Asch's Conformity Experiment

Solomon Asch, with experiments originally carried out in the 1950s and well-replicated since, highlighted a phenomenon now known as "conformity".  In the classic experiment, a subject sees a puzzle like the one in the nearby diagram:  Which of the lines A, B, and C is the same size as the line X?  Take a moment to determine your own answer...

The gotcha is that the subject is seated alongside a number of other people looking at the diagram - seemingly other subjects, actually confederates of the experimenter.  The other "subjects" in the experiment, one after the other, say that line C seems to be the same size as X.  The real subject is seated next-to-last.  How many people, placed in this situation, would say "C" - giving an obviously incorrect answer that agrees with the unanimous answer of the other subjects?  What do you think the percentage would be?

Three-quarters of the subjects in Asch's experiment gave a "conforming" answer at least once.  A third of the subjects conformed more than half the time.

Interviews after the experiment showed that while most subjects claimed to have not really believed their conforming answers, some said they'd really thought that the conforming option was the correct one. 

Asch was disturbed by these results:

"That we have found the tendency to conformity in our society so strong... is a matter of concern.  It raises questions about our ways of education and about the values that guide our conduct."

It is not a trivial question whether the subjects of Asch's experiments behaved irrationally.  Robert Aumann's Agreement Theorem shows that honest Bayesians cannot agree to disagree - if they have common knowledge of their probability estimates, they have the same probability estimate.  Aumann's Agreement Theorem was proved more than twenty years after Asch's experiments, but it only formalizes and strengthens an intuitively obvious point - other people's beliefs are often legitimate evidence.

If you were looking at a diagram like the one above, but you knew for a fact that the other people in the experiment were honest and seeing the same diagram as you, and three other people said that C was the same size as X, then what are the odds that only you are the one who's right?  I lay claim to no advantage of visual reasoning - I don't think I'm better than an average human at judging whether two lines are the same size.  In terms of individual rationality, I hope I would notice my own severe confusion and then assign >50% probability to the majority vote.

In terms of group rationality, seems to me that the proper thing for an honest rationalist to say is, "How surprising, it looks to me like B is the same size as X.  But if we're all looking at the same diagram and reporting honestly, I have no reason to believe that my assessment is better than yours."  The last sentence is important - it's a much weaker claim of disagreement than, "Oh, I see the optical illusion - I understand why you think it's C, of course, but the real answer is B."

So the conforming subjects in these experiments are not automatically convicted of irrationality, based on what I've described so far.  But as you might expect, the devil is in the details of the experimental results.  According to a meta-analysis of over a hundred replications by Smith and Bond (1996):

Conformity increases strongly up to 3 confederates, but doesn't increase further up to 10-15 confederates.  If people are conforming rationally, then the opinion of 15 other subjects should be substantially stronger evidence than the opinion of 3 other subjects.

Adding a single dissenter - just one other person who gives the correct answer, or even an incorrect answer that's different from the group's incorrect answer - reduces conformity very sharply, down to 5-10%.  If you're applying some intuitive version of Aumann's Agreement to think that when 1 person disagrees with 3 people, the 3 are probably right, then in most cases you should be equally willing to think that 2 people will disagree with 6 people.  (Not automatically true, but true ceteris paribus.)  On the other hand, if you've got people who are emotionally nervous about being the odd one out, then it's easy to see how a single other person who agrees with you, or even a single other person who disagrees with the group, would make you much less nervous.

Unsurprisingly, subjects in the one-dissenter condition did not think their nonconformity had been influenced or enabled by the dissenter.  Like the 90% of drivers who think they're above-average in the top 50%, some of them may be right about this, but not all.  People are not self-aware of the causes of their conformity or dissent, which weighs against trying to argue them as manifestations of rationality.  For example, in the hypothesis that people are socially-rationally choosing to lie in order to not stick out, it appears that (at least some) subjects in the one-dissenter condition do not consciously anticipate the "conscious strategy" they would employ when faced with unanimous opposition.

When the single dissenter suddenly switched to conforming to the group, subjects' conformity rates went back up to just as high as in the no-dissenter condition.  Being the first dissenter is a valuable (and costly!) social service, but you've got to keep it up.

Consistently within and across experiments, all-female groups (a female subject alongside female confederates) conform significantly more often than all-male groups.  Around one-half the women conform more than half the time, versus a third of the men.  If you argue that the average subject is rational, then apparently women are too agreeable and men are too disagreeable, so neither group is actually rational...

Ingroup-outgroup manipulations (e.g., a handicapped subject alongside other handicapped subjects) similarly show that conformity is significantly higher among members of an ingroup.

Conformity is lower in the case of blatant diagrams, like the one at the top of this page, versus diagrams where the errors are more subtle.  This is hard to explain if (all) the subjects are making a socially rational decision to avoid sticking out.

Added: Paul Crowley reminds me to note that when subjects can respond in a way that will not be seen by the group, conformity also drops, which also argues against an Aumann interpretation.

YouTube video of a conformity experiment:

---

Asch, S. E. (1956). Studies of independence and conformity: A minority of one against a unanimous majority. Psychological Monographs, 70.

Bond, R. and Smith, P. B. (1996.) Culture and Conformity: A Meta-Analysis of Studies Using Asch's ( 1952b, 1956) Line Judgment Task.  Psychological Bulletin, 119, 111-137.

Comments

[James Bach]

I don't see this exercise as being so much about rationality as it is about our relationship with dissonance. People in my community (context-driven software testers) are expected to treat confusion or controversy as itself evidence of a potentially serious problem. For the responsible tester, such evidence must be investigated and probably raised as an issue to the client.

In short, in the situation given in the exercise, I would not answer the question, but rather raise some questions.

I drive telephone surveyors nuts in this way. They just don't know what to do with a guy who answers "no opinion" or "I don't know" or "can't answer" to every single question in their poorly worded and context-non-specific questionnaires.

[James Annan]

Robert Aumann's Agreement Theorem shows that honest Bayesians cannot agree to disagree - if they have common knowledge of their probability estimates, they have the same probability estimate.

Um, doesn't this also depend on them having common priors?

James

[Vladimir Nesov]

It feels like there was no explicit rule not to ask questions. It's interesting what percentage of subjects actually questioned the process.

If people are conforming rationally, then the opinion of 15 other subjects should be substantially stronger evidence than the opinion of 3 other subjects.

I don't see how moderate number of other wrong-answering subjects should influence decision of rational subject, even if it's strictly speaking stronger evidence, as uncertainty in your own sanity should be much lower than probability of alternative explanations for wrong answers of other subjects.

[Paul Crowley]

The video notes that when the subject is instructed to write their answers, conformity drops enormously. That suggests we can set aside the hypothesis that they conform for the rational reason you set out.

[anonymous]

90% of drivers can be better than the average.

[Chris]

'This may come as some surprise' to Asch & Aumann, but rationality is not the design point of the human brain (otherwise this blog would have no reason to exist), getting by in the real world is. And getting by in the real world involved, for our ancestors through tens of millenia, group belonging, hence group conformity. See J. Harris, 'No Two Alike', Chaps. 8 & 9 for a discussion which references the Asch work. This does not mean of course that group conformity was the only adaptation factor. Being right and being 'in' both had (and have...) fitness value, and it's pefectly natural that both tendencies exist, in tension.

[Steve Shervais]

At an applied level, this reminds me of Dr. Jerry B. Harvey’s discussion of the "Abilene Paradox" in management, where groupthink can take over and move an organization in a direction that no-one really wants to go. All it takes is one dissenter to break the spell.

[Recovering irrationalist]

Surely there's more than social conformity/conflict aversion at work here? In the experiment in the video, an expectation of pattern continuation is set up. For most questions, the 4 spoken words the subject hears before responding do correspond to the apparently correct spoken word response. I'd expect subconcious processes to start interpreting this as an indicator of the correct answer regardless of social effects and be influenced accordingly, at least enough to cause confusion which would then increase susceptibility to the social effects.

I'd expect this effect to also be reduced where the subject is writing down his answers, as that takes out of the equation the close connection between hearing spoken numbers and speaking spoken numbers.

[Caledonian]

Aumann's Agreement Theorem was proved more than twenty years after Asch's experiments, but it only formalizes and strengthens an intuitively obvious point - other people's beliefs are often legitimate evidence.

No, other people's beliefs are often treated as evidence, and very powerful evidence at that.

Belief is not suitable as any kind of evidence when more-direct evidence is available, yet people tend to reject direct evidence in order to conform with the beliefs of others.

The human goal usually isn't to produce justified predictions of likelihood, but to ingratiate ourselves with others in our social group.

What are you attempting to do, Eliezer?

[Stuart Buck]

FYI, if you look at Asch's 1955 Scientific American article, the lines on the cards were a little closer in length than in the example shown above.

[Steve]

my vision is so bad that i answered 'none of the above'. i had to decide to measure the lines. that meant i first had to get to where i did not think the trick was the question. that took a cup of tea.
'trust the ruler, not the vision' has been added to my list of -ings.

[Nominull]

Isn't it reasonable to find it more likely that people are lying than that something has gone that flagrantly wrong with my ability to judge sizes of lines?

[Unknown Healer]

"Belief is not suitable as any kind of evidence when more-direct evidence is available, yet people tend to reject direct evidence in order to conform with the beliefs of others."

Caledonian, this is just wrong. Our ability to interpret evidence is not infallible, and is often fallible in ways that are not perfectly correlated across individuals. So even if we share the same 'direct evidence' as other observers of equaly ability their beliefs are still relevant.

[Psy-Kosh]

Except we'd have to take into account the idea that the others who's beliefs we are using as evidence may themselves have been using the same idea... That results weighting of the beleifs of an initial group being greatly amplified above and beyond what it should be, no?

[Sebastian Hagen]

Robert Aumann's Agreement Theorem shows that honest Bayesians cannot agree to disagree - if they have common knowledge of their probability estimates, they have the same probability estimate.

In addition to what James Annan said, they also both have to know (with very high confidence) that they are in fact honest bayesians. Both sides being honest isn't enough if either suspects the other of lying.

[Richard Kennaway]

In terms of individual rationality, I hope I would notice my own severe confusion and then assign >50% probability to the majority vote.

Noticing your own severe confusion should lead to investigating the reasons for the disagreement, not to immediately going along with the majority. Honest Bayesians cannot agree to agree either. They must go through the process of sharing their information, not just their conclusions.

[Dave]

What are the odds, given today's society, that a randomly selected group of people will include any honest Bayesians. Safer to assume that most of the group are either lying, self-deluded, confused, or have altered perceptions. Particularly so in a setting like a psychology experiment.

[Jason Brennan]

Check out this paper:

Gregory S. Berns, Jonathan Chappelow, Caroline F. Zink, Giuseppe Pagnoni, Megan E. Martin-Skurski, and Jim Richards, “Neurobiological Correlates of Social Conformity and Independence During Mental Rotation,” Biological Psychiatry 58 (2005), pp. 245-253.

It claims that the conformists can, under some conditions, actually come to see the world differently.

[Psy-Kosh]

Oh, one other thing. I know it's been brought up before, but as far as the agreement theorem, I don't feel I can safely use it. What I mean is that it seems I don't understand exactly when it can and cannot be used. Specifically, I know that there's something I'm missing here, some understanding because I don't know the correct way to resolve things like agreement theorem vs quantum suicide.

It's been discussed, but I haven't seen it resolved, so until I know exactly why agreement theorem does not apply there (or why the apparently straightforward (to me) way of computing the quantum suicide numbers is wrong), I'd personally be really hesitant to use the agreement theorem directly.

[Unknown]

Perhaps Eliezer or someone else can check the math, but according to my calculations, if you use Nick Bostrom's SSSI (Strong Self-Sampling Assumption), and make the reference class "observers after a quantum suicide experiment", then if the prior probability of quantum immortality is 1/2, after a quantum suicide experiment has been performed with the person surviving, both the outside observer and the person undergoing the risk of death should update the probability of quantum immortality to 4/7, so that they end up agreeing.

This seems odd, but it is based on the calculation that if the probability of quantum immortality is 1/2, then the probability of ending up being an observer watching the experiment is 17/24, while the probability of being an observer surviving the experiment is 7/24. How did I derive this? Well, if Quantum Immortality is true, then the probability of being an observer watching the experiment is 2/3, because one observer watches someone die, one observer watches someone survive, and one observer experiences survival. Likewise if QI is true, the probability of being an observer surviving the experiment is 1/3. On the other hand, if QI is false, the probability of being an observer watching the experiment is 3/4 (I will leave this derivation to the reader), while the probability of being an observer surviving the experiment is 1/4.

From this it is not difficult to derive the probabilities above, that the probability of being a watcher is 17/24, and the probability of being a survivor 7/24. If you apply Bayes's theorem to get the probability of QI given the fact of being a survivor, you will get 4/7. You will also get 4/7 if you update your probabilities both on the fact of being a watcher and on the fact of seeing a survivor. So the two end up agreeing.

Intuitive support for this is the fact that if a QI experiment were actually performed, and we consider the viewpoint of the one surviving 300 successive trials, he would certainly conclude that QI was true, and our intuitions say that the outside observers should admit that he's right.

[Unknown]

In the above calculation I forgot to mention that for simplicity I assumed that the experiment is such that one would normally have a 50% chance of survival. If this value is different, the values above would be different, but the fact of agreement would be the same (although there would also be the difficulty that a chance other than 50% is not easy to reconcile with a many-worlds theory anyway.)

[Nick Tarleton]

Quantum suicide vs. Aumann has been discussed a couple times before, and yes, it's very confusing.

Intuitive support for this is the fact that if a QI experiment were actually performed, and we consider the viewpoint of the one surviving 300 successive trials, he would certainly conclude that QI was true, and our intuitions say that the outside observers should admit that he's right.

My intuitions say outside observers should not update their estimates one bit, and I'm pretty sure this is correct, unless they should also increase their probability of MWI on making the equivalent observation of a coin coming up heads 300 times in a row.

(although there would also be the difficulty that a chance other than 50% is not easy to reconcile with a many-worlds theory anyway.)
http://www.hedweb.com/everett/everett.htm#probabilities
http://hanson.gmu.edu/mangledworlds.html

[steven]

IMHO quantum immortality and quantum suicide (unlike MWI) are nonsense, but I'm still trying to figure out a way to say this that convinces other people.

For probabilities in MWI I recommend the work of David Wallace.

[Unknown]

Nick, my argument didn't depend on intuition except for support; so it doesn't bother me if your intuition differs. What was your opinion of the argument (or did I simply omit too many of the details to judge)?

[Someone]

I think the most interesting question that arises from these experiments is what's the difference in personality between people who dissent and people who conform (aside from the obvious).

[Psy-Kosh]

Unknown: Hrm, hadn't thought of using the SSSI. Thanks. Ran through it myself by hand now, and it does seem to result in the experimenter and test subject agreeing.

However, it produces an... oddity. Specifically, if using the SSSI, then by my calculations, when one takes into account that the external observer and the test subject are not the only people in existance, the actual strength of evidence extractable from a single quantum suicide experiment would seem to be relatively weak. If the ratio of non test subjects to test subjects is N, and the probability of the subject surviving simply by the nature of the quantum experiment is R, the likelihood ratio is (1+N)/(R+N), (which both the test subject and the external observer would agree on).
Seeing a nonsurvival gives a MWI to ~ MWI likelihood ratio of N/(R+N).
At least, assuming I did the math right. :)

Anyways, so it looks like if SSSI is valid, quantum suicide doesn't actually give very strong evidence one way or the other at all, does it?

Hrm... I wonder if in principle it could be used to make estimates about the total population of the universe by doing it a bunch of times and then analyzing the ratios of observed results... *chuckles* May have just discovered the maddest way to do a census, well, ever.

[Psy-Kosh]

Actually, if considering the SSSA instead of just the SSA, one has to take into account all the observer-moments, past and future, right? So there well be, in addition to the specific observer moments of "immediately post experiment test subject (or not), experimenter, everyone else...", there'll be past and future versions theirof, and of other entities, so you'll have K1 total "others" (other observer-moments, that is) in a MW universe, and K2 << K1 "others" in a single world universe.

This'll make the calculation a bit more confusing.

[Sam]

"... then what are the odds that only you are the one who's right?"

If this is the reasoning for people choosing the same answer then surely it becomes a question of confidence rather than conformity?

Choosing the same answer as the group in your argument is because you aren't confident in your answer and are willing to defer to the majority answer. Not necessarily the same as conformity. By your own rationing you are going with the group because you think their answer is "better" not because you want to be part of the group. I know you can argue that that is just your rationale for conformity, but I feel that conformity is more about doubting something you are sure you know, to side with a group, rather than doubting something you think you might know.

I feel possibly a more accurate test (using this reasoning for conformity) would be to take a group and tell all the members individually that only they will know the right answer. Then give all bar one the same answer and one a different answer and see if they will conform with the group.

[Leeroy Jenkins]

I believe that the subjects were of those of a non-matured state, thus making them of a "childish" mind and not able to process the situation. The subjrects would simply say anything their peers would say or do. I am testing this experiment on my classmates. I am in the 10th grade and will respond back with the solution. I blieve that a matured mind would not give in so easily with a simple question. It is not the question at hand that is making the subjects say something completely incorrect, it is the group pressure and the maturity of the subjects. If a child's mind thinks he or she is to believe that of another subject, then it shall think of that at hand. Children's minds are so open and naive thatt they will believe something as simple as Santa Clause comming down the chimney every year, then they will not hesitate to think of an answer to the question of this experiment. It is a simple and most uneducated experiment I had to present and test. A matured mind will think not of the group pressure but that of the question. I will be back with my results. Thank you.

Leeroy Jenkins