A new colleague sits through his first design review. The night before he had spotted a hole in the proposal and written down two lines about it. When his turn comes, the five people ahead of him have all said it looks fine. What he says is: "Roughly the same view here."
This page takes that one second apart. It is easy to read as "crowds are just sheep", and the real machinery is finer than that: the switch that makes you change your answer usually is not the head count; group boundaries often fail to line up with cultural differences; and as for favouring your own side, a group made by a coin toss will do it in ten minutes.
In 1951 Solomon Asch ran an experiment at Swarthmore College so simple it is almost comic. You see a card with one standard line, and another with three lines; which one matches? Alone, you would almost never get it wrong. But seven or eight people sit in the room with you, all of them briefed confederates, and on twelve critical rounds they announce the same wrong answer in unison. About a third of critical responses went along with the majority, and three-quarters of participants conformed at least once.
The useful part is the variations that followed. Asch changed unanimity rather than size: if one member of the majority gives a different answer, conformity drops from roughly a third to around five percent. And it does not matter whether that person is right — he only has to differ from the rest. So the lever is unanimity, not numbers. What pins you down is not five people; it is the position of being the only one.
There are also two quite different reasons to change your answer, separated by Morton Deutsch and Harold Gerard in 1955. One is that you genuinely think so many people cannot all be wrong, so you must have misjudged: informational influence. The other is that you know they are wrong but do not want to be the lone deviant: normative influence. The split earns its keep, because the remedies do not transfer — the first calls for independent evidence, the second calls for making disagreement cost nothing.
How stable is any of this? In 1996 Rod Bond and Peter Smith pooled 133 studies using this task across 17 countries: conformity tracks how collectivist a society is, and the American studies show conformity declining steadily since the 1950s. So the conformity rate is not a human constant; it is a social quantity that moves with place and period.
One notch up is obedience. In the experiments Stanley Milgram began at Yale in 1961, 26 of 40 people (65%) in the baseline condition raised the voltage to the top of the board. That number now needs footnotes: in 2013 Gina Perry went through the Yale archives and found the experimenter frequently improvising off-script pressure, and many participants doubting on the spot that the shocks were real; in 2009 Jerry Burger ran a version stopping at 150 volts, where 70% chose to continue against Milgram's comparable 82.5%. What has been overturned is not "people obey" but the explanation. Milgram's agentic state — the person hands over responsibility and executes like a machine — has few takers now; the more accepted reading is engaged followership: people stay because they have been persuaded they are helping with something worthwhile. The two readings imply guarding against different things: authority in one case, a narrative in the other.
And one study that was once cited more than any of these has to come off the list. The 1971 Stanford prison experiment was long treated as proof that a situation can turn ordinary people into abusers; in 2019 Thibault Le Texier reconstructed it from archives, unreleased recordings and interviews with 15 participants: the guards had been explicitly told to be tough, and were prompted when they were not tough enough. It can no longer be used as evidence. Ranked by evidence: Asch's paradigm replicates well, Milgram's numbers need discounting and his explanation has been replaced, Stanford goes to zero.
From 1960 to 1961 the anthropologist Michael Moerman lived for a year in a village near Chiang Kham in northern Thailand, meaning to settle a question that looks easy: who are the Lue? He tried the criteria one by one — what they speak, what they wear, how the houses are built, how they farm, what they worship — and each either swept in the neighbouring Thai or cut out part of the village. The only reliable criterion left was: the people who call themselves Lue. And the same person, in another setting, will say he is Thai. In 1965 he published the paper under that same question as its title.
This is not "they are confused about who they are". It is that an ethnic name is not a label for a distinct cultural inventory; it is a position in a system of classification — and the position shifts with the setting.
In 1969 Fredrik Barth's edited volume Ethnic Groups and Boundaries turned this into a general rule: study the boundary, not the "cultural stuff" it encloses. His key observation is exactly what the right panel shows — personnel flow across, and the boundary holds. His own fieldwork on the Afghan-Pakistani frontier records an asymmetry: southern Pathans who lose their land can become Baluch, but not the reverse. The reason lies in how each side reckons membership — Pathan identity hangs on genealogy and land, so losing the land drops you out; the Baluch system is a hierarchy of clientage that can take outsiders in.
The line that holds: a group boundary is not the natural result of cultural difference; it is maintained. Which swaps in a different set of questions: who maintains this line, what does crossing cost, what kind of person gets across. All of them can be checked against fieldwork.
Evidence: both are ethnographic cases, not falsifiable hypotheses; what they rest on is how often later fieldwork has found them usable. The strongest objections deserve stating: a boundary-first view underrates the inertia of cultural content — language and habit are not swapped at will; and Barth leans on individual strategy, which cannot account for boundaries that stand unchanged for generations with almost nobody crossing.
In 1971 Henri Tajfel and colleagues took a set of fourteen- and fifteen-year-old boys in Bristol. They looked at paintings by two artists and were told the grouping followed which painter they preferred — in fact it was random. Then each boy allocated points to other participants, identified only by a code and a group, none of whom he knew, and from which he himself got nothing. On paper there is no reason to favour anyone.
They favoured anyway, and in a peculiar way. Rather than "give my own group the most", they more often chose the allocation that maximised the gap — even when their own group ended up with less than another option would have given it. The paradigm has been replicated many times over; it is one of the few very stable results in this corner.
The conclusion has to be stated precisely or it slides. Marilynn Brewer's 1999 paper puts it firmly: what minimal groups generate is mainly favouring your own side, not hostility to the other. Psychologically and in their consequences these are not the same thing, and they do not have to travel together — reading these experiments as "people are innately xenophobic" overreads them. Much real-world favouritism needs no hostility at all: who gets thought of, who gets recommended, who gets trusted by default.
So where does hostility come from? The textbook answer is Muzafer Sherif's 1954 Robbers Cave camp: two groups of boys form separately, a tournament is arranged, hostility follows fast, and it eases only through tasks neither group can complete alone — fixing the water line, hauling a stalled truck. The current status has to be stated: in 2018 Gina Perry found in the archives that Sherif had run a nearly identical design in New York state in 1953 and it failed — cross-group friendships were too strong for the boys to turn on each other, however hard they were prodded — after which he barely mentioned it again; and the experimenters themselves did a good deal to push the conflict along. So Robbers Cave should no longer be used as clean evidence that hostility arises on its own.
The contact side actually has the harder evidence. In 2006 Thomas Pettigrew and Linda Tropp synthesised 515 studies, 713 independent samples and more than 250,000 people: intergroup contact correlates with lower prejudice at around −0.21 on average, stably in that direction, with checks for publication bias and self-selection. The framing to keep: most of the data is cross-sectional, and within any single study you cannot rule out that tolerant people seek contact in the first place. What is solid is the direction and magnitude, not "contact necessarily reduces prejudice".