There is much that I can admire in Bayesian Rationality (Oaksford & Chater Reference Oaksford and Chater2007, henceforth BR) and the lengthy research programme that it describes. There is also much that I agree with: for example, that human beings resemble Bayesian reasoners much more closely than ones engaging standard logic. I also share part of the authors' research programme, in that I have argued against logicism and excessive use of the deduction paradigm in the psychology of reasoning (Evans Reference Evans2002), as well as proposed that ordinary conditionals are much better modelled by probability logic than by standard propositional logic (Evans & Over Reference Evans and Over2004). So why is it that when I read the Socratic dialogue presented in Chapter 8 of BR, my sympathy is constantly with the sceptic?
The problems for me lie with “rational analysis” and the manner in which Oaksford & Chater (O&C) go about doing their science. Rarely do we see anything approaching a descriptive or algorithmic account of reasoning (the model of syllogistic reasoning being the closest) or even – in spite of their philosophy – a genuinely computational one. What we typically get is an alternative norms account. Thus, choices on the Wason selection task look erroneous from the viewpoint of standard logic, but can be seen as rational attempts to maximise information gain (given a misreading of the task instructions that makes it about categories instead of four individual cards). Similarly, we can predict people's conditional inference rates on the assumption that they are somehow reflecting the probability of the conclusion given the minor premise (despite the oddity that uncertainty resides exclusively in the major premise), and so on. The object of the exercise seems to be to show that there is some normative account of behaviour that can be offered. As a cognitive psychologist, I find this unsatisfying, because I want to know what people are actually doing and how.
The basic philosophy of the authors is set out in Chapter 2, on which I will focus. They start out by describing essentially the same paradox as the one that inspired our earlier book on rationality (Evans & Over Reference Evans and Over1996a). Why is it that people's reasoning and decision making seems essentially rational, intelligent, and adaptive in the real world, but results in numerous violations of logic and other standard normative systems when they are tested in the laboratory? We suggested that while people have some limited capacity for explicitly following normative rules (rationality2), they could often achieve everyday goals by implicit processes such as associative learning (rationality1). O&C (BR, p. 29) object, however, to any form of rationality that cannot be justified by a normative system. But elsewhere in the book, they seem to be describing something very similar to rationality1. For example, they describe some examples of “rational” behaviour with the comment: “Such behaviour may be built in by evolution or acquired via a long process of learning – but need not require on-line computation of the optimal solution” (BR, p. 36). Examples discussed here include Gigerenzer's research programme on fast and frugal heuristics (Gigerenzer Reference Gigerenzer, Koehler and Harvey2004). But Gigerenzer appeals to ecological and not normative rationality, so the link between the two programmes is unclear.
The authors overstate everyday rationality when they ask (BR, p. 30), “why do the cognitive processes underlying everyday rationality consistently work?” Well, they don't, and to say otherwise is to ignore a massive body of psychological literature. For example, in the study of judgement and decision making, many cognitive biases have been shown to occur reliably in everyday contexts and with expert judges: phenomena such as outcome bias, hindsight bias, overconfidence, and the planning fallacy, to name just a few (for evidence, see papers from two recent collections on these topics: Gilovich et al. Reference Gilovich, Griffin and Kahneman2002; Koehler & Harvey Reference Koehler and Harvey2004). There is also a massive accumulation of evidence for dual-process accounts of reasoning and decision making (Evans Reference Evans2007; Reference Evans2008), something which plays no visible role at all in the account of human reasoning offered by O&C. Nor do O&C feel moved to explain why the achievement of standard normative solutions to the task for which they provide alternative normative accounts is associated with those of high general intelligence (Stanovich Reference Stanovich1999). It is particularly disappointing that they make no reply to Stanovich's carefully constructed challenge to their position.
A final issue is whether their methodology provides computational or normative accounts of behaviour, because these are not the same thing (Elqayam Reference Elqayam, Vosiniadou, Kayser and Protopaps2008). The authors try to combine them as follows: “Formal rational principles spell out the optimal solution for the information processing problem that the agent faces. The assumption is that a well-adapted agent will approximate this solution to some degree” (BR, p. 38). But the adaptive mechanisms of nature – evolution and learning – do not optimise. If evolution had an optimal solution, we would not have billions of different species of every conceivable form and function. If learning was optimised, animals would not, for example, generally match the probability of their foraging to that of food sources. But O&C are quite explicit (BR, pp. 30–31) that they are not worried about the performance of a priori normative theories in providing such quasi-computational accounts. On the contrary, they are looking for empirical rationality, which means in essence: observe some behaviour, assume that it is rational, find a normative theory that deems it to be so, and then … nothing else, apparently.
I can understand the benefits of rational analysis when it is rooted in ecological rationality, so that by assuming that behaviour is adapted to the environment, we can look to the structure of that environment for cues as to how our cognitive mechanisms are designed. However, I really don't understand why O&C feel the need to fit a normatively rational model to human reasoning, and still less why that should put an end to the inquiry, with no algorithmic account to follow. Nevertheless, I agree strongly with them that the Bayesian model is a far more appropriate reference for real world reasoning than one based on truth-functional logic, and that it is a standard much more likely to be approximated in the inferences that people actually make.
