The Bias That Divides Us (2021) is about myside bias, the supposed bias whereby we generate and test hypotheses and evaluate evidence in a way that is biased toward our own prior beliefs. Myside bias prevents convergence in beliefs: if people evaluate evidence divergently, due to divergent prior beliefs, then two agents faced with exactly the same evidence may move further apart. Keith Stanovich thinks that this kind of divergence explains partisan polarization and prevents resolution of important political issues. He also thinks that myside bias is irrational. He offers solutions to the problems he diagnoses: we should set our priors at 0.5 in the kinds of circumstances in which myside bias is irrational…. [please read below the rest of the article].
Levy, Neil. 2021. “Is Myside Bias Irrational? A Biased Review of The Bias that Divides Us.” Social Epistemology Review and Reply Collective 10 (10): 31-38. https://wp.me/p1Bfg0-6d6.
🔹 The PDF of the article gives specific page numbers.
The Bias that Divides Us: The Science and Politics of Myside Thinking
Keith E. Stanovich
The MIT Press, 2021
These proposals are difficult to implement, as Stanovich no doubt would agree. Unsurprisingly, the book is shot through with his own biases. The last chapter, where the book is most overtly political, falls well below the standard of the rest: it is credulous toward crude take downs of contemporary ‘woke’ thought and to culture wars polemic. In other words, it exhibits the errors that can arise when someone gathers and evaluates evidence in ways that favor their own prior (political) beliefs. But the book is one worth taking very seriously, even if parts of it are not. The discussion of myside bias brings to it all the sophistication and intelligence that Stanovich has displayed throughout a long and influential career in psychology and the arguments and the conclusions it reaches deserve a wider hearing.
Myside bias is an especially important bias for us to consider, Stanovich argues. We, the readers of a forum like this, belong to what Stanovich calls the “cognitive elites.” Most of us are very highly educated and we tend to score well above median on tests of intelligence. There’s good news for us elites: most biases correlate negatively with measures of cognitive functioning. Cognitive elites tend to do better on tests of belief bias (the bias whereby the validity of an argument is influenced by the truth of its conclusion). Similarly, cognitive elites display less anchoring bias and less hindsight bias. But cognitive elites do not display less myside bias (on some studies, cognitive elites actually display more myside bias). Myside bias is not only a very important bias, Stanovich argues, it is also a bias which rages at full strength in the readers of his book.
Stanovich stipulates he will use ‘myside bias’ to refer to a psychological disinclination to abandon a favored hypothesis. He distinguishes this from confirmation bias, which is often conflated with or held to be identical to myside bias. As he uses the term, the confirmation bias is a testing strategy: someone with that bias looks for evidence supportive of a focal hypothesis. As Stanovich points out, the confirmation bias (so understood) can be perfectly rational. There’s nothing wrong with looking for confirming evidence. It’s only when the confirmation bias leads to or is combined with myside bias that we see departures from rationality: that is, when we no longer deal appropriately with disconfirming evidence. Or so Stanovich suggests.
But is he right? Is myside bias genuinely irrational? It’s hard to shake the idea that something is going wrong when people diverge in the face of one and the same body of evidence. A set of evidence can’t support p and ~p at one and the same time, it seems, yet such divergences occur time and time again in the lab, and seem ubiquitous outside it as well. As Stanovich shows, however, it’s surprisingly difficult to identify any rational failing on the part of any of the agents in many of the parade ground examples of divergent belief updating.
The key to seeing how divergent beliefs update on the same set of evidence may be rational is to apply Bayesian thinking. When an agent gets unexpected evidence that apparently disconfirms their prior belief, they must update their credences. But such updating is sensitive not only to their estimate of how likely the belief is, given the evidence, but also how likely the evidence is, given the belief. They may have stronger reason to regard the evidence as highly unreliable given their confidence in the belief than to appreciably lower their confidence in it. All by itself, this mechanism can explain why two agents may rationally move further apart in the face of the same evidence. If the agents have widely divergent credences in a hypothesis, then the one with lower confidence may move much further in response to apparently disconfirming evidence than the one with higher confidence. Many iterations of such occurrences can leave the party with widely divergent beliefs.
Moreover, such incidents can lead to evidence being preempted (see Begby, 2020) on future occasions. If a particular class of agents (say scientists, or Democrats, or opinion writers for Breitbart) regularly offer you evidence that you have good reason to think is unreliable in light of your priors, you may come to expect such (apparent) evidence from them. Expected evidence gives you no reason at all to update your beliefs. Accordingly, different groups of agents may rationally respond very differently to one and the same set of evidence. One agent may see the evidence as a strong reason to lower their confidence that p, another as a strong reason to lower their confidence that the source of the evidence is reliable and a third as presenting no reason to change their beliefs at all.
It follows, Stanovich points out, that myside processing can be rational. It can be rational to doubt the veracity of evidence rather than to update one’s credence in the focal hypothesis. He suggests that myside bias isn’t rational when probabilities are presented numerically, but that’s too quick. Participants in experiments may (rationally) discount the information experimenters provide them with. In such cases, they may not actually be asking the answering the question asked of them (“which conclusion does this data support”) but rather assessing the conclusion in the light of the plausibility, to them, of the evidence.
In any case, Stanovich recognizes that such experiments, with probabilities that are supposed to be accepted by all parties, are not good models for partisan polarization in the world outside the laboratory. Of course, Fox viewers and MSNBC viewers don’t accept the same data (about vaccine efficacy, say), so we can’t use these experiments to model their belief updating.
Nevertheless, Stanovich insists that myside bias is often irrational. Myside biased processing is rational only when the agent came by their priors honestly. In such cases, it’s rational to engage in what he calls “knowledge projection”; that is, in assessing new evidence in a way that is sensitive to your credences. But, too often, we don’t come by our priors honestly. They don’t arise from genuine evidence but rather reflect our ‘worldview’ or ‘convictions’. In such a case, knowledge projection is not rational, and such updating is not justifiable.
Honest Priors and Convictions
Stanovich never clearly spells out the distinction between honest priors and convictions. He seems to treat it as identical to another distinction he borrows from Abelson (1986), between testable and distal beliefs. The idea, roughly, is that testable beliefs are those that have arisen from evidence for or against them whereas distal can “neither be directly verified by experience, nor can they easily be confirmed by turning to evidence or scientific consensus” (8). In the latter class, Stanovich appears to lump all normative beliefs. Philosophers will of course be quick to point out that scientific theorising is itself shot through with normative assumptions. For instance, there is no normatively neutral way to balance the risk of false negatives versus that of false positives, so even when designing experiments or assessing evidence for some non-normative issue, we need to make decisions about which risk matters more, and this is an inherently normative task (Douglas, 2000).
For this reason (but not for this reason alone), Stanovich’s claim that myside processing is rational when it reflects priors come by honestly seems vulnerable to a regress argument. How did I come to acquire my current credence that p? It may be true that it reflects many episodes of updating on evidence in the kind of way he approves of, but my initial credence was not a function of evidence. Rather, it is a function of some combination of developmentally canalized expectations and the social context which was formative for me.
Perhaps Stanovich would accept this point and distinguish between testable and distal beliefs on the basis of (often multitudinous) further episodes of updating that have occurred since this first shaping. It’s reasonable to believe that for a wide range of initial priors, we will approach a credence that matches reality given enough (good enough) evidence, so we needn’t worry too much about how our initial priors are set. Our testable beliefs will come to reflect good evidence, but our convictions may float free, insensitive to such evidence.
Stanovich situates the distinction between convictions and testable beliefs within a broader theory of what he calls memeplexes. Memes are selfish, in the same way that genes are selfish: the properties and conditions that favor their replication may dissociate from their hosts’ interests. Memeplexes are sets of memes that have an immune system: they are hostile to new credences that might conflict with them. I found this idea intriguing but both unnecessary and wholly unpersuasive. It is unnecessary because it doesn’t seem to do any explanatory work that the distinction between testable and distal beliefs doesn’t already do. Unpersuasive because Stanovich doesn’t provide, and I am unable to imagine, a mechanism whereby memeplexes with the capacity to recognize and fight off undesirable credences could develop (the only mechanism would seem to be myside bias itself, but insofar as memeplexes are supposed to underlie myside bias or explain its properties, we can’t invoke it without circularity).
Perhaps the true value of the memeplex idea for Stanovich isn’t explanatory. He believes that we can best avoid myside bias, in its irrational form, by distancing ourselves from our convictions, and thinking of them as memes that serve their own interests and not ours might provide the necessary distance. When we recognize we’re dealing with a conviction and not a testable belief, we should set our prior probability at 0.5 and evaluate new evidence accordingly.
Science vs. Social Learning
The core of Stanovich’s view, then, is a distinction between credences which we formed on the basis of reliable evidence, either evidence we’ve gathered or the testimony of a scientific community, and those which we have not “thought our way to”; the distal beliefs we hold “largely as a function of our social learning within the valued groups to which we belong and our innate propensities to be attracted by certain types of ideas” (94). A great deal rests, therefore, on the claim that we have not “thought our way” to these distal convictions. Why should we believe that?
One reason Stanovich gives for thinking that we do not think our way to our convictions is that they are rarely things we have “consciously thought through and made an intentional decision to believe” (87). This, he suggests, is the ordinary view of our convictions: they are beliefs we hold as the upshot of conscious reflection and decision. I’m sceptical that this is the folk view: ordinary people are well aware that nonconscious cognition is common, and the thesis that beliefs cannot be voluntarily willed seems quite intuitive. In any case, there’s no reason for us, who are well aware that beliefs are often formed as the upshot of “largely unconscious social learning” (87), sometimes on the basis of developmentally canalized priors, to conclude that they are therefore irrational. For all that’s been said, largely unconscious social learning might lead to credences just as accurate as those shaped by conscious reflection.
In fact, I suggest this is indeed the case. While Stanovich recognizes the importance of testimony (from scientists and other experts), his focus is largely on first-order evidence. We ought to form beliefs for ourselves, he suggests, not defer to the consensus. Much of the last chapter is devoted to exhorting us cognitive elites to be inconsistent, in the way in which non-elite people are. That is, rather than taking the lead of political actors and adopting the views they espouse across a range of (often very disparate) issues, we ought to follow the example of those who pay much less attention to political cues and adopt policy positions that mix and match conservative and liberal strands. We might agree with Democrats on climate change, but that’s no reason to follow their lead on gun control and the minimum wage and mandatory vaccines and affirmative action. Those people who identify strongly with political parties and are also political junkies defer across the board, but that’s irrational (Stanovich claims). Since partisan views reflect political necessity and jostling for votes, we do better either to follow the evidence for ourselves, or when we can’t gather it to set our priors at 0.5.
However, there are good reasons to think social referencing—taking epistemic cues from people who are not epistemic authorities, as well as those who are—is rational. There is plentiful evidence we defer to prestigious individuals (Chudek et al. 2012; Henrich and Gil-White 2001), to consensus or majority opinion and to individuals we identify with (Harris 2012; Levy 2019; Sperber et al. 2010). We defer in these kinds of ways because the opinions of others provide us with genuine evidence.
The Epistemic Significance of Disagreement
We might best be able to bring this out through a consideration of the literature on the epistemic significance of disagreement (Christensen and Lackey 2013; Matheson 2015). The recognition that your epistemic peer disagrees with you about some moderately difficult issue provides you with a reason to lower your confidence in your belief. The disagreement is evidence that at least one of you has made a mistake, and you have no reason to think that it’s more likely that they’re mistaken than you are. Conciliationism—the thesis that we ought to reduce our confidence in the face of peer disagreement—has some seemingly unpalatable consequences, and philosophers have often tried to resist these implications by defining ‘epistemic peer’ narrowly. These philosophers accept that we should lower our confidence in peer disagreement cases, but argue that each of us has few epistemic peers, because few agents have precisely the same evidence and reasoning skills as you do (or, on another account, who are equally likely to be correct about the disputed issue, setting it and the issues it implicates aside).
But this move is unsuccessful, because non-peer disagreement also provides us with some reason to conciliate. Suppose I add up a row of numbers and come to the conclusion they total 945. The discovery that the 28 individuals who have also attempted the sum have each independently come to the conclusion the total is 944 places me under rational pressure to conciliate, even if I am a maths whiz and they are all 5th graders. The numbers count, which indicates that each individual instance of dissent by a non-peer often provides me with some evidence (in Bayesian terms, I should reduce my confidence in my belief to some, perhaps small, degree even in one person cases, because I am not fully confident in my conclusion and my credence that this person is unreliable on problems like this is not sufficiently high). Conversely, peer and non-peer agreement also provides me with evidence (if those 28 5th graders each came to the same conclusion as me, my confidence in my response should rise).
Of course, there are many complications to consider in assessing whether and how much non-peer disagreement provides us with evidence. Is the question one on which there is a body of specialized expertise that makes individuals who possess it significantly more reliable than those who do not? If so, does dissent or agreement stem from such an expert? If the person is a non-expert, to what degree is the question one on which non-experts are more likely than chance to be correct? Was their opinion reached independently of others? If not, are they at least somewhat discerning in echoing opinions (see Coady 2006 and Goldman 2001 for differing perspectives on this question)? Are they embedded in testimonial networks that contain or are anchored in expert opinion? I don’t have the space to discuss these many complications here. Nevertheless, it is very plausible that when political elites come to a view on an opinion outside the sphere of my expertise, I do better to defer to them than either to research the matter on my own or to be agnostic. These elites share my broad values, and therefore are likely to filter and weigh expert testimony in the sorts of ways in which I would, were I knowledgeable about the sources and implications of that testimony. They are likely plugged into testimonial networks that include or are sensitive to genuine expertise. Their public expression of their view exposes it to potential dissent by my fellow partisans, such that a lack of dissent indicates its plausibility. In adopting their view on the matter, I’m responding rationally to evidence.
In response, Stanovich will surely and plausibly insist that this kind of deference must often lead us astray. He points out that political parties yoke together a variety of disparate and sometimes conflicting currents, such that those who are attracted to a party because they share some of its values but not others (say they share the GOP’s high valuation of traditional institutions, but not the same party’s high valuation of free markets) would find themselves adopting positions contrary to their own values in deferring to party elites. He also points to the fact that it’s highly unlikely that such party elites have true beliefs across the board, given the extrinsic pressures under which they adopt positions. All of this is true, but it shows neither that we do better epistemically to be agnostic or do our own research, on the one hand, or—more pointedly—that we’re not being rational in deferring.
First accuracy. I can be confident that there are some issues on which my party is wrong and the opposition is right. But—since I’m an expert in so few of the topics on which they disagree—I have little to no sense which particular issues might be affected. On any particular issue, I do better—by my own lights—by adopting the party position than by being agnostic (I also do better, by my own lights, to defer than to do my own research, I believe: see Levy, forthcoming, for a defence).
In any case, we’re concerned here with the rationality of deference and not its accuracy. While rational cognition is non-accidentally linked to accurate cognition—the norms of rationality are justified because they are truth-conducive—rationality and accuracy often dissociate in particular cases. If my evidence is misleading, then rational belief update may lead me astray. We saw above how Bayesian reasoning may lead different agents rationally to diverge in response to one and the same set of evidence, after all. So pointing out that being guided by our convictions is guaranteed sometimes to lead us astray does not show that we’re irrational to be so guided. The opinions of party elites constitute genuine evidence for me, and it is rational for me to be guided by them. What goes for deference to party elites is true much more generally: we defer rationally when we engage in social referencing; that is, when our convictions are shaped by such cues (again, see Levy forthcoming).
We see the dissociation between accuracy and rationality in The Bias that Divides us Itself. As I’ve already noted, the final chapter contains little of value. A big part of the problem is that Stanovich defers to sources like James Lindsay and Peter Boghossian; sources that those who have genuine expertise in his bêtes noires, like “critical race theory,” know to be extremely low quality. In so deferring, he goes badly wrong. But it certainly doesn’t follow he’s guilty of irrationality. He may well be deferring appropriately, given his own priors.
Myside Bias, Irrationality, and Academic Psychology
In this opinionated review, I’ve focused on Stanovich’s claim that myside bias is often irrational. I’ve ignored many of his other claims. One of his aims in the book is to demonstrate the academic psychology is often inadvertently biased against conservatives, because it suffers from the same flaw that many of the experimental studies that purport to demonstrate irrationality in participants exhibit: confounding prior beliefs with processing errors (see Tappin and Gadsby 2019 for a discussion of how pervasive this error is in psychological research on bias and irratonality). Just as participants may give the ‘wrong’ answer not because they fail appropriately to process the evidence given but because they regard it as unreliable, so conservatives may show more resistance to change (say) than liberals because the items selected for a scale are those that conservatives don’t want to change and liberals do. Change the selection of items and it is liberals who become change resistant. He is convincing in suggesting there is no strong evidence that conservatives exhibit more psychological bias than liberals. He is somewhat less convincing in arguing that conservatives show no more out-group dislike than liberals (he is surely right that to some extent the data is confounded by the groups toward which attitudes are measured), and less convincing again in arguing there’s no strong evidence of more racism on the conservative side. Still, his points here are valuable and ought to be taken on board both by psychologists and those who consume their work.
There is much of value in The Bias that Divides Us. Stanovich succeeds to a considerable degree in demonstrating bias against conservatives in psychological research. His defence of the irrationality of myside bias is to my mind less convincing, but it is one that deserves a serious hearing. It’s a pity that the book exhibits the very phenomenon it decries, at least in the form it manifests here. If I’m right, though, in deferring to such unreliable sources Stanovich nevertheless behaves rationally.
Neil Levy, firstname.lastname@example.org, Macquarie University.
Abelson, Robert P., 1986. “Beliefs Are Like Possessions.” Journal for the Theory of Social Behaviour 16 (3): 223–250.
Begby, Endre. 2020. “Evidential Preemption.” Philosophy and Phenomenological Research 102 (3). https://doi.org/10.1111/phpr.12654
Christensen, David and Jennifer Lackey. 2013. The Epistemology of Disagreement: New Essays. Oxford University Press.
Chudek, Maciej, Sarah Heller, Susan Birch, Joseph Henrich 2012. “Prestige-Biased Cultural Learning: Bystander’s Differential Attention to Potential Models Influences Children’s Learning. Evolution and Human Behavior.” 33: 46–56.
Coady, David. 2006. “When Experts Disagree.” Episteme 3 (1-2): 68–79.
Douglas, Heather. 2000. “Inductive Risk and Values in Science.” Philosophy of Science 67 (4): 559–579.
Goldman, Alvin I., 2001. “Experts: Which Ones Should You Trust?” Philosophy and Phenomenological Research 63 (1): 85–110.
Harris, Paul L. 2012. Trusting What You’re Told. Harvard University Press.
Henrich, Joseph, and Francisco J Gil-White. 2001. “The Evolution of Prestige: Freely Conferred Deference as a Mechanism for Enhancing the Benefits of Cultural Transmission.” Evolution and Human Behavior 22 (3): 165-196.
Levy, Neil. forthcoming. Bad Beliefs: Why They Happen to Good People. Oxford University Press, Oxford.
Levy, Neil 2019. “Due Deference to Denialism: Explaining Ordinary People’s Rejection of Established Scientific Findings.” Synthese 196: 313–327.
Matheson, Jonathan. 2015. The Epistemic Significance of Disagreement. Palgrave Macmillan.
Sperber, Dan, Fabrice Clement, Christophe Heintz, Olivier Mascaro, Hugo Mercier, Gloria Origgi, and Deirdre Wilson. 2010. “Epistemic Vigilance.” Mind & Language 25 (4): 359–393.
Tappin, Ben M. and Stephen Gadsby. 2019. “Biased Belief in the Bayesian Brain: A Deeper Look at the Evidence.” Conscious Cogn 68: 107–114. doi: 10.1016/j.concog.2019.01.006.
Categories: Books and Book Reviews