Overgeneralising has a bad rap. It seems an obvious, if understandable, mistake to generalise beyond what our evidence supports. A child playing in the park may assume, based on their experiences of their friendly pet golden retriever, that all dogs are friendly. But they would be wrong to approach an XL Bully, a breed of dog effectively outlawed in the UK recently following a string of attacks on children. When someone makes unwarranted generalisations about social groups, we condemn them for the sin of stereotyping. It is wrong to conclude after just one experience with a rude Brit that all Brits are rude.
We hear less about the problems of _under_generalising. Generalising comes so naturally that few of us are at any risk of undergeneralising, so we give it little thought. The focus on overgeneralising is understandable given its potential costs, such as the child assuming that a dangerous dog is just like the friendly dogs they’ve encountered to date. Generalising itself, though, isn’t a bias that needs to be corrected. It is a cognitive shortcut and a fundamental part of our mental life. We’d be lost without it. It gives our experience structure and frees us up to focus on the new at the expense of the old. The child who approaches the XL Bully and the person who stereotypes Brits risk making a mistake, but the source of their error is not their desire to generalise.
Our story begins in Cambridge, 1930. Searching through the papers of his recently deceased friend, the philosopher Richard Braithwaite discovered a sheaf of handwritten notes on the nature of general beliefs, such as the belief that all humans are mortal. The friend was Frank Plumpton Ramsey, the remarkable philosopher and mathematician described as a ‘genius by all tests for genius’ by Paul Samuelson, the Nobel Laureate in economics. The notes were written in September 1929, just four months before Ramsey’s death from a mysterious illness that he may have contracted from swimming in the River Cam. Ramsey died aged just 26 but would go on to leave an oversized legacy in philosophy, mathematics, economics and logic.
In those notes, later titled ‘General Propositions and Causality’ by Braithwaite, Ramsey advocated a view of generalisations that renounced both his own previous views and those of Ludwig Wittgenstein, on whom the young Ramsey, until only recently an undergraduate, exerted a significant influence during their time together in Cambridge in the late 1920s.
Ramsey wanted to know what it means to say that a person believes that all humans are mortal. He concluded that, if you believe a generalisation, you have a habit or disposition. If you believe that all humans are mortal, you will take care not to run over your friends with your car and to make sure they have air to breathe, because you are disposed to treat any human you meet as mortal. Together, these habits and dispositions make up the system with which we ‘meet the future’. They are how we orient ourselves in the face of new observations and experiences.
Two people can disagree in their beliefs about generalisations, even when they agree on all the facts about particulars
What is the alternative? Wittgenstein, with whose writing Ramsey was very familiar, having worked on the translation of the fiendishly cryptic _Tractatus Logico-Philosophicus_ (1921) from German into English while still an undergraduate, had argued that to believe a generalisation is to believe an infinite conjunction of facts about particulars: Alan is mortal and Bella is mortal and Carl is mortal and so on for every human being you could possibly mention. Ramsey’s new view saw beliefs in generalisations as involving a disposition to treat any human you encounter as mortal. That’s a bit different. After all, the disposition is a good one to the extent that it gets things right when tested. The disposition to treat any human I meet as mortal has not led me astray so far, but the disposition to treat any human as having blonde hair would falter quickly. That is why I don’t believe that all humans have blonde hair.
Why did Ramsey reject his earlier view? Ramsey saw that two people can disagree in their beliefs about generalisations, even when they agree on all the facts about particulars. Imagine that I am debating a techno-utopian who believes that mortality is only a contingent feature of humanity’s present existence and that, in principle, humans could live forever. I can agree with them that every human up to now has been mortal, but at the same time disagree over the question of whether all humans are mortal. The techno-utopian would disagree with me that all humans are mortal, even though no human has yet achieved immortality. If believing the generalisation involved believing only that each previous human had been mortal, we could not disagree without disagreeing about a particular human’s mortality. But we agree about all past humans having been mortal. That’s not contested. We disagree only about possible future humans, without disagreeing about any particular human. My disposition to treat humans as mortal would come unstuck in a possible future in which some tech wizard unlocks the secret of immortality. Our disagreement does not concern the facts about the mortality of any actual human, but rather the possible future applications of the disposition.
This new perspective helps to explain why Ramsey revised his earlier view of generalisations. Rather than asking _What does the world need to be like for a belief in a generalisation to be true?_, Ramsey asked _What is a belief in a generalisation for?_ His answer, as I’ve already suggested, is: _to meet the future_. Why? Human experience is far too complex for us to attend to every detail. We must be selective in the things we pay attention to. But how do we decide what to focus on and what to ignore? Beliefs in generalisations help to solve this dilemma. When I see a double-decker red vehicle in London, I can use my general knowledge about buses to predict that it is going to pull in at the next bus stop to pick up the person signalling with their arm. I don’t have to wait to see what the vehicle does, which frees me up to focus my attention elsewhere. These predictions don’t need to be infallible to be useful. When a bus is full or off duty, it won’t stop for passengers. Even if my beliefs sometimes lead me to make a mistake, I use them in generalisations to make predictions about what the future will be like, and I use these predictions to decide what to ignore.
Not all our generalisations are put to the same use. For Ramsey, whether we should believe a generalisation depends on the virtues of the habit that constitutes the belief. Universal generalisations admit no exceptions: upon meeting a single hairless Sphynx cat, I should no longer believe that all cats are furry. This feature of universal generalisations means that, if they are to be useful, they must be formulated with a high degree of precision. My belief that all cats are furry is too coarse-grained. I should perhaps have believed that all cats _except Sphynx cats_ are furry instead. But even that won’t be quite right, once I learn about Ukrainian Levkoys. And, as they become more precise, these beliefs become less useful, because they become more complex, with too many _ad hoc_ clauses. We go from the simple ‘All cats are furry’ to ‘All cats except Sphynx cats and Ukrainian Levkoys and this one and that one… are furry.’
Sphinx cat. Courtesy Wikimedia
Ramsey, a mathematically inclined philosopher, was interested in precisely formulated universal generalisations. But many of the generalisations that we believe don’t demand this kind of precision. I believe that cats are furry. How many hairless cats do I have to meet before I should abandon this belief? Generalisations like ‘Cats are furry’, without a quantifying word such as ‘all’ or ‘most’, are now known to philosophers and linguists as _generic_ generalisations. They are about what is generally true of a category (cats), but do not say what proportion of the category has the property in question (being furry). I can answer the question _How many cats are furry?_ by saying _Most cats are furry_, but not just _Cats are furry_. These are more common in everyday life than the universal generalisations that philosophers and logicians like to deal with.
My disposition to treat any cat as furry can lead me to error, but it gets things right enough of the time
As with the universal generalisations Ramsey discussed, to believe a generic generalisation is to be disposed to treat members of a category as having some property. Where they differ is that generic beliefs need not be revised in the face of one or even many counterexamples. I can hold on to my belief that cats are furry in the knowledge that many cats have no fur at all. I don’t need to revise my belief that buses pull in to pick up passengers just because they don’t when they are full. Things can get even more extreme: as the philosopher and cognitive scientist Sarah-Jane Leslie has argued, people will accept that mosquitoes carry the West Nile virus even when they are aware that only 1 per cent of mosquitoes are carriers. Sometimes, an incredibly strong correlation is not enough to warrant a belief in a generic generalisation: I don’t believe that humans are right-handed, even though around 90 per cent of us are.
All of this makes it extremely difficult to answer the question: what does the world need to be like for a generic generalisation to be true? Taking inspiration from Ramsey, we can ask instead: what are generic generalisations _for?_ They are generalisations about a category, but they don’t apply univocally to every member of the category. Because they are tolerant of exceptions, they demand less of us than universal generalisations do. They don’t require us to revise our beliefs in the face of counterexamples. This surprising and rarely mentioned feature is what makes them so useful: I can form many generic beliefs about categories without worrying about getting things precisely right in every case. My disposition to treat any cat I meet as furry will sometimes lead me to error, but it is still worth holding on to that disposition because it gets things right enough of the time. And when I get things wrong, it is unlikely that there will be much of a cost. Before encountering it, I might mistakenly assume that my friend’s new cat will be furry and then discover that it is a Sphynx cat, but errors of this kind are not going to keep me up at night.
So far, so good for generic generalisations. They simplify the world by abstracting away from individual differences. They are insensitive to proportions, so we don’t need precise statistical evidence to support them. They admit exceptions so we don’t continually need to revise them. But the very features that make generic generalisations so useful are also hallmarks of stereotypes. Universal generalisations don’t make for good stereotypes – they are too easy to falsify by finding a single counterexample. Generic generalisations are better suited to this task, since we can hold on to them even when we know of a large number of counterexamples.
To stereotype a category is to associate certain properties with the category and its members. Like other beliefs in generalisations, stereotypes are part of the system with which we meet the future. They impose structure on our experiences. We cannot attend to every individual difference but instead use these beliefs to simplify our experience and make predictions about the future. And we do not revise our beliefs in these generalisations every time we meet an exception, for fear of abandoning the structure they provide.
Stereotyping social groups, including those defined by nationality, gender or skin colour, is typically thought morally wrong, because it undermines people’s autonomy by focusing on their membership of a group, rather than their qualities as individuals. It is common to hear that stereotyping is not only a moral failing, but an _epistemic_ one, a failure of rationality or to fulfil our duties as knowers. We are sometimes even told that a tendency to stereotype is a consequence of low cognitive ability.
There are many ways that stereotyping might contribute to an epistemic failure. It often leads us to error, in part because the evidence from which we generalise is bad. Stereotypical portrayals of members of historically oppressed ethnic groups in film and television can form the basis for stereotypical generalisations about those groups, which in turn perpetuate the stereotype in future portrayals. Stereotyping might cause us to systematically misinterpret ambiguous evidence in a way that leads us to develop pernicious false beliefs. Black people in the UK are significantly more likely to be stopped and searched by police officers than white people. Holding a racial stereotype about that group might lead a person to conclude incorrectly that its members have a higher innate propensity to commit crimes.
Generalising means meeting the future with assumptions that make the world around us intelligible
Important though these troubling aspects of stereotyping are, I am more interested in the question of whether there is something inherently irrational in the epistemic attitude that leads to stereotyping.
To answer that question, we must consider why we stereotype at all. Stereotyping is a form of generalising. People stereotype because they need to simplify their experience. And because we all need to simplify our experience, we should not rule out stereotyping as inherently irrational. It may often lead us to error, but our mistake is not our desire to generalise.
Human cognition is limited. That is why mental shortcuts like generalising are an essential part of our mental lives. In Daniel Kahneman’s 2002 Nobel Prize-winning work on bounded rationality, summarised in his _Thinking, Fast and Slow_ (2011) and originally developed with Amos Tversky, he found that people employ a range of cognitive heuristics when asked to make complex judgments. Heuristics are ways of making decisions under conditions of uncertainty or limited resources. Because they allow us to ignore a lot of information, heuristics are not always accurate. In a 1983 study by Kahneman and Tversky illustrating the ‘conjunction fallacy’, people were presented with the following vignette:
> Linda is 31 years old, single, outspoken and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations.
Their participants, including those with sophisticated knowledge of statistics, judged ‘Linda is a bank teller and is active in the feminist movement’ to be more probable than ‘Linda is a bank teller’, given the information in the vignette. But the former entails the latter and so cannot be more probable. Instead of applying precise statistical reasoning, people are employing a ‘representativeness’ heuristic, where the first option is judged more probable than the second because the description of Linda is more representative of their idea of a _feminist_ bank teller than of a bank teller. It is tempting to interpret this as evidence of widespread irrationality, but to dismiss these heuristics out of hand is to ignore the value of such cognitive shortcuts. We often need to reason quickly or with only partial evidence, and heuristics are invaluable in such circumstances. A person who never employed heuristic thinking would be overwhelmed by competing cognitive demands. Heuristics, though fallible in systematic ways, are essential tools for navigating the complexity of experience.
Generalising is another kind of cognitive shortcut. Like Tversky and Kahneman’s heuristics, it may lead us to errors. This is because generalising inherently goes beyond our evidence. My dispute with the techno-utopian over the mortality of humans concerns what is possible for humans; we don’t disagree about the evidence that is available.
Generalising means meeting the future with assumptions that make the world around us intelligible, but at the expense of paying less attention to individual differences. Stereotyping is a way of making the social world intelligible. That does not make stereotyping right. It can be accurate or wildly inaccurate, and it lends itself to abuse. It can harm members of stereotyped groups in many ways, in part because it is a failure to see people as individuals, rather than simply as members of a category. The cognitive utility of stereotyping does not make it rational: holding on to a stereotype in the face of overwhelming evidence against it is surely irrational. But it does make stereotyping understandable and perhaps even morally excusable, since it reflects our need to make predictions about the future, based on the categories that things, including people, fall into.
How do we find a balance between undergeneralising and overgeneralising? One place to start is to seek out counterexamples to our generalisations. We can then rule out at least some of our general beliefs as false. This response to the problem of generalisations was proposed by the philosopher of science Karl Popper. He famously argued that, strictly speaking, we are _never_ justified in believing that universal generalisations are true on the basis of experience, since there is always another potential counterexample lurking beyond our observation. We are, however, justified in believing that a universal generalisation is false when we know of a single counterexample. A single immortal human would falsify the generalisation that all humans are mortal.
This proposal works well enough for universal generalisations, since we can easily state the conditions in which they are false. Popper’s concern was the practice of science, where the precision of universal generalisations is welcome. But many of our everyday general beliefs are not beliefs in universal generalisations. They are, instead, generic beliefs. The problem with these beliefs is that neither ordinary speakers nor trained linguists have a clear idea of the conditions under which generic generalisations are true or false. They are tolerant of exceptions, so the mere fact that there are some hairless cats does not mean my belief that cats are furry is false. And if we can’t say when they are false, then Popper’s suggested search for counterevidence will not tell us what (not) to believe.
Without clear criteria for truth and falsity, how can we say what evidence should lead us to revise our generic beliefs? Here’s my answer. We should, like Ramsey, approach this question pragmatically. There are two factors that we should take into account when evaluating our generic beliefs. Both rest on Ramsey’s idea that beliefs in generalisations are dispositions to be judged by their consequences. For any particular generic generalisation we should ask two questions. First, how often does it lead us into error? Second, how costly are the consequences of these errors? Our replies will not give us a definitive answer, as Popper hoped, to the question of whether we should revise our beliefs in generalisations. But, if we follow Popper’s advice and actively seek out counterexamples, and try to falsify them to test their strength, then they can act as a guide, helping us to weigh up the importance of getting things right and the consequences of getting things wrong.
My generic beliefs about social categories need to meet a higher standard than those about mosquitoes
If I believe that mosquitoes carry the West Nile virus, I am likely to be wrong most of the time when I judge individual mosquitoes to be carriers of the disease. But if I do get bitten by a mosquito carrying the virus, I may develop a serious illness. Although I am likely to be wrong a lot of the time, my error is worth it because I will take precautions against all mosquitoes, infected or not. This reduces my likelihood of contracting the disease.
Next, consider a stereotype about a social group: Scottish people drink too much. According to a press release summarising a 2025 report from Public Health Scotland, ‘people in Scotland are drinking 50 per cent above safe limits’. The stereotype may lead me to make accurate judgments in many cases but, when we consider the costs of the errors that I do make, my failure to treat those Scots as individuals and focus on them and their personal circumstances harms them. Indeed, this failure is sometimes claimed to constitute the intrinsic wrong of stereotyping. The costs of my errors need not be borne by me for me to take them into account. My generic beliefs about social categories may, therefore, need to meet a higher standard than those about, say, mosquitoes, since there is more at stake, morally speaking, when I make errors about other people.
The child who approaches the XL Bully expecting it to behave like their friendly golden retriever is making a dangerous mistake because they have overgeneralised from their experiences with their tame pet. The costs of that mistake might be very high. But what this example obscures is that many of the generalisations that the child makes on the basis of their experiences with their pet are true: dogs bark, dogs have four legs, dogs eat meat, dogs are furry. If the child failed to draw those conclusions from their admittedly meagre evidence, their knowledge of the world would be severely impoverished. What matters is that the child tests these generalisations against new evidence. And while it might be reckless, it is by approaching the XL Bully that they learn of their mistake. Popper would have approved.
Our need to generalise is in tension with the requirement that we respect each other’s autonomy. This is a tension between a cognitive obligation – to simplify experience and so make it comprehensible – and a moral one – to treat one another as individuals in all their rich specificity. It is easy to confuse a moral failing for a cognitive one but, when we criticise someone for overgeneralising or even stereotyping, we should remember how essential that cognitive shortcut is. Even stereotypes are part of the system with which we meet the future. To abandon that system wholesale would be to lose the in-built tools that enable us to focus on what is truly new.
_This essay is the winner of the inaugural_ _Aeon Philosophy Prize__._