Why We Generalise

KNOWLEDGE, JUDGMENT & SOCIAL CATEGORIES

The mind cannot treat every encounter as entirely new. Generalisations make experience manageable, but the same mental economy can flatten people into stereotypes. The challenge is to judge a shortcut by its evidence, its errors and who must pay for them.

A child who has known only an affectionate family dog may expect the next dog to welcome a hand held out for a pat. Usually that expectation will save the child from having to study every tail, tooth and movement from the beginning. On another occasion, however, the assumption could be dangerous. The very habit that makes the world navigable can also send someone confidently toward a mistake.

We are quick to recognise the problem as overgeneralising. One friendly animal does not prove that every animal of its kind is friendly; one discourteous tourist does not reveal the character of an entire nation. In social life, an unsupported generalisation about a nationality, gender, ethnicity or class is likely to be condemned as a stereotype. The criticism may be epistemic, because the judgment rests on poor evidence, and moral, because a person is reduced to a category.

Far less attention is paid to the opposite failure. Someone who never carried a lesson from one case to another would have to relearn at every doorway that handles open doors, at every kerb that moving cars can injure, and at every conversation that words can reassure or offend. Such a person would not be admirably free of bias. They would be stranded among disconnected details.

The central question, then, is not whether we should generalise. We cannot avoid doing so. It is how to distinguish a reasonable compression of experience from a careless or harmful one. That requires separating several issues that are often run together: whether a pattern is statistically strong, whether believing it helps us act, whether it survives contrary cases, and whether acting on it respects the individuals it concerns.

Why the Mind Needs Compression

Every scene offers more information than attention can absorb. Walking along a city street, we register faces, sounds, vehicles, signs, shopfronts and changes in the weather, yet focus on only a small portion. Past experience tells us which details can temporarily recede. A familiar bus is expected to follow its route; a red light is expected to stop traffic; a darkening cloud may prompt us to carry an umbrella. These expectations can fail, but without them ordinary action would slow to paralysis.

Generalisations perform this work by grouping encounters. They allow us to treat a new object as another bus, dog or storm rather than as a phenomenon with no history. Once the category is activated, a cluster of expectations follows. We predict what might happen, decide what deserves scrutiny, and reserve scarce attention for whatever departs from the pattern.

This economy is not merely laziness. A mind with finite time and memory must leave many facts in the background. Accuracy competes with speed, detail and the need to choose. An expectation that is right often enough may be practically valuable even if it is not exceptionless. A bus may pass a stop because it is full or out of service, but the expectation that buses collect waiting passengers remains useful.

Of course, usefulness does not turn an error into truth. Nor does the necessity of categorisation license every category we inherit. It explains why the temptation to generalise is deeper than prejudice or intellectual weakness. We should correct bad shortcuts, but a campaign to eliminate shortcutting itself would attack one of the conditions of intelligent action.

Beliefs Directed Toward the Future

In 1930, after Frank Ramsey died at only 26, his friend R. B. Braithwaite prepared for publication a manuscript Ramsey had completed in 1929. Issued posthumously as General Propositions and Causality, it reconsidered what it means to believe a general claim such as “all humans are mortal”. Ramsey’s answer shifted attention away from a static list of facts and toward a person’s readiness to act.

An older picture, associated with Ramsey’s earlier thinking and with Ludwig Wittgenstein’s Tractatus, treated a universal generalisation as an indefinitely long conjunction: this person is mortal, and that person is mortal, and so is every other person one could name. Ramsey came to see a difficulty. Two people might agree about the mortality of every human who has lived so far and still disagree about whether humanity is necessarily mortal.

Imagine one person who expects mortality to remain part of human life and another who thinks future technology could abolish it. They dispute the generalisation even though neither identifies an already immortal individual. Their disagreement concerns how present evidence should guide responses to cases not yet encountered. It is about the reach of a pattern, not a past item on a list.

Ramsey’s alternative was to understand a general belief as a habit or disposition. To believe humans are mortal is to be prepared to treat any human one meets as vulnerable to death: to avoid running them down, to provide air when they cannot breathe, and to plan under the assumption that a life can end. General beliefs belong to the system with which, in Ramsey’s memorable phrase, we “meet the future”.

Rather than asking only what arrangement of facts would make a generalisation true, Ramsey invites us to ask what the belief enables a person to do. A disposition earns confidence when it guides new encounters well. Treating humans as mortal has proved dependable. Treating every human as blond would collapse almost immediately. The contrast emerges when the habit is applied, not merely when a sentence is contemplated.

This practical account explains why generalising always ventures beyond the evidence already in hand. Its purpose is prospective. We use what has happened to prepare for what has not. If every application had to wait for complete information about the particular case, the generalisation would offer no savings. It would cease to organise attention precisely when guidance was needed.

That leap into the future creates the possibility of error. Yet fallibility is not by itself proof of irrationality. Maps omit trees and doorways so that roads remain legible. Categories omit differences so that recurring structure becomes visible. The question is whether the omission is suited to the task—and whether its mistakes can be tolerated.

All Cats, Cats in General

Not every generalisation demands the same degree of precision. A universal claim, such as “all cats have fur”, is defeated by one hairless cat. To preserve it, we could add exceptions: all cats except Sphynx cats have fur. The appearance of another hairless breed would require another clause. Soon the once-helpful statement becomes a cumbersome register of qualifications.

Everyday speech often uses a looser form. “Cats have fur” ordinarily does not mean that every cat does, or even explicitly state a proportion. Philosophers and linguists call this a generic generalisation. Generics attribute a characteristic to a kind while tolerating exceptions. “Buses stop for passengers” can remain acceptable when a full bus drives past. “Birds fly” survives penguins and ostriches.

The relationship between numbers and generics is surprisingly weak. Philosophers note that speakers accept “mosquitoes carry West Nile virus” even when only a small fraction are infected in the familiar example. Sarah-Jane Leslie has proposed that a dangerous or striking property helps explain why the generic feels natural. By contrast, roughly nine in 10 humans are right-handed, yet “humans are right-handed” sounds like an odd description of our species. Frequency matters, but it does not settle which generic claims speakers accept.

This looseness makes generic thinking remarkably efficient. We do not need exact population data before forming an expectation, and we need not discard the expectation whenever an exception appears. If a friend says she has adopted a cat, picturing fur is sensible. Discovering that the animal is a Sphynx produces a minor surprise, not a crisis in reasoning. The shortcut was inexpensive, and the correction is easy.

But tolerance of counterexamples has a darker implication. A belief may persist through many encounters that fail to fit it. Because the claim never specified how many members must possess the feature, evidence against it can be dismissed as another exception. The resilience that makes a generic useful can also make it stubborn.

From Pattern to Stereotype

Social stereotypes exploit the same flexible architecture. They associate a category of people with a trait and prepare us to see an unfamiliar member through that association. Like other general beliefs, they organise experience and support prediction. Like other generics, they can survive obvious exceptions. What changes is not necessarily the cognitive form, but the object of judgment and the human stakes of getting it wrong.

Stereotyping is often criticised as a failure to see a person as an individual. A stranger’s nationality, skin colour, gender or class becomes the basis for expectations about intelligence, honesty, temperament or danger. The individual has not supplied the evidence for that judgment, yet may have to answer for it. Opportunities can be withheld, behaviour interpreted suspiciously, and autonomy narrowed before the person has acted.

The supporting evidence is also frequently corrupt. Fictional portrayals may repeat the same image of a historically oppressed group until familiarity begins to feel like observation. Selective news coverage can make rare conduct appear characteristic. Institutional practices then produce apparently confirming data. If police stop one group more often because officers expect criminality, the resulting records cannot straightforwardly prove that the original expectation described an innate propensity.

A stereotype can also steer perception after an encounter begins. Ambiguous confidence may be read as leadership in one person and arrogance in another. The same mistake may be treated as a lapse by an expected high performer and as confirmation of incapacity by someone burdened with a negative group image. The belief does not merely predict evidence; it helps determine which evidence observers notice and how they classify it.

These feedback loops show why a stereotype cannot be assessed by asking only whether a correlation exists somewhere in the data. Categories may be broad, the comparison group unclear, and the observed pattern shaped by discrimination, geography, wealth or opportunity. Moving from a population average to a claim about an individual adds another inferential step. Moving again from that claim to unequal treatment adds a moral decision.

Still, it would be too simple to say that the urge to stereotype must arise from a uniquely irrational faculty. Stereotyping draws on the ordinary need to compress social information. That account makes the behaviour intelligible; it does not make the result accurate, fair or permissible. A causal explanation is not a moral acquittal.

Nor is every general statement about a human group automatically a stereotype in the same sense. Demographers, historians and public-health researchers need categories to describe disparities and direct resources. Trouble grows when an aggregate finding is treated as an essence, when a complex category is assumed to explain itself, or when the statistic is carried into a personal judgment without relevant evidence. The purpose and level of the inference matter: diagnosing an institutional pattern is different from assigning a character to a stranger.

Fallible Heuristics, Limited Minds

The study of heuristics helps clarify this distinction. Daniel Kahneman and Amos Tversky mapped how people use fast rules of thumb when reasoning with uncertainty and limited resources. Some produce systematic mistakes. Yet a thinker who attempted a complete statistical calculation before every everyday choice would be unable to function. Human judgment is bounded by time, information and attention; recognising those limits is not an insult.

One famous experiment described Linda as bright, outspoken, philosophically educated and involved in campaigns against discrimination and nuclear weapons. Participants often judged it more likely that she was both a bank teller and active in the feminist movement than that she was a bank teller. The first option seemed more representative of the description, but it cannot be mathematically more probable: everyone who is both is also a bank teller. Later debate has examined how wording and presentation affect interpretation of the task, without altering that basic set relation.

The error reveals the pull of resemblance. People asked to judge probability may instead ask which story fits a familiar type. In a laboratory task, that substitution violates elementary logic. In daily life, however, sensitivity to representative patterns can generate quick, useful hypotheses. The responsible response is not to pretend that intuitive pattern-matching can be removed, but to know when the decision requires a slower standard.

Generalisations are similarly double-edged. They are efficient because they ignore some information and extend beyond what has been observed. Those same features make them vulnerable to false confidence. The appropriate threshold should therefore vary with the decision. Guessing that a newly adopted cat has fur demands little investigation. Deciding whom to hire, search, insure, diagnose or punish demands much more.

This is where the language of “rational” and “irrational” can mislead. A shortcut may serve a recognisable cognitive purpose while producing an indefensible judgment in a particular setting. It may be understandable as a response to limited attention yet unreasonable after contrary evidence appears. It may even improve the predictor’s narrow success rate while imposing unjust costs on those being predicted.

So cognitive utility cannot finish the moral argument. Explaining why a belief is tempting tells us where correction must begin. It does not tell a person harmed by the belief that the treatment was acceptable. Understanding the machinery of stereotyping should increase accountability, not dissolve it.

The relevant comparison is rarely between an imperfect shortcut and omniscient judgment. Usually it is between several feasible methods: pausing for another question, checking a record, using a more relevant category, postponing a decision or admitting uncertainty. A heuristic becomes less defensible when a modest effort could replace it with substantially better individual evidence. Limits on cognition are real, but so are our capacities to design routines that compensate for them.

Count Errors—and Their Costs

A promising way to judge a generic belief is to examine the disposition it creates. Two questions matter. How often does the belief lead to error? And what happens when it does? Neither answer alone is enough. A rarely mistaken prediction can still be unacceptable if its occasional failures are catastrophic, while a frequently mistaken precaution may be sensible if it is cheap and prevents serious harm.

Consider mosquitoes and West Nile virus. Most mosquitoes are not infected, yet where the virus circulates, the possibility of an infected bite can justify low-cost precautions. Repellent and protective clothing impose modest burdens compared with the illness an infected bite may cause. The cautious generic can prompt reasonable action even though most mosquitoes encountered will not carry the virus.

The calculation changes when the category consists of people. Suppose population-level evidence shows unusually high alcohol consumption in a region. It does not follow that a particular person from that region drinks heavily. Treating the stranger as though the group pattern were their biography can humiliate them, distort a medical or employment decision, and require them to disprove an assumption they did not create.

Crucially, the person making the inference may not bear most of the cost. A shopkeeper may feel safer after monitoring customers from a stereotyped group, an employer may feel efficient after filtering applicants by a crude proxy, and an official may feel prudent after directing scrutiny unevenly. The burdens—suspicion, lost opportunity, anxiety and diminished standing—fall elsewhere. An honest assessment must count those distributed costs, not only the decision-maker’s convenience.

Accuracy itself also needs careful description. A belief might modestly improve prediction across thousands of cases and remain a poor guide to any individual. It might track a temporary social condition rather than a stable trait. Or the attempt to act on it may change behaviour and institutions in ways that sustain the pattern. A bare success rate cannot tell us whether the category is relevant, the mechanism understood, or the treatment justified.

For this reason, generic beliefs about social groups should usually face a higher standard than generics about harmless objects. People possess plans, histories and the capacity to explain themselves. They are not merely instances from which an observer extracts a regularity. Respect for autonomy gives us reason to replace a category-based guess with individual evidence whenever the decision materially affects a life.

How to Challenge a Generalisation

Karl Popper famously placed the search for counterexamples at the centre of scientific criticism. At the level of logic, no finite series of white swans can prove that every swan is white, but one black swan refutes the universal claim. Actual scientific practice is less tidy: measurement error and supporting assumptions can mean that one anomalous result does not immediately overturn a theory. Still, universal statements clearly advertise what would make them false.

Generic claims are harder. One hairless cat does not refute “cats have fur”, and perhaps a hundred would not. Ordinary speakers rarely know the exact point at which a tolerated exception becomes evidence against the generalisation. A person who wants to preserve a social stereotype can exploit that vagueness indefinitely, classifying every counterexample as unusual while every apparent confirmation is treated as typical.

A better discipline begins by seeking disconfirmation even when no single case can force surrender. Record the misses as seriously as the hits. Ask whether the sample reflects the population or only the places you happen to look. Examine whether a vivid story, inherited image or institutional practice is doing more work than direct evidence. And specify the decision the belief is being used to support.

The frequency-and-cost test can then be made more demanding. How common is the supposed trait, and compared with what baseline? Does the category improve prediction after relevant circumstances are considered? How severe are false positives and false negatives? Who receives the benefit of caution, and who bears the mistake? Can the person affected offer evidence about their own case before the judgment becomes consequential?

Different settings will produce different answers. Public-health planning may legitimately use group-level patterns to allocate outreach while still avoiding assumptions about every patient. A researcher may study an average difference without claiming that it defines each participant. In a personal encounter, where the individual’s conduct and testimony are available, continuing to privilege a crude group expectation becomes much harder to defend.

Revision also requires attention to confidence. We need not replace one absolute with another. A weak prior expectation can be held lightly, updated quickly and prevented from determining high-stakes action. The more harmful a mistaken application would be, the stronger the evidence required and the more opportunities there should be for review. This is not the elimination of generalisation; it is the governance of it.

Language can help. Saying “a higher rate has been measured in this sample under these conditions” leaves room for mechanism, variation and uncertainty. Saying “those people are like that” turns a contingent pattern into a character. The second form is easier to remember and deploy, which is precisely why it deserves suspicion.

See the Pattern, Meet the Person

The need to generalise and the duty to respect individuality remain in tension. The first answers a cognitive limitation: no one can process every case without categories. The second answers a moral reality: a person can be harmed when an observer’s economy of thought becomes the frame through which their choices are interpreted. Neither obligation simply cancels the other.

Recognising this tension prevents two opposite errors. One is to treat every group generalisation as proof of stupidity, ignoring why predictive shortcuts are unavoidable. The other is to treat cognitive inevitability as permission, as though whatever arises naturally from limited attention must be morally innocent. We can understand a mechanism while demanding better evidence and fairer action.

The child approaching an unfamiliar dog illustrates the balance. From a beloved pet, the child has learned many sound generalities: dogs bark, need food, communicate through posture and are capable of attachment. Without carrying those lessons forward, the child’s knowledge would be impoverished. The dangerous step is not learning from one dog at all, but extending the friendliest feature without enough attention to the cost of being wrong.

Human encounters demand an additional correction. In many such encounters, the person can describe intentions, contest a category and supply a history. We can begin with broad knowledge, but we should allow the individual to displace it. The point of a first impression is to orient inquiry, not to close it.

Institutions bear an even greater responsibility because their shortcuts are repeated at scale. A casual expectation may become a hiring filter, risk score, policing priority or medical protocol. Repetition gives small biases cumulative force, while automated or bureaucratic systems can hide the judgment inside a procedure. Regular audits, transparent criteria, appeal routes and attention to uneven error rates are ways of making institutional generalisations answerable.

None of these safeguards will produce a mind without categories. That is neither a realistic nor a desirable aim. The better ambition is a mind conscious of what categories cost: one that knows its compressed picture is incomplete, notices when the stakes rise, seeks cases that resist the pattern, and yields to evidence about the person standing in front of it.

Generalisations let us move through a complicated world because they tell us what can, for the moment, be ignored. Wisdom lies in knowing when the ignored detail is actually the important one. A stereotype becomes most dangerous when it conceals that choice—when a convenient expectation is mistaken for the whole truth about someone else. We will continue to meet the future with assumptions. The ethical task is to keep those assumptions revisable and never make other people bear the unexamined costs of our need to simplify.