What Makes Us Human

Have you ever wondered how humans can solve such complex problems—and yet sometimes make absurdly simple mistakes? In What Makes Us Smart, Samuel Gershman argues that both our remarkable intelligence and our fallibility stem from the same underlying computational principles. We are not perfectly rational machines, but adaptive organisms navigating a world full of uncertainty, scarcity, and ambiguity. Gershman’s provocative thesis is that our quirks and biases aren’t design flaws—they’re the price of intelligence given limited data and limited computational capacity.

According to Gershman, to understand the human mind, we need to think like engineers of an imperfect system: one that evolved under constraints. His framework draws on cognitive science, Bayesian statistics, information theory, and neuroscience to explain how the brain approximates rational thought while staying frugal with energy and computation. Across thirteen chapters, he constructs a grand theory linking perceptual illusions, learning biases, social conformity, moral reasoning, and even language design through two central ideas: inductive bias and approximation bias.

Why Our Brains Are Biased—And Why That’s Smart

The human brain does not have the luxury of unlimited information or unlimited processing power. To make sense of the world, it must rely on biases: assumptions about how things usually work or shortcut methods that make decisions fast but occasionally flawed. Inductive biases are the deep knowledge structures—like our sense of causality, object permanence, or linguistic rules—that help us generalize from few examples. Approximation biases, in contrast, are the small, computational shortcuts our brains take to save effort, like compressing sensory input or relying on limited memory samples.

Together, these biases form what Gershman calls the computational logic of cognition: the idea that intelligent systems must balance accuracy, speed, and cost. Our minds are forced to trade precision for efficiency, and in doing so, they exhibit predictable systematic errors that can actually be understood as signatures of optimal design under constraints. In one sense, illusions and biases are rational responses to an imperfect world.

From Bayesian Brains to Social Minds

Gershman begins with perception as the most fundamental case study: how the mind interprets uncertain sensory data. Borrowing from Bayesian probability, he shows how the brain behaves like a statistical reasoner—combining prior beliefs with new evidence to infer the most probable explanation. Visual illusions like the Ponzo illusion or the Moon illusion occur because our “priors” about perspective and depth are usually reliable, even if they misfire in contrived settings.

These same inferential mechanisms extend beyond sight into how we reason about others, make moral judgments, and construct scientific theories. We don’t simply process sensory data; we construct and revise intuitive theories of the world—what Gershman calls “mental models” or “intuitive theories.” These include intuitive physics (how objects move), intuitive psychology (what people want or know), and intuitive sociology (how groups behave). Each of these theories relies on built-in biases that enable rapid learning and reasoning but also resistance to change, explaining why people cling to beliefs even in the face of contradictory evidence.

Learning, Language, and Rational Illusions

Throughout the book, Gershman moves seamlessly between experiments, anecdotes, and thought-provoking examples. Children over-regularizing past tenses (saying “runned” or “goed”) reveal compositional learning rules that generalize from limited input. Adults imitating unnecessary actions (overimitation) show the deep inferential roots of social learning. Even our tendency to conform, exhibit optimism bias, or fall for confirmation bias can be reframed as efficient strategies for learning under uncertainty.

Later chapters extend this logic into economics and linguistics: the brain as an information-compressing device (efficient coding) and language as an optimized medium balancing informativeness and effort. Gershman concludes by returning to the original paradox—our simultaneous brilliance and foolishness—arguing that both emerge from the same computational architecture. If you understand those algorithms, you not only understand why people err—you also understand what makes them uniquely intelligent.

Why It Matters

In a cultural moment dominated by AI and debates about machine intelligence, Gershman’s argument has profound implications. It suggests that the mystery of human intelligence is not its perfection but its adaptivity. We are smart because we are biased, brilliant because we are limited. By modeling those limitations, cognitive science can bridge natural and artificial intelligence. The result is a deeply optimistic view: understanding our flaws is the surest path to understanding our genius.

Why does your brain see the moon as larger on the horizon than overhead? Why do you misjudge how long it will take to climb a hill or even mishear a word? Gershman argues that these perceptual errors—so-called illusions—are not failures of logic but necessary handiwork of a smart, economical brain.

Perception as Bayesian Inference

Gershman explains perception through the lens of Bayesian inference: the brain forms hypotheses about what it sees, combining prior expectations with incoming sensory evidence. For instance, when you see converging railway tracks in the Ponzo illusion, your brain infers depth. It concludes that identical lines must differ in physical length, leading to the illusion that the upper line is longer. Far from being irrational, this is exactly what a Bayesian agent would predict if it prioritized probable interpretations of ambiguous input.

This principle explains countless phenomena—from the Craik-O’Brien-Cornsweet illusion (where equal brightness surfaces appear different due to shading) to the ventriloquist effect, where a puppet’s lips dictate where you think the sound comes from. The mind continuously performs probabilistic blending, weighing the reliability of each cue (vision versus hearing) and producing the most likely unified interpretation.

Explaining Away and Integrating Evidence

At the core of Gershman’s argument is the concept of explaining away: when multiple hypotheses could explain an observation, confirming one naturally weakens the others. Your brain discounts less probable interpretations automatically. This same logic extends to reasoning beyond perception. When you see someone act rudely, you infer whether it’s their personality or the situation—what psychologists call the fundamental attribution error—but you tend to overweight personality because behavior feels stable and predictable.

In both cases, illusions demonstrate an optimization trade-off. The brain sacrifices accuracy for stability and predictability under noisy conditions. As Gershman parallels from Kahneman’s Thinking, Fast and Slow, our perceptual system is a “fast system” honed for survival, not perfect truth.

Confidence, the Hard–Easy Effect, and Rational Bias

Even our metacognitive errors—like being overconfident on difficult tasks and underconfident on easy ones—emerge from rational learning. Gershman uses Bayesian updating to show that confidence reflects posterior belief given noisy evidence. When tasks are unexpectedly easy, we underweight success because it exceeds our prior expectations; when they’re hard, we over-interpret partial success. The result is the hard–easy effect. Far from being irrational, this effect is mathematically predicted by optimal probabilistic inference with uncertain priors.

Illusions are not bugs but features of design.

They reveal how your brain constructs perception through logical, adaptive shortcuts to handle a world of limited information.

In essence, Gershman redefines what it means to be rational. Rationality is not about getting everything right—it’s about making the best possible inference with finite sensory evidence. Illusions prove that your mind is not foolish; it’s frugal.

If you’ve ever wondered how children can learn the rules of language after hearing only a handful of sentences, or how people can infer causality from sparse observations, the answer lies in inductive bias. Gershman defines inductive bias as the mind’s built-in assumptions about how the world works—innate predispositions that allow us to generalize rapidly from limited data.

Causality: The Glue of Human Understanding

We naturally assume the world is governed by causes, not coincidences. Experiments by Albert Michotte showed that people perceive causality even in simple animations of shapes bumping into each other. Likewise, Fritz Heider and Marianne Simmel found that moving geometric patterns were spontaneously interpreted as social agents with goals and feelings. Gershman calls this our causal bias—the belief that events are interlinked by physical or psychological mechanisms. This bias allows us to infer invisible forces, predict consequences, and build “intuitive theories” of the world, much like scientists construct models from limited data.

Compositionality: Building Infinite Knowledge From Finite Parts

Another universal bias is compositionality: the mind’s assumption that complex things are made of simpler components. Language exemplifies this beautifully—children use rules like adding “-ed” to form past tense (“cook” → “cooked”), overgeneralizing (“run” → “runned”) before correcting. This same principle extends to vision and concept formation. Psychologist Irving Biederman proposed that objects are recognized from combinations of 36 basic shapes, or “geons.” Modern research by Brenden Lake and colleagues found that humans recognize new handwritten characters by understanding their generative parts—how strokes combine to form letters—which reveals how compositional learning enables quick generalization from one example.

Objects and Structure Learning

We cannot help but perceive a world of discrete objects. Our “object bias” filters sensory chaos into stable entities that persist in space and time. Infants show this bias within months—tracking occluded objects and expecting continuity even when shapes change briefly (the “tunnel effect”). Gershman explains that such object-based reasoning allows us to generalize knowledge reliably because objects, unlike raw sensations, behave predictably.

Over time, these simple biases grow into sophisticated inferential structures through hierarchical Bayesian learning: we learn not just facts about the world but how to learn future facts more efficiently. For example, children develop a “shape bias” when learning object names, concluding that words generalize by shape rather than color—a principle inferred from experience with previous words. This is learning-to-learn at work.

In short, inductive biases are not flaws—they are the hidden syntax of intelligence. They make cultural transmission, language, and science possible by letting us infer deep structure from sparse evidence. Without bias, learning would be impossible.

Humans are social learners. We imitate, conform, take advice, and follow trends—not always wisely, but often optimally given the information available. Gershman shows how social behavior, from conformity to overimitation, reflects rational inference in social environments.

Conformity and Cascades

In Solomon Asch’s classic line-judgment experiments, people conformed to obviously incorrect group opinions. At first glance, this looks irrational. But Gershman reframes it as Bayesian reasoning under uncertainty. When an entire group agrees on a perception, that social evidence can outweigh one’s private evidence. This logic extends to social networks: in information cascades, once a few individuals adopt a belief, others rationally follow because their private information adds little compared to collective consensus—a process driving fads, financial bubbles, and revolutions.

Overimitation and Rational Trust

Chimpanzees are adept at emulation—skipping unnecessary steps to get a reward—whereas human children (and adults) tend to copy actions precisely, even irrelevant ones. This “overimitation” puzzled psychologists, but Gershman interprets it as rational reasoning about others’ intentions. Observers assume demonstrators are knowledgeable and efficient; if they performed extra steps, those actions must be subtly necessary. Thus, copying faithfully becomes a rational inference, not blind mimicry. Humans, unlike chimpanzees, infer mental causality—that others have goals worth imitating.

Advice Taking and Reliability

When we receive advice, we often struggle to weigh it properly. Studies find we underweight advice that diverges from our opinion—but Gershman notes that this too accords with Bayesian logic. Highly discrepant advice signals that the expert’s belief may stem from different assumptions or unreliable data. Conversely, advice from a trusted expert or consistent source will be integrated proportionally to perceived reliability. Even “doing the opposite of bad advice” is rational if negative correlation is known.

Through these lenses, social irrationalities appear rational once context is considered. We conform, imitate, and heed (or ignore) others not because we are sheep, but because it is statistically reasonable to rely on social information when private evidence is weak. Culture itself, Gershman suggests, is cumulative Bayesian inference distributed across minds.

Why do people cling to false beliefs? From anti-vaccine conspiracies to religious faith, Gershman argues that belief perseverance can be rational given how minds assign credit or blame to evidence. When facts conflict with our theories, we can always tweak auxiliary assumptions—supporting explanations—to protect the core belief. This idea parallels how scientists rescued Newton’s laws by hypothesizing unseen planets like Neptune.

The Bayesian Logic of Stubbornness

Using Bayesian confirmation theory, Gershman shows that beliefs are networks of hypotheses. We assign lower probability to central theories changing compared to peripheral auxiliaries. Thus, when disconfirming evidence appears, rational agents revise the less costly assumptions. For example, a religious believer might question the credibility of a witness instead of the existence of God; a scientist adjusts calibration assumptions before discarding core physics. This makes belief systems robust yet flexible.

Sparsity and Determinism in Intuitive Theories

Gershman identifies two principles that make human theories resistant to contradiction: sparsity (we assume few causes) and determinism (we expect regular, predictable links). These biases, while leading to resilience, are beneficial. They allow fast learning when one cause explains many effects. Children and adults both employ sparse, near-deterministic reasoning when forming causal models—a strategy balancing simplicity with explanatory power.

Explaining Polarization and Self-Protection

From stem-cell debates to climate politics, Gershman uses this framework to explain polarization: two individuals exposed to the same data diverge because they have different priors about reliability or hidden variables (auxiliaries). Optimistic people discount bad news by assuming invalid feedback; conspiracy theorists explain contradictions through hidden plots. Even the sense of a “true, good self” persists because we reinterpret immoral actions as situational, not dispositional.

To change minds, chip away at auxiliaries—not core beliefs.

Gershman concludes that persuasion must target peripheral assumptions—the scaffolding that makes beliefs resilient—rather than attacking the central dogma directly.

Resisting disconfirmation is not inherently irrational; it’s a design feature of coherent thinking. Robust theories, whether scientific or personal, must withstand noisy data. When human stubbornness stabilizes belief against every anomaly, it’s simply the Bayesian brain doing its job too well.

Your brain consumes a meager 20 watts—less than a light bulb—yet performs computations that no supercomputer can match. How? By embracing frugality. Gershman argues that the brain achieves efficiency through approximation biases: it compresses, samples, and ranks information rather than processing it exhaustively.

Efficient Coding and Rank Representation

Borrowing from Claude Shannon’s information theory, Gershman likens neurons to communication channels. To maximize efficiency, brains encode frequently encountered stimuli with shorter, more economical “codes.” This produces rank encoding: subjective magnitude (like brightness or loudness) is proportional to an item’s rank within context, not its absolute value. This explains the Fechner and Weber laws of perception—why we notice relative, not absolute, differences.

Rank encoding also shows up in economics: people judge their salary not by amount but by comparison to peers. Likewise, happiness depends less on wealth than on relative rank. These contextual effects, far from irrational, are what a resource-efficient brain should produce.

Range Effects and Adaptive Illusions

Because encoding is local, perception adapts to the range of current input. At night, your eyes become more sensitive to dim light; in bright sun, they desensitize. The same logic generates visual illusions like the Ebbinghaus effect—context shifts perceived size—and even economic biases such as loss aversion. Kahneman and Tversky’s famous “value function,” where losses loom larger than gains, mirrors the brain’s efficient coding curve for uneven reward distributions.

Frugality Beyond Perception

Efficiency governs memory, decision-making, and language. Judgments depend on samples of experience (the “Law of Small Experiences”), leading to anchoring and range effects in value estimates. Language itself exhibits compression: common words are shorter (Zipf’s law). Gershman merges psychology and engineering to show how all cognition arises from the quest to minimize energy, time, and entropy.

The frugal brain, in short, is an algorithmic marvel—rational not because it’s perfect, but because it makes the best possible use of scarce resources. Every illusion and bias is a footprint of efficiency.

Why are natural languages messy, ambiguous, and irregular—yet so powerfully expressive? Gershman proposes that language embodies the same trade-off between informativeness and effort that shapes all cognition. The goal of a good language is not clarity alone but efficiency: conveying maximum information with minimal cost.

Ambiguity as a Feature, Not a Bug

Artificial languages like Lojban or Esperanto tried to remove ambiguity by enforcing strict rules. They failed because perfect clarity demands unbearable effort. Natural language thrives precisely because ambiguity lets context do the heavy lifting. As Gershman notes via the Benjamin Franklin “John Thompson” sign anecdote (“John Thompson sells hats”), efficiency favors shorter, contextually inferable expressions. Ambiguity only wastes effort when context can’t disambiguate it—so languages evolve to balance redundancy and economy.

Color, Word Order, and Meaning

Languages across cultures carve the world into semantic categories shaped by communicative efficiency. For instance, every language’s color terms partition the color spectrum near optimal information boundaries: languages with fewer color words favor perceptually distinct or frequently referenced regions like red and yellow. Similarly, differences in word order (SOV, SVO) reflect efficiency trade-offs between noise resistance and word-marking redundancy. Even adjective order (“small red car,” not “red small car”) follows rational design: more subjective adjectives (“small”) come earlier to protect meaning under memory constraints.

Compositionality and Evolution

The miracle of language learning and cultural evolution is compositionality: infinite expressiveness from finite vocabulary. Gershman shows that languages evolve under iterative transmission to become both compressible and expressive—simple enough to learn, rich enough to communicate. Laboratory simulations (Simon Kirby’s studies) confirm that compressed-but-expressive languages emerge spontaneously when speakers must communicate meaning efficiently across generations.

Language isn’t chaotic; it’s the outcome of natural selection for cognitive efficiency. Every irregular verb, ambiguous noun, or syntactic quirk is evidence of human design by constraint. As George Zipf foresaw, effort and meaning dance together: from neurons to grammar, we say just enough—but no more.

Randomness seems like the opposite of intelligence. But Gershman shows that stochastic thinking—strategic randomness—is essential to adaptive learning, creativity, and decision-making. The brain, like modern AI systems, often uses sampling and noise to explore, approximate, and invent solutions efficiently.

Monte Carlo Minds

Inspired by physicist Stanislaw Ulam’s discovery of Monte Carlo simulations, Gershman suggests that brains perform probabilistic inference by randomly sampling possibilities rather than exhaustive calculation. Neural noise, often seen as a defect, may enable this process. The brain’s apparent variability—different reactions to the same input—reflects sampling from probability distributions of beliefs, allowing flexible, approximate reasoning.

Posterior Probability Matching

In decision experiments, people’s choices match outcome probabilities instead of always maximizing them—a behavior long seen as suboptimal. Gershman argues it’s exactly what you’d expect from a sampling system: a few samples yield proportional frequencies matching the underlying probabilities. From children inferring word meanings to adults judging categories, the mind’s randomness mirrors Bayesian sampling precision scaled to effort.

Exploration and Strategic Noise

Randomness also powers exploration. In uncertain environments—the “multi-armed bandit” problems—humans mix exploitation (choosing known good options) with exploration (trying unknown ones). Strategies like Thompson sampling and softmax decision-making mirror biological patterns: explore more when uncertain, settle as knowledge grows. Even competitive contexts, from soccer penalty kicks to animal defenses, exploit unpredictability to achieve equilibrium. Randomness, far from chaos, is a computational strategy for adaptability.

The final paradox of Gershman’s book is that noise—the very thing engineers try to suppress—is fundamental to intelligence. Our brains are not broken signal processors; they’re Monte Carlo machines humming with purposeful uncertainty.