(Cross-posted from my blog.)

Since some belief-forming processes are more reliable than others, learning by what processes different beliefs were formed is for several reasons very useful. Firstly, if we learn that someone's belief that p (where p is a proposition such as "the cat is on the mat") was formed a reliable process, such as visual observation under ideal circumstances, we have reason to believe that p is probably true. Conversely, if we learn that the belief that p was formed by an unreliable process, such as motivated reasoning, we have no particular reason to believe that p is true (though it might be - by luck, as it were). Thus we can use knowledge about the process that gave rise to the belief that p to evaluate the chance that p is true.

Secondly, we can use knowledge about belief-forming processes in our search for knowledge. If we learn that some alleged expert's beliefs are more often than not caused by unreliable processes, we are better off looking for other sources of knowledge. Or, if we learn that the beliefs we acquire under certain circumstances - say under emotional stress - tend to be caused by unreliable processes such as wishful thinking, we should cease to acquire beliefs under those circumstances.

Thirdly, we can use knowledge about others' belief-forming processes to try to improve them. For instance, if it turns out that a famous scientist has used outdated methods to arrive at their experimental results, we can announce this publically. Such "shaming" can be a very effective means to scare people to use more reliable methods, and will typically not only have an effect on the shamed person, but also on others who learn about the case. (Obviously, shaming also has its disadvantages, but my impression is that it has played a very important historical role in the spreading of reliable scientific methods.)

A useful way of inferring by what process a set of beliefs was formed is by looking at its structure. This is a very general method, but in this post I will focus on how we can infer that a certain set of beliefs most probably was formed by (politically) motivated cognition. Another use is covered here and more will follow in future posts.

Let me give two examples. Firstly, suppose that we give American voters the following four questions:

1. Do expert scientists mostly agree that genetically modified foods are safe?
2. Do expert scientists mostly agree that radioactive wastes from nuclear power can be safely disposed of in deep underground storage facilities?
3. Do expert scientists mostly agree that global temperatures are rising due to human activities?
4. Do expert scientists mostly agree that the "intelligent design" theory is false?

The answer to all of these questions is "yes".* Now suppose that a disproportionate number of republicans answer "yes" to the first two questions, and "no" to the third and the fourth questions, and that a disproportionate number of democrats answer "no" to the first two questions, and "yes" to the third and the fourth questions. In the light of what we know about motivated cognition, these are very suspicious patterns or structures of beliefs, since that it is precisely the patterns we would expect them to arrive at given the hypothesis that they'll acquire whatever belief on empirical questions that suit their political preferences. Since no other plausibe hypothesis seem to be able to explain these patterns as well, this confirms this hypothesis. (Obviously, if we were to give the voters more questions and their answers would retain their one-sided structure, that would confirm the hypothesis even stronger.)

Secondly, consider a policy question - say minimum wages - on which a number of empirical claims have bearing. For instance, these empirical claims might be that minimum wages significantly decrease employers' demand for new workers, that they cause inflation and that they significantly reduce workers' tendency to use public services (since they now earn more). Suppose that there are five such claims which tell in favour of minimum wages and five that tell against them, and that you think that each of them has a roughly 50 % chance of being true. Also, suppose that they are probabilistically independent of each other, so that learning that one of them is true does not affect the probabilities of the other claims.

Now suppose that in a debate, all proponents of minimum wages defend all of the claims that tell in favour of minimum wages, and reject all of the claims that tell against them, and vice versa for the opponents of minimum wages. Now this is a very surprising pattern. It might of course be that one side is right across the board, but given your prior probability distribution (that the claims are independent and have a 50 % probability of being true) a more reasonable interpretation of the striking degree of coherence within both sides is, according to your lights, that they are both biased; that they are both using motivated cognition. (See also this post for more on this line of reasoning.)

The difference between the first and the second case is that in the former, your hypothesis that the test-takers are biased is based on the fact that they are provably wrong on certain questions, whereas in the second case, you cannot point to any issue where any of the sides is provably wrong. However, the patterns of their claims are so improbable given the hypothesis that they have reviewed the evidence impartially, and so likely given the hypothesis of bias, that they nevertheless strongly confirms the latter. What they are saying is simply "too good to be true".

These kinds of arguments, in which you infer a belief-forming process from a structure of beliefs (i.e you reverse engineer the beliefs), have of course always been used. (A salient example is Marxist interpretations of "bourgeois" belief structures, which, Marx argued, supported their material interests to a suspiciously high degree.) Recent years have, however, seen a number of developments that should make them less speculative and more reliable and useful.

Firstly, psychological research such as Tversky and Kahneman's has given us a much better picture of the mechanisms by which we acquire beliefs. Experiments have shown that we fall prey to an astonishing list of biases and identified which circumstances that are most likely to trigger them. 

Secondly, a much greater portion of our behaviour is now being recorded, especially on the Internet (where we spend an increasing share of our time). This obviously makes it much easier to spot suspicious patterns of beliefs.

Thirdly, our algorithms for analyzing behaviour are quickly improving. FiveLabs recently launched a tool that analyzes your big five personality traits on the basis of your Facebook posts. Granted, this tool does not seem completely accurate, and inferring bias promises to be a harder task (since the correlations are more complicated than that between usage of exclamation marks and extraversion, or that betwen using words such as "nightmare" and "sick of" and neuroticism). Nevertheless, better algorithms and more computer power will take us in the right direction.

In my view, there is thus a large untapped potential to infer bias from the structure of people's beliefs, which in turn would be inferred from their online behaviour. In coming posts, I intend to flesh out my ideas on this in some more details. Any comments are welcome and might be incorporated in future posts.

* The second and the third questions are taken from a paper by Dan Kahan et al, which refers to the US National Academy of Sciences (NAS) assessment of expert scientists' views on these questions. Their study shows that many conservatives don't believe that experts agree on climate change, whereas a fair number of liberals think experts don't agree that nuclear storage is safe, confirming the hypothesis that people let their political preferences influence their empirical beliefs. The assessment of expert consensus on the first and fourth question are taken from Wikipedia.

Asking people what they think about the expert consensus on these issues, rather than about the issues themselves, is good idea, since it's much easier to come to an agreement on what the true answer is on the former sort of question. (Of course, you can deny that professors from prestigious universities count as expert scientists, but that would be a quite extreme position that few people hold.) 

In Policy Debates Should Not Appear One-Sided, Eliezer Yudkowsky argues that arguments on questions of fact should be one-sided, whereas arguments on policy questions should not:

On questions of simple fact (for example, whether Earthly life arose by natural selection) there's a legitimate expectation that the argument should be a one-sided battle; the facts themselves are either one way or another, and the so-called "balance of evidence" should reflect this.  Indeed, under the Bayesian definition of evidence, "strong evidence" is just that sort of evidence which we only expect to find on one side of an argument.

But there is no reason for complex actions with many consequences to exhibit this onesidedness property.

The reason for this is primarily that natural selection has caused all sorts of observable phenomena. With a bit of ingenuity, we can infer that natural selection has caused them, and hence they become evidence for natural selection. The evidence for natural selection thus has a common cause, which means that we should expect the argument to be one-sided.

In contrast, even if a certain policy, say lower taxes, is the right one, the rightness of this policy does not cause its evidence (or the arguments for this policy, which is a more natural expression), the way natural selection causes its evidence. Hence there is no common cause of all of the valid arguments of relevance for the rightness of this policy, and hence no reason to expect that all of the valid arguments should support lower taxes. If someone nevertheless believes this, the best explanation of their belief is that they suffer from some cognitive bias such as the affect heuristic.

(In passing, I might mention that I think that the fact that moral debates are not one-sided indicates that moral realism is false, since if moral realism were true, moral facts should provide us with one-sided evidence on moral questions, just like natural selection provides us with one-sided evidence on the question how Earthly life arose. This argument is similar to, but distinct from, Mackie's argument from relativity.)

Now consider another kind of factual issues: multiple factor explanations. These are explanations which refer to a number of factors to explain a certain phenomenon. For instance, in his book Guns, Germs and Steel, Jared Diamond explains the fact that agriculture first arose in the Fertile Crescent by reference to no less than eight factors. I'll just list these factors briefly without going into the details of how they contributed to the rise of agriculture. The Fertile Crescent had, according to Diamond (ch. 8):

1. big seeded plants, which were
2. abundant and occurring in large stands whose value was obvious,
3. and which were to a large degree hermaphroditic "selfers".
4. It had a higher percentage of annual plants than other Mediterreanean climate zones
5. It had higher diversity of species than other Mediterreanean climate zones.
6. It has a higher range of elevations than other Mediterrenean climate zones
7. It had a great number of domesticable big mammals.
8. The hunter-gatherer life style was not that appealing in the Fertile Crescent

(Note that all of these factors have to do with geographical, botanical and zoological facts, rather than with facts about the humans themselves. Diamond's goal is to prove that agriculture arose in Eurasia due to geographical luck rather than because Eurasians are biologically superior to other humans.)

Diamond does not mention any mechanism that would make it less likely for agriculture to arise in the Fertile Crescent. Hence the score of pro-agriculture vs anti-agriculture factors in the Fertile Crescent is 8-0. Meanwhile no other area in the world has nearly as many advantages. Diamond does not provide us with a definite list of how other areas of the world fared but no non-Eurasian alternative seem to score better than about 5-3 (he is primarily interested in comparing Eurasia with other parts of the world).

Now suppose that we didn't know anything about the rise of agriculture, but that we knew that there were eight factors which could influence it. Since these factors would not be caused by the fact that agriculture first arose in the Fertile Crescent, the way the evidence for natural selection is caused by the natural selection, there would be no reason to believe that these factors were on average positively probabilistically dependent of each other. Under these conditions, one area having all the advantages and the next best lacking three of them is a highly surprising distribution of advantages. On the other hand, this is precisely the pattern that we would expect given the hypothesis that Diamond suffers from confirmation bias or another related bias. His theory is "too good to be true" and which lends support to the hypothesis that he is biased.

In this particular case, some of the factors Diamond lists presumably are positively dependent on each other. Now suppose that someone argues that all of the factors are in fact strongly positively dependent on each other, so that it is not very surprising that they all co-occur. This only pushes the problem back, however, because now we want an explanation of a) what the common cause of all of these dependencies is (it being very improbable that they all would correlate in the absence of such a common cause) and b) how it could be that this common cause increases the probability of the hypothesis via eight independent mechanisms, and doesn't decrease it via any mechanism. (This argument is complicated and I'd be happy on any input concerning it.)

Single-factor historical explanations are often criticized as being too "simplistic" whereas multiple factor explanations are standardly seen as more nuanced. Many such explanations are, however, one-sided in the way Diamond's explanation is, which indicates bias and dogmatism rather than nuance. (Another salient example I'm presently studying is taken from Steven Pinker's The Better Angels of Our Nature. I can provide you with the details on demand.*) We should be much better at detecting this kind of bias, since it for the most part goes unnoticed at present.

Generally, the sort of "too good to be true"-arguments to infer bias discussed here are strongly under-utilized. As our knowledge of the systematic and predictable ways our thought goes wrong increase, it becomes easier to infer bias from the structure or pattern of people's arguments, statements and beliefs. What we need is to explicate clearly, preferably using probability theory or other formal methods, what factors are relevant for deciding whether some pattern of arguments, statements or beliefs most likely is the result of biased thought-processes. I'm presently doing research on this and would be happy to discuss these questions in detail, either publicly or via pm.

*Edit: Pinker's argument. Pinker's goal is to explain why violence has declined throughout history. He lists the following five factors in the last chapter:

• The Leviathan (the increasing influence of the government)
• Gentle commerce (more trade leads to less violence)
• Feminization
• The expanding (moral) circle
• The escalator of reason