In Bayesian Rationality (Oaksford & Chater Reference Oaksford and Chater2007, henceforth BR) Oaksford & Chater (O&C) present both positive arguments in support of a probabilistic approach and negative arguments against a non-probabilistic “deductive” approach to human reasoning. The negative arguments are comparative in nature and hinge on the presumed acceptability of a direct comparison between the theories' computational-level descriptions, without taking account of their algorithmic-level specifications. I argue that this central premise is fundamentally mistaken: It is potentially misguiding to compare and evaluate theories on the basis of just their computational-level descriptions.
Cognitive psychology is about explaining what people do. Ignoring the algorithmic level is only acceptable when there is a “triumphant cascade” through the levels of description (Dennett Reference Dennett1987, p. 227), that is, when there is a functional match between the input-output function computed at the computational level and the function computed at the algorithmic level. In these cases, there is no added explanatory value in going from the computational level to the algorithmic level and, thus, no loss in explanatory value by not considering the algorithmic level. Cognitive theories are, however, swamped with functional mismatches between the levels (Franks Reference Franks1995; Reference Franks1999). It follows that when one critiques the idealized computational-level description of such theories on the basis of their explanatory (in)adequacy, one creates a straw-man argument, that is, a negative argument aimed at an oversimplified straw-man version of a theory (Schroyens, in press).
One could try to justify a comparison at the computational-level descriptions of theories by assuming that they are normative. The underlying assumption is that a theory can be normatively justified on the basis of what people actually do: “the apparent mismatch between normative theories and reasoning behaviour suggests that the wrong normative theories may have been chosen; or that the normative theories may have been misapplied” (BR, p. 30). O&C take the poor explanatory adequacy of deductive logic to indicate it provides the wrong norm. (Of course, when one does not consider that the normative theories may have been misapplied, one has jumped to a conclusion). There are several problems with the “is-ought” argumentation in O&C's book.
Given that cognitive psychology is not concerned with how people should reason, the “is-ought” analysis in BR signifies that O&C are not engaged in cognitive psychology, but are making an epistemological, philosophical argument. As an argument about normative rationality, the “is-ought” analysis is fallacious. “This fallacy consists in assuming that because something is now the practice, it ought to be the practice. Conversely, it consists in assuming that because something is not now the practice, it ought not to be the practice” (Damer Reference Damer2005, p. 127). That is, one must be careful not to commit the “naturalistic fallacy” (to use G. E. Moore's term).
The “is-ought” analysis is also fallacious in its appeal to common opinion. “This fallacy consists in urging the acceptance of a position simply on the grounds that a large number of people accept it or in urging rejection of a position on the ground that very few people accept it” (Damer Reference Damer2005). Virtually all people were once ignorant about the currently accepted fact that the earth is a round celestial object that revolves around the sun; and even today there is a large segment of the world population that remains unenlightened with regard to this. Does this mean that Copernicus was wrong, and was rightly excommunicated by the Church? One can have a personal opinion about how people should reason, and I think one actually should have such an opinion. However, when one is considering the psychology of human reasoning, it is irrelevant whether one can label 5% or 95% of the participants as “rational” because their behavior corresponds to some norm. Whether 5% or 95% of participants' responses can be labeled “correct” is irrelevant to the task of describing and explaining what they are actually doing. The cognitive psychologist's task is first and foremost that of developing descriptively adequate processing models that allow us to understand how people reason. All performance has to be accounted for, whether logically correct or incorrect.
O&C give body to their “is-ought” ideology, by adopting Anderson's (1990) rational analysis approach. However, it is a mistake to think that the “the essence of rational analysis [Anderson 1990] is to view cognitive processes as approximating some normatively justified standard of correct performance” (Oaksford & Chater Reference Oaksford and Chater1998, p.18). As O&C stated more correctly elsewhere: “Anderson [1990] argues that we must distinguish normative from adaptive rationality” (Oaksford & Chater 1998, p. 174; see also, Chater et al. Reference Chater, Oaksford, Nakisa and Redington2003). Indeed, “rationality is being used in two senses and that rationality in the adaptive sense, which is used here, is not rationality in the normative sense that is used in the studies of decision making and social judgement” (Anderson Reference Anderson1990, p. 31). There is no unconditional “should”: there are no absolute normative standards of correct performance. This goes against the very notion of adaptive rationality, which is relative and “defined with respect to a specific environment” (Anderson Reference Anderson1990, p. 35).
O&C have fallen into the same conceptual trap many “mental logicians” fell into: a conceptual mistake – called “reasoning imperialism” by Rips (Reference Rips and Elio2002) – for which O&C, with much merit to their work, criticized such logicians. Bayesian Rationality provides prominent and valuable contributions to the literature that echoes Anderson's (1990) idea that deductively irrational behavior might not be adaptively irrational. That is, the book successfully questions the normative status of logic within the rational analysis approach of Anderson (Reference Anderson1990). However, instead of adopting and following the notion of adaptive rationality to its full extent, by questioning the feasibility of a single theory (whether probabilistic or logical) to have absolute normative status, they simply replace one normative theory with another. Adaptation is defined with respect to a specific environment and a true rational analysis accordingly takes account of this environment. It does not concern itself with absolute normative issues, because this denies the environmental relativity of adaptively rational behavior.
