously, built with different techniques, and use disagreement between them as a signal that something needs more investigation rather than a nuisance to be resolved by picking whichever model you already liked.

What people get wrong about building mental models

The most common mistake isn't building a bad model, it's mistaking model-collection for model-building. Reading a list of Munger's roughly eighty to ninety recommended mental models, or a blog post summarizing cognitive biases, gives you the vocabulary of mental models without the underlying skill the vocabulary is supposed to represent. This is the illusion of explanatory depth again, one level up: it's possible to feel like you understand "mental models" as a topic while being unable to apply any single model of yours to a genuinely new problem.

A second common mistake is treating analogies as literal descriptions rather than partial maps. An analogy such as "electricity is like water flowing through pipes" is genuinely useful for building initial intuition about current and resistance, but it breaks down badly if pushed toward explaining voltage sources, AC frequency, or semiconductor behavior. The failure isn't in using the analogy, it's in forgetting that an analogy is scaffolding for a mechanism, not a substitute for learning the mechanism itself.

A third, subtler mistake is confusing confidence with correctness. Because confirmation bias makes an untested model feel just as solid as a tested one from the inside, the feeling of "this makes sense to me" is not evidence the model is right. It is, at best, evidence that the model is internally consistent and hasn't yet met a case designed to break it. The only reliable signal is what happens when you deliberately go looking for that case.

What this means

A better mental model isn't one built from more facts, it's one built from facts connected by an explicit mechanism, tested against at least one case where the simple version breaks, and held loosely enough to be checked against other models rather than defended reflexively. That structure is what separates knowledge you can recite from knowledge you can actually use to reason about something new. It is the same structure Craik proposed in 1943, that Johnson-Laird formalized in 1983, that Kahneman and Tversky's research shows most people's models silently drift away from under the pull of confirmation bias and availability, and that Munger spent a career arguing has to be built in plural, as a latticework, rather than relied on one model at a time.

---

Read next:

• How to read a research paper without getting lost
• Why analogies help and hurt science learning
• Why facts need context to be useful

FAQ

Q: What is a mental model, and who first proposed the concept?

A mental model is an internal, simplified representation of how a system works, used to predict outcomes rather than just recall facts. Psychologist Kenneth Craik first proposed the concept in his 1943 book The Nature of Explanation, arguing that the mind builds small-scale models of reality to anticipate events before they happen.

Q: What is the illusion of explanatory depth?

It's a documented cognitive bias, first described by researchers Leonid Rozenblit and Frank Keil in a 2002 study in Cognitive Science, where people rate their understanding of everyday mechanisms as much higher than it actually is, until asked to explain the mechanism step by step, at which point their confidence in their own understanding drops sharply.

Q: How do confirmation bias and the availability heuristic distort mental models?

Both biases, documented extensively by psychologists Daniel Kahneman and Amos Tversky, skew which information a model gets built from and which information reinforces it afterward. Availability makes vivid or recent cases feel more common than they are, and confirmation bias then filters new information to fit the model that vivid data already produced.

Q: What did Charlie Munger mean by a "latticework of mental models"?

In a 1994 speech at the University of Southern California, investor Charlie Munger argued that facts only become usable knowledge when organized across multiple mental models from different disciplines, rather than relied on as isolated facts or a single dominant framework. He warned that depending on only one model produces "man with a hammer syndrome," where every problem gets forced to fit the one tool available.

Q: Why do mechanisms transfer to new situations better than memorized facts?

A memorized fact only applies to the specific case it was learned in. A mechanism describes the causal pathway connecting cause and effect, which allows someone to reason about new, related situations they haven't directly encountered before, a distinction central to Philip Johnson-Laird's cognitive science research on mental models.

Q: How can someone test whether they truly understand a concept?

Try generating a new example on your own or predicting what would happen if a key condition changed, rather than just recognizing the original example as correct. Difficulty doing either suggests the understanding is closer to memorization than a working causal model.

Q: Why are counterexamples important when building a mental model?

Counterexamples reveal the boundaries of a mechanism, showing where a simplified explanation stops applying. A model that only accounts for typical cases often fails precisely in the edge cases that matter most for real decision-making, and deliberately seeking out counterexamples is one of the few reliable checks against confirmation bias.