The Evaluating Self

25Studies of Independence and Conformity

Solomon E. Asch, "Studies of Independence and Conformity: I. A Minority of One Against a Unanimous Majority", Psychological Monographs 70(9), whole no. 416, 1956

Linked to 27 chapters: supports 11 · contradicts 2 · mechanism 18 · same claim 4 · pairing 7

A minority of one against a unanimous majority The task A third yielded,a quarter never Interviews Nine variations Not suggestion 18 comparisons, 12of them critical. Amajority of 7 to 9,unanimous and wrong. 36.8 per cent of thecritical estimateswent with the group;24 per cent never did. Independence andyielding each in threekinds; concern wasuniversal, unrelated. Speaking aloud, thesize of the error andof the standard: allmove the effect. No argument, rewardor punishment. Thepressure came frominside the subject. Asch built the situation to study independence, and reports the two thirds who kept it first. What the interviews add is that the same error count can stand for quite different acts.
Figure 25.0 — The monograph in five moves: a disagreement manufactured about a plain fact, the size of the effect, the interviews behind it, nine variations that move it, and Asch's argument about what it is not.

Solomon Asch set out to study independence and is remembered for its opposite. He seated one student among seven to nine others who had met the experimenter beforehand, and asked the group to say aloud which of three lines matched a standard line on a second card. The task was so easy that people judging alone were right on more than 99 per cent of the comparisons. On twelve of the eighteen trials the others announced, unanimously and without strain, a line that was plainly wrong. Across 123 subjects at three colleges, a third of those critical judgements went the majority's way. Asch's interest was in the other two thirds, and in what the interview afterwards showed about both. A quarter of his subjects never moved at all. Yielding took three forms and independence took three, and they did not line up with the error counts. A person's stand was set within the first two or three trials and then held. The effect fell to a third of its size when the subject wrote his answer instead of speaking it. Asch rejected the explanations then available — uncritical acceptance, suggestion, reward and punishment — as restatements rather than accounts, and argued that what he had built was not a demonstration of conformity but a situation in which one person had to decide what to do about a contradiction he could not understand.

Plan of the investigation

The materials were plain. A white card carried one black vertical line, the standard; a second card, forty inches away, carried three comparison lines numbered one to three. Standards were three, five and eight inches, and the two wrong lines differed from the standard, and from each other, by amounts a person could not miss. Eighteen comparisons took about twenty minutes.

The single critical subject was recruited by a member of the majority for what was described as an experiment in visual discrimination, and arrived to find the group already waiting. He was seated next to last in the second row, so that he heard all but one of the others before his turn. The experimenter read the instructions aloud to everyone, and two confederates asked clarifying questions, so that all present appeared equally new to the task. The first member of the majority read the prearranged answers off a small card and the rest followed his lead. They were told to answer clearly and firmly, not to argue with the critical subject, not to look at him, and not to feign surprise at his answers. The majority was, in Asch's description, impersonal rather than aggressive.

The comparison, and where the dissenter sits majority's answer correct 1 2 3 standard the three comparison lines, forty inches away The order of announcement the only naive subject 1 2 3 4 5 6 7 8 Judging alone, people got these comparisons right more than 99 per cent of the time.
Figure 25.1 — The task and the seat. The critical subject hears six to eight unanimous wrong answers before he speaks, and the wrong answer is one he can see is wrong.

The series was built to let several things be read off it. The first two trials were neutral, the majority answering correctly, which lent it credibility; a third neutral trial fell in fifth position. The twelve critical trials were three basic comparisons shown four times each. On two showings of a comparison the majority chose the wrong line nearer the standard, on two the wrong line further from it, so that identical stimuli could be seen under a moderate and an extreme majority error, and its errors ran from three quarters of an inch to one and three quarters. The whole nine-trial set was then repeated without a pause, which let the two halves of the session be compared. The experimenter behaved as an impartial chairman, recording dissenting judgements without a show of surprise, and an individual interview followed immediately, before the subject had been told anything.

The majority effect

Thirty-seven people from a comparable college judged the same lines individually. Thirty-five made no errors at all, and the group's error rate was 0.7 per cent. Among the 123 who faced the majority, the mean was 4.41 errors out of twelve, or 36.8 per cent of the critical estimates. Asch states the headline both ways, and the second way matters as much as the first: the majority deflected a third of the estimates, and two thirds remained independent, which he takes as evidence that the perceived fact was the stronger of the two forces at work.

The distribution is the part most often dropped in retelling. It is not a hump around the mean but something closer to a J, with its mode at the truth. Twenty-nine subjects, 24 per cent, made no error whatever; about the same proportion, 27 per cent, went with the majority eight to twelve times; the remaining half were spread continuously between. On the six neutral trials the experimental groups made three errors in 738 judgements, so the errors were a product of the disagreement and of nothing else.

How the 123 subjects divided 29 8 10 17 6 7 7 4 13 6 6 4 6 0 10 20 0 1 2 3 4 5 6 7 8 9 10 11 12 errors on the twelve critical trials judging alone, 35 of 37 made no error at all The mode is at zero. A second cluster sits at the far end, and half the subjects lie between.
Figure 25.2 — How the 123 subjects divided, from the monograph's Table 3. The tallest bar is at zero errors, and the second cluster is at the other end.

The errors tracked the majority closely. When it erred moderately, every error the subjects made was moderate too; when it erred extremely, 80.7 per cent of theirs matched it and 19.3 per cent fell short. Those short errors came almost entirely from trials where the intermediate line lay between the truth and the majority's choice, which is why Asch reads them as compromises rather than mistakes.

Individuals were consistent with themselves. Those independent in the first half of the series were independent in the second. More than half of those who ever erred did so by the second critical trial, and only 7 per cent first erred in the second half. Pressure did not accumulate; early independence conferred something like immunity, and early yielding perpetuated itself. Asch is careful about what this does not show: such consistency could be a property of the person, or commitment to a course taken for accidental reasons in the first minute, and his data cannot separate them.

Forms of independence and yielding

All subjects noticed the disagreement at once, and their first reaction was puzzlement rather than doubt. They formed hypotheses — that they had misunderstood the instructions, that the lines were an optical illusion, that the group was following whoever spoke first — and gave them up as the disagreement recurred. What replaced the hypotheses was a question about persons: who is right here. Most subjects, including the most independent, felt some doubt about their own accuracy, and the doubt then turned inward. They reported fearing that they were the ones with the defect, that they would be seen as a misfit, a black sheep, a sore thumb, and several described the loneliness of the position more vividly than the disagreement. Suspicion of the setup was rare, and Asch explains why: concern about oneself arrived first and left no room for it.

Of the 94 subjects asked whether a ruler would have vindicated them, 71 per cent said yes, and that confidence tracked performance: 87 per cent of the wholly independent against 41 per cent of those who erred eight to eleven times. Doubt and temptation did the same in reverse. One question showed no such gradient. Eighty-two per cent said they had been concerned about the disagreement, and concern was unrelated to what they did; what differed was how it was handled. Yielding subjects also underestimated how often they had yielded, by four trials on average, and some who had followed the majority ten or twelve times reported two.

Six ways to answer the same question Independence Yielding Confidence the majority is wrong; conflict never grows Report what you see declines to say the majority is wrong Obligation thinks the majority right, answers anyway Perception complies without knowing it; very rare Judgement I am wrong, they are right; gives up the claim Action sees clearly, hides it to avoid standing out Error counts do not sort these. Two subjects with one score may be doing different things.
Figure 25.3 — The six modes Asch draws out of the interviews. Independence and yielding each come in three kinds, and the error count alone does not tell them apart.

Working from the extremes of the distribution, Asch sorts the reactions. Independence took three shapes. Some kept faith in what they saw, held the majority to be mistaken, and were freed by that confidence from most of the conflict; one said he was getting a certain pleasure out of being different. Others declined to pronounce on the majority at all and retreated to what they could not be argued out of, that this was how the lines looked to them. A third group was assailed by doubt, believed the majority was probably right, and reported what they saw anyway, because the task had asked for an accurate answer.

Yielding took three shapes as well. At the rarest, a few subjects appeared to comply without knowing it, and Asch is scrupulous about the limits of the evidence: there is no reason to think the lines looked different to them, only that the distortion happened somewhere in judgement and stayed out of their awareness. Far commoner was yielding at the level of judgement, where confidence collapsed, the subject concluded that he was wrong and the group right, and went as far as to feel he had forfeited the right to report his inferior impressions; several said they went along so as not to spoil the experiment. Third was yielding at the level of action, by people who saw the lines clearly, never confused themselves, and suppressed what they saw because being visibly the odd one out was intolerable. These were in the most pain, since they could not soften the conflict by becoming unsure.

Variations of the experimental conditions

Eight further experiments test what the first depends on, and several change nothing. Judgements of brightness on two colour wheels gave 39 per cent errors; lines a thirty-second of an inch wide, which removed the objection that the subjects had been comparing rectangles rather than lengths, gave 24.5 per cent; tripling the series to thirty-six critical trials gave 37.6 per cent, and no subject who stayed independent through the first eight trials erred afterwards; warning the group that the lines would be measured with a ruler at the end left the effect untouched at 36.7 per cent.

Two manipulations moved it. The first was publicity. When the critical subject was told he had arrived late and should write his estimates instead of announcing them, while the majority went on speaking, errors fell to 12.5 per cent, about a third of the original, and all ten that remained were moderate: nobody followed the majority when it went extreme. The subjects insisted the change had made no difference to them, which Asch notes is exactly what subjects say. Since every one of them assumed the experimenter would compare their answers with the group's, even this was not private, and he says the fully private case remained to be done.

What raised the effect and what lowered it judging alone, no majority writing the answer, not saying it largest majority error, 3/4 in four neutrals to each critical the main experiment no neutral trials after the start smallest majority error, 1/4 in 0.7 12.5 25.7 26.2 36.8 53.0 56.7 0 20 40 60 errors on critical trials, per cent Removing the need to say it aloud takes away most of the effect. A bigger error takes away some.
Figure 25.4 — Errors on critical trials under the conditions the monograph varied, from Tables 3, 17, 21 and 22. Writing the answer instead of saying it takes away most of the effect.

The second was the size of the disagreement, and it runs the opposite way to intuition. In the sixth experiment, eighteen subjects judged a standard against a single unequal line, with the discrepancy set at a quarter, a half or three quarters of an inch. Errors fell as the majority's error grew: 56.7 per cent, then 33.3, then 25.7. A gross error is easier to resist than a small one. Errors also rose with the size of the standard, from 30.5 per cent at three inches to 48.1 per cent at nine, which Asch admits he cannot explain and for which he offers a name rather than a mechanism: differences well above the threshold still vary in what he calls unquestionability, a dimension that does not show itself until a judgement is challenged.

As the error grows, fewer follow As the line grows, more follow per cent of judgements per cent of judgements 56.7 33.3 25.7 30.5 36.1 39.8 48.1 0 20 40 60 0 20 40 60 0.25 0.50 0.75 3 5 7 9 size of the majority's error (inches) length of the standard (inches) Both panels give the per cent of critical judgements that followed the majority (Table 21).
Figure 25.5 — The sixth experiment, from Table 21. The bigger the majority's error, the fewer the subjects who follow it; the longer the standard line, the more.

Two results the monograph does not contain belong here, because Asch had reported them in 1951 and refers to them throughout. The first is that the effect does not need a crowd. Against one opponent a subject is barely affected, against two he errs on about an eighth of trials, and against three he reaches the full level, which does not rise as the majority grows to eight or fifteen. What matters is not the number but the unanimity. The second is the partner. One other person answering correctly cuts the errors to a small fraction of their level, and the partner need not even agree with the subject: a confederate who gives a third, different wrong answer breaks the effect nearly as well. When the partner answers correctly at first and then switches to the majority, the subject's errors climb back towards the level of a man alone.

Structure of the experimental conditions

Asch closes by laying out what he had actually built, in fifteen numbered properties, because he thought the conclusions people were drawing from it exceeded it. The object of judgement was unambiguous and each person had independent access to it; the subject had no grounds for doubting the eyesight or good faith of the others; the unanimity was known to arise from separate judgements rather than a group decision, which made it hard to dismiss as chance. The conflict was irreconcilable, since the facts would not bend, and it was also unintelligible. In ordinary disagreement we can at least imagine why the other side sees it differently: the matter is complex, someone has overlooked something. Here nothing of the kind was available, which is why so mild a disagreement about lines produced distress.

It also lets him say what the situation was not. There was no persuasion, no argument, no reward and no punishment. The majority never addressed the subject, never disputed with him, never expressed displeasure. It was, he writes, "an entirely neutral chorus of voices." "Whatever pressure the subject felt grew solely within himself." That is the argument against reading the results as suggestion or as uncritical acceptance, the readings he had criticised at the outset for redescribing the behaviour in the guise of explaining it. Nothing was suggested to anyone. The subject supplied the pressure, out of his own reading of what the disagreement meant about him, and it is that supply, not the group's action, that the modes of yielding and independence describe. He ends by marking the boundaries: the participants were peers, temporarily assembled, not interacting; and the resistance at issue was the solitary kind, closer to courage civile than to the courage of people acting together.

Since 1956

The experiment became one of the most repeated in psychology, and the repetitions mostly found it. A meta-analysis by Rod Bond and Peter Smith in 1996 pooled 133 studies from seventeen countries and confirmed the effect along with two patterns Asch could not have seen from three American colleges in one decade. Conformity in the United States declined steadily from the 1950s onwards, so some of what the line task measured was the temper of a period. And it varied with culture: societies that rate interdependence highly produced larger effects, which fits Asch's own view that yielding is about the meaning a person assigns to standing apart rather than about eyesight. Attempts in Britain diverged. Stephen Perrin and Christopher Spencer, testing engineering students in 1980, recorded one conforming response in 396 trials and called the finding a child of the McCarthy years; their later work with unemployed West Indian men facing white majorities produced large effects, which points less at the decade than at what the subject has to lose.

The misreadings have proved more durable than the replications. The study is cited for the proposition that people conform, when its own numbers say two thirds of judgements were independent and a quarter of subjects never yielded once; Ronald Friend, Yvonne Rafferty and Dana Bramel documented in 1990 how textbook after textbook inverted the emphasis Asch had insisted on. It is cited as though the majority argued or threatened, when it did nothing but say numbers out loud. It is cited with the majority's size as the operative variable, when three dissenters do the work of fifteen and one ally undoes most of it. And it is cited as evidence about people in general on the basis of male American college students, tested in public, on a question with no stake attached.

What has lasted is not the third who yielded but the anatomy underneath: that resistance and compliance can wear the same error count and be different acts, that the decisive move happens in the first minute, that a single voice on your side is worth more than any argument, and that the pressure people feel from a unanimous room is something they manufacture themselves out of what they think the room will conclude about them.

Claims drawn from this chapter · 10

evidenceThe majority effect

Asch's wrong majority captured about a third of people's answers

When a unanimous group announced an obviously wrong answer, about a third of the judgements made by Asch's subjects went along with it, on a task people get right more than 99 per cent of the time when alone.

Because123 subjects at three colleges made a mean of 4.41 errors out of twelve, 36.8 per cent of critical estimates, against 0.7 per cent among 37 controls judging individually and three errors in 738 neutral judgements.
LimitsMale American college students, tested in public, on a question with no stake attached; two thirds of judgements were independent and a quarter of subjects never yielded once.
PracticeForce your own judgement into the open before the room's is known — a written estimate, a committed position — because the effect is on the answer given, not on the answer held.
After AIAn assistant is a confident unanimous voice with no visible stake, available on every question. Asch's subjects could see that the majority was wrong, which is the condition that mostly does not hold with a model.
opinionForms of independence and yielding

Counting wrong answers hides why a person agreed with the group

Asch found three different reasons for going along with the group and three for holding out, and counting a person's wrong answers cannot tell them apart. Someone who thinks the majority is right but reports what he sees scores the same as someone untroubled.

Becausethe individual interviews, conducted before debriefing and read from the extremes of the distribution, separate the kinds; Asch is explicit that distortion of perception, if it occurred at all, was the rarest form and cannot be established from the data.
LimitsInterview evidence from a small number of subjects at the extremes of the distribution, interpreted by the experimenter; Asch himself says the perceptual-distortion category cannot be established from his data.
PracticeIn review, watch what you are doing as you do it: holding a position, suppressing a doubt, or having no doubt at all. The behaviour is identical from outside and the three are not the same act.
After AIAccepting a model's output and agreeing with it are indistinguishable in the record you leave behind. Only attention at the time can tell them apart, and only you can supply it.
contestedSince 1956

How much people conform changes with the decade and the country

How much people conform depends on when and where the study was run, as much as on the person. Conformity in the United States fell steadily after the 1950s, and it is stronger in countries that value fitting in with the group.

BecauseBond and Smith's 1996 meta-analysis pooled 133 studies in seventeen countries and found both patterns, while Perrin and Spencer recorded one conforming response in 396 trials with British engineering students in 1980 and large effects with unemployed West Indian men facing white majorities — which points at what the subject has to lose.
LimitsContested: the cross-national comparison confounds culture with decade and with the populations tested, and the Perrin and Spencer results point at stake rather than at nationality.
PracticeAudit the conditions rather than the people. If dissent is rare on your team, the finding to act on is what standing apart costs here, not who is compliant.
After AIWhat a person has to lose by disagreeing is the moving part, and it has not gone anywhere. Disagreeing with a system the organisation has invested in has its own cost, and nobody has measured it.
evidenceThe majority effect (Table 3)

In Asch's experiment most people held out and a minority gave in

Asch's result is not that everyone gave in a little. 24 per cent of people never agreed with the wrong majority, 27 per cent agreed eight to twelve times, and the most common score was zero.

Becausethe J-shaped distribution is in the paper's own table, and Friend, Rafferty and Bramel documented in 1990 how textbook after textbook inverted the emphasis Asch insisted on.
LimitsThe split is observed in one paradigm on one kind of judgement; nothing here establishes that the independent quarter are independent elsewhere, and Asch's own interviews found three different routes to the same error count.
PracticeKeep a list of the judgements you will not delegate to a room or a tool, and treat independence as a disposition you can hold rather than a statistical exception. A quarter of Asch's subjects held it throughout.
After AIThe finding everyone cites — people conform — is not the finding. The one that matters for supervision is that resistance is common, unevenly distributed, and evidently a property of persons.
opinionStructure of the experimental conditions

In Asch's study the pressure came from the subject, not the group

The majority in Asch's experiment only stated answers; they did not argue, urge, reward or punish. Whatever pressure the subject felt he produced himself, out of what he thought standing apart would say about him.

Becausethe fifteen structural properties in the closing section are offered as an argument against reading the results as suggestion or uncritical acceptance, and the conflict is characterised as irreconcilable and, unusually, unintelligible.
LimitsAn argument from the structure of the situation rather than a manipulation of it; the claim that the pressure is self-generated is an interpretation of the interviews, and Cialdini reads the same phenomenon as social proof operating on the subject.
PracticeIn review, separate what the situation contains from what you have added to it. Asch's majority did nothing but say numbers, and a third of the judgements moved anyway.
After AIA model applies no pressure at all, by construction. If you find yourself deferring to it, the deference is yours — which is the most useful thing in this monograph and the least cited.
evidenceVariations of the experimental conditions

One dissenter mattered more in Asch's study than the majority's size

What made people give the wrong answer in Asch's experiment was that everyone else agreed, not how many they were. Three opponents produced the full effect and fifteen produced no more. One ally cut errors sharply, even when the ally was also wrong.

Becausethe partner results come from Asch 1951, reported in the monograph's discussion rather than in its own experiments, and when the partner defects to the majority the errors climb back toward the level of a man alone.
LimitsThe partner effect is reported from earlier work rather than tested here, and the relevant condition is that unanimity is broken — which says nothing about whether the dissent is informative.
PracticeArrange for one dissenting voice before a decision rather than more reviewers. A single ally who breaks unanimity is the cheapest intervention in this literature, and it works even when the ally is also wrong.
After AISeveral assistants agreeing is not several opinions; they were trained on overlapping data by overlapping methods. The unanimity is manufactured upstream, and the partner effect has no analogue inside a model.
evidenceVariations of the experimental conditions

People in Asch's study decided in the first trials and stayed

In Asch's study people settled early: over half of those who ever gave a wrong answer did so by the second trial, only 7 per cent first erred in the second half, and with 36 trials no one new gave in after the eighth.

Becausemore than half of those who ever erred did so by the second critical trial and only 7 per cent first erred in the second half, and tripling the series to 36 critical trials produced 37.6 per cent with no new defectors after the eighth.
LimitsEstablished on a repetitive perceptual task where nothing changed between trials; a setting in which evidence accumulates would not be expected to behave this way.
PracticeSpend your attention on the first exchange rather than on endurance. Whatever stance you take in the first minute of a review, with a room or with a model, is the stance you will keep.
After AIThe first answer an assistant gives sets your posture for the session. Deciding once, early, whether you are checking or accepting is a larger choice than it feels like at the time.
evidenceVariations of the experimental conditions (sixth experiment, Table 21)

People resisted the majority more when its error was obvious

In Asch's sixth experiment, as the majority's wrong answer moved from a quarter of an inch off to three quarters, errors fell from 56.7 per cent to 25.7. The more obvious the error, the easier it was to resist.

Becauseeighteen subjects across the discrepancy range give a monotone fall, and errors also rose with the length of the standard — 30.5 per cent at three inches to 48.1 at nine — which Asch names unquestionability and cannot explain.
LimitsEighteen subjects in one experiment, and the second pattern — more errors on longer standards — is reported without explanation by the author himself.
PracticeReview the near misses rather than the disasters. An output that is slightly wrong is the one you will accept, and the size of the error is inversely related to your chance of catching it.
After AIA model's plausible-but-wrong answer is precisely the small discrepancy. The obvious failures are the safe ones, and the failure rate you experience is concentrated where your resistance is lowest.
evidenceVariations of the experimental conditions (seventh variation)

People stop copying the group once they write their answer down

In Asch's experiment, letting people write their answer down instead of saying it aloud cut wrong answers from 36.8 per cent to 12.5 per cent, and no one gave a wrong answer many times.

Becausethe subjects insisted the change made no difference to them, and since all assumed the experimenter would compare their answers with the group's, the fully private case was never run.
LimitsEven the written condition was not private — every subject assumed the experimenter would compare — so 12.5 per cent is an upper bound on what privacy removes, not a measurement of the residual.
PracticeHave people write their estimate before anyone speaks. It is the same intervention as a cognitive forcing function, it costs a minute, and it removed the extreme errors here entirely.
After AICommitting to your own answer before reading the model's is the same manipulation. It is also the one users reliably dislike, which is why it has to be built into the process rather than left to intention.
evidenceForms of independence and yielding

People who went along with the group forgot how often they did

People who agreed with a wrong majority underestimated how often they had done so, by about four trials. Some who went along ten or twelve times reported two.

Becausethe interview totals set against the trial records show the gap, and confidence tracked performance elsewhere: 87 per cent of the wholly independent said a ruler would vindicate them against 41 per cent of those who erred eight to eleven times.
LimitsRetrospective self-report against a trial record in one session; the underestimate may be a repair of a conflict the subject felt rather than a failure of memory, and the design cannot separate these.
PracticeLog how often you took the tool's answer, as a count, at the time. Your memory of how independent you were will be about a third short, and the direction of the error is known.
After AINobody keeps this count with an assistant. The one measurement we have — developers 19 per cent slower while believing themselves faster — suggests the gap is at least as large as Asch's.

All claims · Source map