Reliance on consensus

By consensus I mean extreme majority but not necessarily universal view within some group - as this seems to be its meaning as it is commonly used today.

Someone competent in a field does not have much use for consensus. A mathematician can examine a proof and discover for himself whether it is valid or invalid, without relying on the opinion of other mathematicians. A scientist who doubts a result can reproduce the experiment for himself, without relying on the opinion of other scientists. This is not to say that scientists do not benefit from communication with other scientists, that they do not benefit from a sanity check provided by other scientists. But consensus can be built up in many different ways, many of them more likely to reinforce error than to discover truth, only some of them an improvement on independent thinking. So consensus - be it general, scientific, field-specific, or what have you, is not by itself very good evidence of truth. As was pointed out in the book by that title, to get the "Wisdom of Crowds" phenomenon it's important that people give their independent opinions, thus avoiding an information cascade - a condition that is not all that often fulfilled.

If you rely on consensus you are making yourself into part of the above-mentioned information cascade. We might hope, therefore, that people who rely on the consensus in a field are themselves all outside that field. That is not certain to be the case, which poses a problem for those who would like to rely on consensus.

If you wish to rely on the consensus of the group of people competent to individually judge claims in a given topic for themselves, you also have the problem of defining that group. If you are not yourself competent to judge, how do you know who is competent to judge? If you rely on someone else's definition of that group, how do you know that they are competent to judge? Ultimately you cannot avoid making a decision for yourself - you cannot avoid choosing who to rely on, and therefore you cannot, in the end, avoid relying on your own competence to make the decision. (I happen to think that many people are not competent, and that consequently they rely on false authority - they are the blind led by the blind. So this ultimate reliance on one's own competence is a real problem and not just of academic interest.)

There are some ways for the competent to identify others to delegate judgment to. For example, if you spot check someone's work, and all that you've checked is flawless, then it is reasonable to rely on the rest of his work without checking on it, treating him as an authority. This, however, presupposes your own competence - your own authority. It does not work for the incompetent.

You don't have to have the same kind of competence as the person whose competence you are testing. I don't know many words of Chinese (Mandarin or other dialect) but I can identify authorities on Chinese. There are ways to "bootstrap" competence. But these employ a certain kind of competence. For instance, you need to have the competence to distinguish the cases where a specific kind of bootstrapping succeeds from the cases where it fails. (So the thoroughly incompetent are thoroughly screwed. The competent may, of course, guide them, but so may anyone else.) I happen to think that many people accept authorities without the benefit of proper "bootstrapping" - possibly as a result of tragically mistaken "bootstrapping" based on false signs of competence.

Once you have identified a group of authorities, then relying on consensus, rather than relying on (say) the judgment of selected individuals from this group, may serve the limited function of checking your own tendency to cherry-pick authorities whose judgments coincide with what you want to be true. And it might, additionally, confer a Wisdom of Crowds benefit. But if there is dissent within your identified group of authorities, that is a significant fact. And of course, relying on consensus (as opposed to relying on one specific authority) is only meaningful (will only even potentially give you a different conclusion than relying on a single authority) if there is dissent. So the very occasions on which reliance on consensus might even possibly confer some benefit, are precisely those occasions where the significant fact of the existence of dissent may give you pause.

Thus, the useful function of reliance on consensus is restricted (it checks your own tendency to cherry-pick), and its value is dubious (since what it amounts to is ignoring the existence of dissent among your chosen authorities). Furthermore it completely fails to distinguish convergence by independent experience and thought from convergence by groupthink, which you should be keenly interested in distinguishing.

There is some evidence that the average ______ is incompetent. The linked evidence (the rate of correct answers was worse than random guessing) is for economics but why should this not be generally true? I suspect it is (based in part on my own experiences here and there). So as a rule of thumb, between eighty and 100 percent of accredited members of a field are incompetent and should not be trusted.

Good luck relying on the consensus within a field (as defined by accreditation) if that's the case!

Underlying the popular habit of relying on consensus may lie the consensus theory of truth. Certain philosophical notions are these days deeply embedded in popular thought, and the consensus theory of truth may be one of them.

Share this

malicious fake consensus.

You are focusing on consensus going wrong spontaneously, through groupthink, conformity, or people feeding their shared biases back to each other and thus perceiving these shared biases as fact.

No doubt that happens a great deal, stupidity being considerably more common than evil - but evil is common enough. Other failures of group think are malicious. One is conspiratorial - a bunch of people meet in secret, agree upon a lie, and then these people supposedly independently report the same lie - thus, for all example, all the supposedly non communist and supposedly anti communist journalists and politicians confidently agreeing that the Moscow trials were clearly legitimate and true.

Another is coercive. You get global warming research results, or you don't get research money.

The recent financial crisis was called in advance by a small minority of people. They were ignored partly through group think, party because they were being "racist" - a false consensus was and is coercively imposed.


a bunch of people meet in secret, agree upon a lie, and then these people supposedly independently report the same lie

Hence for example the obvious suspicion about Ezra Klein's JournoList. Consensus within an impressively large group arrived at by a secret informational cascade is - by virtue of JournoList's secrecy - passed off as consensus arrived at by independent observation and thought, and thus by virtue of this fraud has greater influence than it legitimately ought to have.

It's curious that, prior to learning about JournoList, I had pegged Ezra Klein as an idiot and had wondered aloud at the curious frequency with which he is quoted, given that he is an idiot.

Golf Clap

Very nice write up.

Bootstrapping Competence

Do you have a theory for Bootstrapping Competence?

I can see a lot of applications to human learning, not to speak about Artificial Intelligence. Maybe the opportunity for a great website, too.

I've always thought that indeed there should be a case for bootstrapping confidence in experts through proper random testing of their claims -- the problem being in identifying degrees of "dependence" or "independence" between claims.

Private mail on the topic welcome.

I think we can

I think we can systematically bootstrap confidence with these three tools

- a semantic web
- reification of statements
- a model of trust

A semantic web gives a large amount of easily understandable bits of information. Any scientific paper can have a simple semantic representation representing its hypothesis, findings, etc.

Reification allows to tie statements to people. If X is a statement, (A says X) is a reified statements. Any semantic web must rely on reification in order to build confidence in statements

A model of trust is needed.

A simple model is to assume that people have an intrinsic probability of truth. One can then assign probability weights to a set of statements and use Bayes rule to build an estimate of other person's reliability. One can then use these probabilities to weight these people's statements about other people's reliability. A network of trust is thus created.

Another model could assume people are specialists, their probability of making true statements varies with the topic but remains correlated to a general reliability factor. These can be estimated with Bayesian magic.