Rationality Is the Exception: Humans Were Not Built to Think

Aristotle called man the rational animal, and the Enlightenment built its institutions on that description: give people facts and a free press and they will reason their way to good decisions. It is a flattering picture, and it is wrong. Rationality is not what the human brain does by default. It is a rare, effortful, late-arriving trick that a few people manage some of the time, usually with heavy institutional scaffolding, and that most people never manage at all. It is incidental to what the brain was built for, the way a calculator app is incidental to a phone.

This is not a philosophical opinion. It is the consistent finding of a century of experimental psychology, and it is the plain reading of two decades of social media data. The evidence is worth laying out in one place.

What the Laboratory Found

The experiments below are classics because they replicate, and because the failures they show up are not failures of the stupid or the uneducated. They are failures of Harvard doctors, Stanford undergraduates and the general population alike.

The Wason selection task (Wason 1966). Four cards, one simple rule ("if a card has a vowel on one side, it has an even number on the other"), and the question which cards you must turn over to test it. Fewer than 10% of people get it right. Reframe the identical logic as a social rule ("if you are drinking beer, you must be over 18") and about 75% get it right (Griggs and Cox 1982; Cosmides 1989). The brain has no general logic module. It has a cheater-detection module, and logic problems that happen to look like cheating get solved.
The conjunction fallacy (Tversky and Kahneman 1983). Linda is 31, outspoken, and was a philosophy major concerned with social justice. Is she more likely to be a bank teller, or a bank teller who is active in the feminist movement? Roughly 85% of respondents picked the second, which cannot be more probable than the first, because it is a subset of it. A story that sounds right beats a probability that is right.
Base-rate neglect (Casscells, Schoenberger and Grayboys 1978). Sixty staff and students at Harvard Medical School were asked: a disease has a prevalence of 1 in 1,000, the test has a 5% false-positive rate, a patient tests positive, what is the chance they have the disease? The correct answer is about 2%. The most common answer was 95%. Only 18% got it right. These were the people who read your test results.
Anchoring, framing and availability (Tversky and Kahneman 1974, 1981). Spinning a rigged wheel of fortune before asking what percentage of UN members are African countries moved the answers. Describing the same public-health programme as "200 of 600 will be saved" or "400 of 600 will die" flipped the majority's choice. The judgement was not computed from the facts; it was computed from the packaging.
The Cognitive Reflection Test (Frederick 2005). A bat and a ball cost $1.10 together and the bat costs a dollar more than the ball; what does the ball cost? Most people, including most students at selective universities, say ten cents. The right answer is five, and it takes one line of arithmetic to check. People do not check. The intuitive answer arrives first and feels like knowledge.
Asch's conformity experiments (1951). Asked which of three lines matches a reference line, with confederates confidently giving the wrong answer, subjects went along with an obviously false answer on about a third of trials, and three quarters of them did so at least once. The lines were not ambiguous. The room was.
Telling more than we can know (Nisbett and Wilson 1977). Shoppers asked to pick the best of four identical pairs of stockings chose the rightmost pair four to one, and every one of them gave a reason about quality. Nobody mentioned position, and when it was suggested, they denied it. The reasons we give for our decisions are generated after the decision, by a part of the mind that was not present for it.
Choice blindness (Johansson et al. 2005; Hall et al. 2012). Subjects chose which of two faces was more attractive, were handed the other one by sleight of hand, and then explained fluently why they had chosen it. The same trick on a political survey had voters arguing for the opposite of the positions they had just ticked. If reasoning were driving belief, this could not happen.
Moral dumbfounding (Haidt 2001). Given a scenario of consensual, protected, secret, one-time sibling incest that harmed no one, people insist it is wrong, fail to produce a reason that survives scrutiny, and then say "I don't know why, it just is." The verdict comes first; reasoning is the press office.
When prophecy fails (Festinger, Riecken and Schachter 1956). A small group predicted that a flood would end the world on 21 December 1954. When the date passed, the members who had given up their jobs and possessions did not lose their faith. They became more committed and began proselytising. The disconfirmation was absorbed as confirmation.
The Dunning-Kruger effect (1999). Those with the least competence in logic, grammar or humour rated themselves well above average, because the skill needed to judge one's performance is the same skill that was missing.

Keith Stanovich, who spent a career measuring this, concluded that rational thinking is a distinct ability from intelligence, that IQ tests do not capture it, and that on tasks of rational thinking performance is poor across the whole intelligence range. Hugo Mercier and Dan Sperber went further in The Enigma of Reason (2017): reasoning did not evolve to find truth. It evolved to win arguments and to evaluate other people's, which is why we are excellent at spotting the flaw in an opponent's case and nearly blind to the flaw in our own. Confirmation bias is not a bug in the reasoning module. It is the reasoning module doing what it was selected for.

Why Mathematics and Science Are So Hard to Teach

Every child on earth learns to speak without a teacher. No child on earth learns algebra without one. That asymmetry is the whole story.

David Geary (1995) distinguished biologically primary abilities from biologically secondary ones. Primary abilities (language, face recognition, reading other people's intentions, intuitive physics, small-number sense) come for free through play, in every culture, without instruction. Secondary abilities (reading, arithmetic beyond a handful, algebra, formal logic, the scientific method) are cultural inventions that must be forced onto a brain with no circuitry for them, through years of effortful, motivated, externally imposed practice. Language is universal to every human society ever recorded. Writing was invented perhaps four times. Science was invented once.
Stanislas Dehaene showed what "no circuitry" means. Reading works by neuronal recycling: it hijacks a patch of visual cortex evolved for recognising objects. Arithmetic hijacks an approximate magnitude system in the parietal lobe that gives us exact quantities only up to about four and a blurry logarithmic sense after that. The Mundurucu of the Amazon, who have number words only up to five, do approximate arithmetic on large quantities as well as French adults and exact arithmetic not at all (Pica et al. 2004). Exact arithmetic is not something humans do. It is something a few cultures trained themselves to do against the grain.
Our built-in physics is Aristotelian. Michael McCloskey (1983) asked students to draw the path of a ball leaving a curved tube. A large fraction, including students who had passed university physics, drew it continuing to curve. Asked where a package dropped from a moving plane lands, they drew it falling straight down. The Force Concept Inventory later showed that a full semester of traditional physics lecturing moves these intuitions barely at all (Hake 1998): the average gain in conventional courses was about a quarter of what there was to gain.
Probability is the worst of all. When Marilyn vos Savant published the correct solution to the Monty Hall problem in 1990, she received around ten thousand letters telling her she was wrong, about a thousand of them from PhDs. Paul Erdős, one of the great mathematicians of the century, refused to accept it until shown a simulation.
Literature, by contrast, asks the brain to do what it does at the dinner table: track who wants what, who is lying, who is in love, who will betray whom. That is folk psychology, a biologically primary skill, running on the same circuitry as gossip. Every human culture has stories; Jerome Bruner called narrative the native mode of human thought. A literature class can be taught to a room of fifteen-year-olds because the fifteen-year-olds arrived with the hardware. A physics class has to override the hardware.

This is why people dislike mathematics and science, and the dislike is not a character flaw. Sian Beilock's group found that in people with high maths anxiety, merely anticipating a maths problem activates the same brain regions that register bodily pain (Lyons and Beilock 2012). Science is a collection of claims that contradict every intuition you have: the ground is moving at a thousand miles an hour and you cannot feel it, invisible animals make you sick, your ancestors were fish, the solid table is mostly empty space. Nobody's brain finds that comfortable. That science exists at all, that it appeared late, in one place, and nearly died several times before it took, is the strongest evidence that it is not a natural product of human cognition.

Why People Believe Things That Are Almost Certainly False

Take religion, because it is the largest and best-documented case.

Humanity has worshipped thousands of gods. The major living religions are mutually exclusive: the truth of one entails the falsity of the others. Any honest prior for "the one I happened to be raised in is the true one" starts somewhere near zero and does not improve, because the single best predictor of a person's religion is not any argument or evidence but the religion of their parents and the latitude and longitude of their birth. A born-again Baptist in Alabama would be a devout Hindu had he been born in Varanasi, and he would be equally certain. Beliefs that track geography rather than evidence were not produced by evidence.
Nobody performs this calculation, and that is itself the finding. The question "what are the odds that of all the religions, mine is the correct one?" is almost never asked by believers, not because it is answered but because it does not occur. Reasoning is not consulted.
Cognitive science explains why belief comes so easily. Humans have what Justin Barrett called a hyperactive agency detector: a rustle in the grass is treated as a predator, because a false alarm is cheap and a missed tiger is fatal. The same machinery sees faces in clouds, intent in weather, and a mind behind the universe (Guthrie 1993; Boyer 2001). Deborah Kelemen showed that children are intuitive theists, explaining rocks as pointy "so that animals can scratch on them" and rivers as existing "so that boats have something to float on". Teleology is the default; it has to be trained out, and in most people it never is.
Terror management theory (Greenberg, Solomon and Pyszczynski 1986): remind people of their own death and they cling harder to their worldview and punish outsiders more severely. Religion's central promise is that death is not the end. A belief that removes the worst fear a mind can have does not need to be true to be held.
Religion is also membership. Richard Sosis found that religious communes with costlier requirements survived longer than secular ones. A belief held as a badge of belonging is not held as a hypothesis, so evidence against it is not evidence against anything the believer cares about. It is an attack on the tribe.

None of this is special to religion. The same machinery produces astrology, homeopathy, conspiracy theories and nationalism. Religion is just the case with the longest record.

Hope Beats Facts

Tali Sharot's group found that people update their beliefs asymmetrically: told that a bad outcome is more likely than they thought, they barely move; told it is less likely, they move a lot (Sharot, Korn and Dolan 2011). The brain tracks good news and discards bad news, and this shows up in the neural signal, not just in the questionnaire. Around 80% of people show the bias.
Neil Weinstein (1980) found that students rated themselves as far less likely than their peers to be divorced, fired, or diagnosed with cancer, and far more likely to own a home and live past eighty. Everybody cannot be below average risk.
Patients with metastatic lung or colorectal cancer, receiving chemotherapy that every oncologist knew was palliative, were asked whether the treatment might cure them. 69% of the lung patients and 81% of the colorectal patients said yes (Weeks et al. 2012, New England Journal of Medicine). They had been told. They had heard something else.
Students asked how long their thesis would take estimated 34 days on average. It took 56. Asked for a worst-case estimate under the assumption that everything went wrong, they said 49 days, which was still less than reality (Buehler, Griffin and Ross 1994). The planning fallacy has not been cured in the thirty years since.
Americans spend about a hundred billion dollars a year on lottery tickets, against an expected return that every buyer could compute and none does.

There is an evolutionary reason hope wins. Alloy and Abramson (1979) found that mildly depressed people judge their own control over events more accurately than healthy people do. Accuracy about your prospects is depressing, and depression does not forage, court, or fight. Natural selection built optimists, and optimism means discarding facts that point the wrong way.

Facts Are Ammunition, Not Input

If people were reasoning from facts to conclusions, giving them more facts would move them. It does not. It arms them.

Biased assimilation (Lord, Ross and Lepper 1979). Supporters and opponents of the death penalty read the same two studies, one finding deterrence and one finding none. Each side judged the study supporting its view to be rigorous and the other to be flawed. After reading identical evidence, both sides were more confident than before. The same facts polarised.
Motivated numeracy (Kahan et al. 2013). Subjects were given a data table and asked whether it showed a skin cream working. The numerate got it right. The same numbers, relabelled as gun-control data, split them: highly numerate liberals and conservatives each got the answer right when it supported their side and wrong when it did not, and the more numerate they were, the larger the gap. Skill at reasoning was recruited to defend the tribe. The same team had already found that greater science literacy makes people more polarised on climate change, not less (Kahan et al. 2012).
Motivated scepticism (Taber and Lodge 2006). Given arguments on affirmative action and gun control, subjects spent longer reading and rebutting opposing arguments than agreeing ones, rated congenial arguments stronger, and, when allowed to choose, sought out their own side. The politically sophisticated did all of this more, not less.
Drew Westen (2006) put committed partisans in a scanner and showed them their own candidate contradicting himself. The regions associated with cold reasoning stayed quiet. The regions associated with emotion and conflict lit up, and when the subject found a way to excuse the contradiction, the reward circuitry fired. Rationalisation is literally rewarding.
Corrections do not reliably backfire, as was once feared (Wood and Porter 2019), but they do very little: they shift stated beliefs slightly, the shift decays within weeks, and it almost never changes the vote or the behaviour that the belief was defending.

Jonathan Swift wrote in 1721 that reasoning will never make a man correct an ill opinion which by reasoning he never acquired. Ziva Kunda formalised it in 1990 as motivated reasoning: we ask "must I believe this?" of unwelcome evidence and "can I believe this?" of welcome evidence, and the answers are almost always no and yes. A fact that supports the feeling is kept. A fact that contradicts it is not weighed; it is dismissed, reinterpreted, or forgotten. Jonathan Haidt's summary is the best one: the emotional dog wags the rational tail.

Facebook's news feed launched in 2006, and Twitter the same year. Since then, several billion people have had every sentence they publish scored for engagement, in real time, by systems that learn what humans respond to and serve them more of it. Nobody designed those systems to reward outrage. They were designed to reward whatever humans reward, and they measured the answer with a precision no psychology lab could approach. The results are in.

Falsehood outruns truth (Vosoughi, Roy and Aral 2018, Science). Across 126,000 stories shared by three million people on Twitter between 2006 and 2017, false news was 70% more likely to be retweeted than true news, reached 1,500 people about six times faster, and spread deeper and more broadly. The effect was strongest for political news. Bots spread true and false stories equally; the asymmetry was entirely human. False stories were more novel and evoked more surprise and disgust.
Moral emotion is the fuel (Brady et al. 2017, PNAS). Across 563,000 tweets on gun control, climate change and same-sex marriage, each moral-emotional word ("attack", "shame", "greed", "evil") increased the expected retweet rate by about 20%. The spread was almost entirely within the ideological group, not across it.
Attacking the other tribe works best of all (Rathje, Van Bavel and van der Linden 2021, PNAS). Across 2.7 million posts by news outlets and politicians, each additional word referring to the political out-group increased the odds of a share by 67%. Out-group language was the single strongest predictor of engagement, far stronger than praising one's own side, and the most shared posts were the angriest.
Exposure to the other side makes it worse (Bail et al. 2018, PNAS). Paying Twitter users to follow a bot that retweeted the opposing side for a month made Republicans substantially more conservative and Democrats slightly more liberal. The Enlightenment prescription of exposure to diverse views was tested and it backfired.
Facebook's own researchers agreed (the Haugen documents, 2021). After the 2018 shift to "meaningful social interactions", the angry reaction was weighted five times a like, and internal research reported that publishers and parties had shifted to outrage because it was what the ranking rewarded. Facebook's own data scientists warned that the change was making the platform angrier. The metric said it was working.
Nobody is even checking (Pennycook and Rand 2019, 2021). People who share false headlines mostly do not believe them; asked directly, they rate the headlines as inaccurate. They shared them because sharing is a social act, not an epistemic one, and the question of truth simply did not arise. A brief prompt to think about accuracy improved the quality of sharing a little, which shows how far from the default it is.

The algorithm did not make people tribal and emotional. It measured that they were, at a scale of billions, and then gave them what they measurably preferred. If humans were driven by facts, the sober, sourced, fact-checked post would win the engagement auction. It has been losing that auction every day for twenty years. The optimisation ran on us, and it found what the psychologists found: emotion first, tribe second, facts as decoration.

What Follows

Rationality is real. Science works, courts sometimes reach true verdicts, and the occasional person changes their mind on evidence. But every one of those cases runs on scaffolding: peer review, adversarial procedure, double-entry bookkeeping, controlled experiments, the norm that a claim is worthless until someone tried to break it. Strip the scaffolding and the reasoning collapses back to what the laboratory shows. Rationality is a fire that a few institutions keep lit with continuous effort. It is not the human condition. It is an incidental achievement of a small minority of humans, sustained by structures that most people never enter and do not miss.

The mistake of the last three centuries was to take that minority as the description of the species and to design public life for it. Give the public the facts, we said, and they will reason. They did not, and the last twenty years measured exactly how much they did not. Institutions that only work if the crowd reasons will fail, because the crowd does not. Institutions that work anyway, that make the rational move the low-effort move and that do not depend on anyone being persuaded, are the only kind worth building.

And the reader is not exempt. Neither is the writer. That is the last thing the evidence says: the person most sure they are the rational exception is running on the same hardware as everyone else.