Reconsidering the normative argument from bounded rationality

The Normative Argument from “Bounded Rationality”

A number of scholars have tried to draw important normative conclusions from descriptive research on “bounded rationality.” For instance, Gigerenzer and Goldstein (1996) write that “Bayes’ rule and other rational algorithms quickly become complex and cognitively intractable, at least for ordinary human minds” (p. 651), that “unbounded rationality is a strange and demanding beast,” and that “cognitive algorithms need to meet more important constraints than internal consistency: they need to be psychologically plausible” (p. 665). In brief, according to what I am calling here the normative argument from bounded rationality, we should be careful in interpreting people’s failure to adhere to rational norms as evidence of their irrationality, since following such norms can be too demanding for cognitively limited agents. These considerations are often found in the literature, and the idea that people should not be charged with irrationality when failing to comply with overly demanding norms has attracted a lot of attention (e.g., Cherniak, 1986; Good, 1993; Harman, 1986). In light of this, it should not come as a surprise that authors like Keith Stanovich (1999) have actually presented these considerations as grounding one of the main positions on offer in the “rationality debate.” Specifically, Stanovich associated this view with the label “Apologist.”

A first important element that is worth highlighting is that the argument has been presented as an attack on pessimistic assessments of human rationality. Notably, according to Stanovich’s (1999) presentation of this view, “individuals should not be deemed irrational because of the large deviations of their responses from the normative.” In particular, the “Apologist takes seriously the caveat about computational limitations … and this view emphasizes that when human performance deviates from a normative model, irrationality cannot necessarily be imputed” (p. 7).

Moreover, the normative argument from bounded rationality is typically presented as an independent argument in the “rationality debate.” More specifically, Stanovich clearly distinguishes this concern from criticisms of the external validity of findings from bias research and defenses of the conduciveness to success of biased behavior. But Stanovich is not alone in offering this characterization. For instance, also Annika Wallin (2013), in a recent discussion, treats the demand that normative standards be psychologically plausible as independent of other arguments in the literature. In particular, she writes that:

I have of course, ignored the issue of frugality here, and merely focused on success, but I think that if it were the case that following coherence consistently led to success for real people and not merely for rational angels, the issue of frugality would become secondary. (p. 473)

What the quote suggests is that, also on Wallin’s view, the concern that normative standards are too demanding for actual cognizers has to be distinguished from considerations about whether following norms of rationality such as principles of coherence leads to successful behavior. The question that arises, however, is how exactly can considerations about people’s cognitive limitations lead to normative conclusions.

Bounded Rationality and the Ought-Implies-Can Principle

It is interesting to note that most discussions of the normative relevance of literature on bounded rationality seem to appeal to versions of the ought-implies-can principle. For instance, Shira Elqayam (2011) writes that:

Simon’s model of bounded rationality refers mainly to the way that human rationality is constrained by what Cherniak calls “the finitary predicament”; that is, physical and cognitive limitations on processing. The fact that humans do not live forever, that our brain capacity is finite, dictates that only tractable computations could count as rational. The underlying rationale is the age old “ought-implies-can” attributed to Kant. (p. 402)

Construed this way, the normative argument from bounded rationality is indeed an independent argument in the “rationality debate,” clearly distinguished from concerns about the adaptiveness of behavior departing from the “standard picture of rationality.” Specifically the argument would be resting on the ought-implies-can principle, which was first used in ethics, but has since been applied quite broadly. In this context, the principle states that people can be charged with irrationality only in cases where they could have done otherwise. While other authors do not explicitly refer to such principles, the principle seems to capture quite well the concerns that several commentators in the “rationality debate” seem to have in mind. For instance, in his discussion of the normative relevance of research on cognitive limitations, Stich (1990) has pointed out that “it seems simply perverse to judge that subjects are doing a bad job of reasoning because they are not using a strategy that requires a brain of the size of a blimp” (p. 27). Moreover, when Stanovich (2004) tries to spell out the position of the “Apologist” he writes that:

It seems perverse to call an action irrational when it falls short of optimality because the human brain lacks the computational resources to compute the most efficient response. Ascriptions of irrationality seem appropriate only when it was possible for the person to have done better. (p. 157)

Notably, the ought-implies-can principle is not uncontroversial (see Graham, 2011; Stern, 2004; Vranas, 2007 for some critical discussions of the principle). However, commentators in the “rationality debate” generally defend a version of this principle. But a remark is in order; consider that the plausibility of the principle seems to depend on what counts as exceedingly demanding. Specifically, following rules of rationality might be too demanding in the sense that it requires sacrifices and is laborious, or in the sense that we literally cannot follow them. Arguably, the second and stronger sense of demandingness seems to lend more robust grounding to the ought-implies-can principle. After all, while following norms of the “standard picture of rationality” might be too demanding, in the weaker sense of the term, the person who claims this should expect to be told that this is the nature of rationality, and that we have to accept its demandingness. Commentators on the “rationality debate” generally take the stronger version of the ought-implies-can principle to be plausible. At the same time, they also argue that the argument based on this principle does not work as a challenge to pessimistic strands in bias research.

Here I will not provide an exhaustive list of concerns that have been expressed about this view, but rather a selection. A first objection offered by Stein (1996) is that the abovementioned concerns about the demandingness of the “standard picture of rationality” do not apply to the tasks and problems examined in bias research. To see this, consider a person trying to follow the consistency preservation principle. We can follow Stein here, and characterize it along these lines: “Suppose we have the intuition that before a person commits herself to some belief p, she should check to make sure that p is logically compatible with all her other beliefs” (Stein, 1996, p. 163). Notably, this person would never have enough time to acquire new beliefs. The computational problem will emerge once a person begins to monitor and to check the consistency of her beliefs. For instance, consider a person with n beliefs. In order to check the pair-wise consistency of n beliefs, n(n-1)/2 pairs have to be compared. But is consistency computationally possible? Arguably, for mere mortals, being fully consistent in environments involving larger ns becomes computationally impossible. It thus seems that, at least in this case, it is literally impossible to conform to the “standard picture of rationality.” However, as Stein has convincingly argued (cf. 1996, p. 248), norms like the conjunction rule, which are often and extensively discussed in bias research, do seem to be reasonable candidates for the implementation in human brains in real time. In particular, according to the conjunction rule of probability theory, the probability of the conjunction of events cannot exceed the probabilities of the constituent events. On the face of it, it does not seem to be true that following norms like the one presented above is not “psychologically plausible.”

Moreover, as Stanovich’s research on individual differences in thinking and decision-making has revealed, at least some people seem to be able to follow those norms, and following standard norms of rationality thus does not seem to go beyond human cognitive capacities, or at least beyond the cognitive capacity of some of the reasoners. To be sure, one might reply that following those norms goes beyond the cognitive capacities of at least some people, though not of others. More precisely, one might insist that those people who fail to comply with the norms of the “standard picture of rationality” do so because of the computational limitations they face, which differ from those of people who do not violate them. However, there are two problems with this view. One problem seems to be the following. As Stanovich’s research seems to suggest, for some biases there is no direct correlation between cognitive capacity and susceptibility to bias, where standard tests of intelligence are often taken to provide an operationalization of cognitive and computational capacity. Thus, it might be difficult to explain differences in performance on the part of reasoners in terms of differences in cognitive capacities. In the words of Stanovich (2011), this strategy:

[w]ill not work for all of the irrational tendencies that have been uncovered in heuristics-and-biases [emphasis added] literature. This is because some of those biases are not very strongly related with measures of intelligence. (p. 357)

One might take issue with the operationalization of cognitive capacity offered by Stanovich. However, even if someone were unconvinced by this strategy, there would be a second and less controversial reply to applications of the use of the ought-implies-can principle in this context. Specifically, it would still be unclear whether people with limited computational power really could not do any better and bring themselves to follow the norms of the “standard picture of rationality.” As a number of theorists have tried to show, at least in some cases (e.g., in the case with the base rate fallacy) it seems possible to make a problem easier to compute by modifying the format of information (e.g., by using natural frequencies instead of single probabilities). If this is correct, it follows that a person’s computational power can be improved. Hence, it is unclear whether people with limited computational power really cannot conform to the norms of the standard picture of rationality once the right resources and conditions apply.

In light of this, it seems fair to follow most commentators in the debate and argue that this particular conceptualization of the normative argument from bounded rationality does not open promising avenues in the “rationality debate.” Specifically, for many important problems explored in the heuristics in biases literature, people do not seem to need unreasonable cognitive resources to find the normatively correct answer.

Bounded Rationality and Adaptive Rationality

One might want to conclude that the argument presented above is unconvincing, as it does not seem to be successfully applicable to findings from bias research. Still, one might also wonder whether what I presented above is the best or the only legitimate conceptualization of what we have called the normative argument from bounded rationality. As it turns out, there seems to be a different interpretation of the normative argument from bounded rationality. Instead of appealing to versions of the ought-implies-can principle, the second version of the argument seems to rest on what I will refer to as “adaptive rationality,” namely the idea that what matters for the assessment of behavior is whether people’s thinking and decision-making lead to the attainment of their epistemic and prudential goals. More precisely, I characterize the framework of “adaptive rationality” along the following lines: the rationality of an agent’s behavior in a particular situation depends on the extent to which the choice method in use is goal-conducive, in the sense that the strategy that she actually uses would help her achieve her prudential and epistemic goals. Adaptive rationality thus departs from standard views on rationality by emphasizing the importance of assessing biological organisms against their goals rather than formal universal principles, and by broadening the range of relevant goals. Specifically, as Gigerenzer (2008) has pointed out, the goals that are considered by the “standard picture of rationality” are too narrow. According to him,

Logic focuses on truth preservation. Consequently, mental logic, mental models …, and other logic-inspired systems investigate cognition in terms of its ability to solve syllogisms, maintain consistency between beliefs, and follow truth table logic. … Probability theory depicts the mind as solving a broader set of goals, performing inductive rather than deductive inference, dealing with samples of information involving error rather than full information that is error-free, and making risky “bets” on the world rather than deducing true consequences from assumptions. (p. 20)

As adaptive rationality theorists correctly emphasize, people also have further goals other than, say, consistency and truth perseveration. For instance, as Gigerenzer and Gaissmaier (2011) write, in the social domains goals “include transparency, group loyalty, and accountability” (p. 471; see Hertwig, Hoffrage, & the ABC Research Group, 2013 for a thorough treatment of people’s social and adaptive rationality), and arguably researchers in the heuristics and biases tradition have often and unduly neglected the social dimension of human rationality. Notably, the importance of considering social norms and goals when assessing behavior has been stressed also by authors outside the Centre for Adaptive Behaviour and Cognition and Centre for Adaptive Rationality (e.g., Mercier & Sperber, 2011; Sen, 1993).

To see that the points above are not at all untethered, let us focus on the conjunction fallacy (Tversky & Kahneman, 1983). Given the story of Linda, a person who took part in antinuclear demonstrations, majored in Philosophy, and some other activities, people judge of that person that it is more probable that she should be a bank teller and active in the feminist movement, than it is that she should be a bank teller. This phenomenon is usually interpreted as an indication of irrationality, because it violates the conjunction rule of probability theory, which states that the probability of a conjunction is always smaller than or equal to the probability of one of its conjuncts. But can this rule be used to measure people’s rationality? On the one hand, scholars interpreting the normative argument from bounded rationality by appealing to the ought-implies-can principle seem to start from the assumption that good reasoning is defined by its compliance with logic, probability theory, and rational decision theory. On the other hand, adaptive rationality theorists reject this premise and contend that imposing logic and probability theory as universal benchmarks of good reasoning is problematic. For instance, the logical standard does not allow for intelligent inferences about the meaning of words (i.e., “probably” and “and”) used in the Linda problem as well as the very purpose of the task (for a review of different interpretations of the conjunction fallacy see Moro, 2009). In particular, as adaptive rationality theorists have pointed out, the conjunction fallacy might actually be a case of genuinely adaptive behavior, and the conversational goal of being informative may have contributed to that finding (Hertwig & Gigerenzer, 1999). Specifically, the basic point here is that participants may be interpreting the goal of the task as a request to be as informative as possible. This seems plausible, because in normal conversation it is assumed that the speaker is cooperative and the cover story might be considered relevant for solving the problem. Hence, if participants are trying to be informative, and are thus ordering options according to their informational value given that profile, it makes perfect sense to choose a conjunction over one of the conjuncts.

But it seems fair to claim that besides important social goals and norms, people often care about saving time and energies as well. And these latter goals are what makes research on bounded rationality and on people’s cognitive limitations particularly relevant to the framework of adaptive rationality. Specifically, the reason why literature on people’s bounded rationality has normative relevance is that, even when following principles of the “standard picture of rationality” is not literally impossible, it can still be very laborious and demanding, and thus requires good amounts of time and energies. This is particularly important: we agents should be sensitive to the fact that following such norms is laborious and costs time, energies, and so forth, because, and as long as, these are goods that people value and care about. In other words, if following norms of the “standard picture of rationality” requires extensive computations and important investments in terms of time and efforts, those norms can hardly be used as benchmarks for the study of adaptive behavior and cognition, and it might well be that strategies that depart from tenets of the “standard picture of rationality” are comparatively more adaptive than those complying with it.

The importance of these considerations about speed and frugality emerges clearly in studies by Herbert Simon. His work on satisficing clearly illustrates the sorts of considerations presented above (e.g., Simon, 1979). In particular, Simon famously argued that decision-makers typically satisfice rather than optimize. A decision-maker normally chooses an alternative that meets or exceeds specified criteria, even when this alternative is not guaranteed to be unique or in any sense optimal. Simon argued that, instead of scanning all the possible alternatives, computing the probability of every outcome of each alternative, calculating the utility of each alternative, and thereupon selecting the optimal option with respect to expected utility, an organism typically chooses the first option that satisfies its “aspiration level.” But other scholars, especially in Gigerenzer et al.’s tradition, have described and formally modeled several other heuristics, such as the priority heuristic (Brandstätter, Gigerenzer, & Hertwig, 2006), which are allegedly used by people and are supposed to save them a good deal of time and energies when making decisions.

First, it is important to note that, while commentators in the debate have often interpreted considerations about the normative relevance of bounded rationality by appealing to the ought-implies-can principle (cf. Hands, 2014, p. 13), and suggesting that “the methods of classical rationality are computationally intractable and time consuming, and thus beyond the bounds of human decision makers” (Newell, 2005, p. 11), it is not clear whether such attribution is entirely correct. To be sure, authors such as Gigerenzer sometimes seem to appeal to versions of the ought-implies-can principle and to literal impossibility to follow rational norms. For instance, Gigerenzer (1996) writes that “Bayes’ rule and other rational algorithms quickly become complex and cognitively intractable, at least for ordinary human minds” (p. 277). But quite often the second version of the normative argument from bounded rationality seems to better capture what scholars offering those considerations had in mind: the problem is not that the norms used to assess behavior in bias research are cognitively intractable, but rather that, while following them does not go beyond the bounds of human decision-makers, it might still lead to maladaptive or non-adaptive behavior. Probably the best way to reconcile these two lines of reasoning is by stressing that whilst the tasks used in bias research do not typically involve cognitively intractable strategies, this is also because bias researchers have favored rather simple situations in their study of human rationality, which are not necessarily representative of the real world. Once we set aside the contexts typically explored in bias research, such as those characterized by monetary gambles, and consider rational behavior in “large worlds,” the computational complexity and related costs hugely increase (cf. Brighton & Gigerenzer, 2012). After all, a number of interesting real-world problems are computationally intractable, conducing to combinatorial explosion.

Second, an important remark here is that according to this second version of the argument, the request that we adhere to standards that are not overly demanding is not in fact independent of other arguments offered in the “rationality debate,” and in particular of the concern that biased thinking and decision-making can be adaptive. Actually, it only makes sense on the background of this view. Specifically, the reason why we should hesitate to accept the norms of the “standard picture of rationality” as normative benchmarks is that behavior that focuses on goals such as consistency but neglects other goals, such as speed and frugality, should not be seen as adaptive. What this suggests is that the normative argument from bounded rationality, construed along these lines, is actually a special case of the more general concern that behavior violating rational norms can be adaptive. Following principles of the “standard picture of rationality” can be non-adaptive or maladaptive for boundedly rational reasoners, as it can entail important losses in terms of time and energies, which are goods people care about. Heuristics and strategies departing from the “standard picture of rationality” can be more adaptively rational than strategies adhering to tenets of logic, probability theory, and rational decision theory.

Third, it is important to note that, at least prima facie, this version of the argument seems to be more promising than the one appealing to the ought-implies-can principle, since it seems to be relevant to phenomena described in bias research, unlike the first construal of the normative argument. But some caveats are nevertheless in order here. It is true, on the one hand, that the perspective of “adaptive rationality” can be questioned. For instance, the idea that mundane goals such as saving energies and time, and epistemic goals such as making accurate choices could in principle be traded is controversial. Specifically, one might argue that it is epistemic accuracy, and not agential success, that matters for normative assessment. It is important to note, however, that researchers from different parties in the “rationality debate” have defended the importance of adaptive behavior for attributions of rationality. For instance, cognitive psychologists Keith Stanovich and Richard West claim that “adaptive decision making is the quintessence of rationality” and that “to think rationally means taking the appropriate action given one’s goals and beliefs” (Stanovich & West, 2014, p. 81). A further concern here is that there are clearly many tasks discussed in bias research that do not seem to be vulnerable to this version of the argument. Consider once again, for example, the Linda problem used to elicit conjunction fallacies: as Gigerenzer (1997) himself acknowledged:

Limited cognitive resources are not an issue in the Linda problem. For finding the “correct” solution, absolutely no knowledge about the environment is needed and no resources for obtaining further information are required; thus the issue of limited does not apply. Similarly, little if any computational capacities are needed, and time constraints or information are of no relevance. (p. 204)

More specifically, even under this construal, advocates of the normative argument from bounded rationality seem to have in mind problems that are different from those that are generally associated with bias research: paradigmatic examples of that research are about how people cope with tasks involving high uncertainty, computational intractability, etc., namely conditions that hinder the “standard picture of rationality” from being reasonably applied. ¹ At the same time, unlike the characterization of the argument resting on the ought-implies-can principle and which did not seem to be applicable to behavior described in bias research, this second version of the argument seems to be relevant to biases and behavior discussed in bias research. Specifically, since the argument does not refer to literal impossibility to adhere to the norms of the “standard picture of rationality,” it seems to be more widely applicable. For instance, in Kahneman and Tversky’s heuristics-and-biases tradition it has been argued that trade-offs are too cognitively complex for people to manage them effectively (Kahneman, Slovic, & Tversky, 1982), and that they might thus end up using heuristics such as the lexicographic heuristic. The term lexicographic signifies that the cues are looked up in a fixed order of validity, like the alphabetic order used to arrange words in a dictionary. People using the lexicographic heuristic compare different options on one key attribute, such as price, size, weight, reliability, durability, calories, sugar, etc., and choose the option that performs the best on that single attribute, while generally ignoring the other attributes. For example, the take-the-best heuristic is an inference strategy in which cues are ordered lexicographically, comparing the cues one after the other and using the first cue that discriminates as the one reason to yield the final decision. Notably, while it is often stressed that people use lexicographic strategies because of their cognitive limitations, some evidence is at odds with this claim, suggesting that in fact in some contexts people who score higher at intelligence tests reason heuristically (Bröder, 2003), although the existence of correlations between cognitive ability and use of lexicographic strategies proposed is still somewhat undertested.

However, for an illustration of the normative relevance of considerations about the demandingness of the “standard picture of rationality,” let us consider conformity to norms of probability theory. In particular, let us examine an example suggested by Elqayam (2012), and consider someone who is not motivated to engage in effortful processing, for she finds hard thinking to be particularly painful and not really productive. She might have goals that pull in different directions. On the one hand, she might want to give an accurate answer to a problem requiring the application of Bayes’ theorem. However, she might also want to spend as little cognitive effort as possible. As it turns out, research in the field of judgment and decision-making has suggested that people are prone to incorrectly applying Bayes’ theorem. But what is important to stress here is that, for an agent who finds hard thinking rather painful, it might not be “adaptively rational” to comply with rules of the “standard picture of rationality,” as this would involve significant losses in time, and huge and painful cognitive effort. As we can see, it is fair to conclude that this version of the normative argument from bounded rationality seems more widely applicable and hence more promising than the one previously examined, which was resting on the application of the ought-implies-can principle.

Fourth, whilst the version of the argument presented above seems to be more promising than that appealing to the ought-implies-can principle, as it seems relevant to the findings discussed in bias research (e.g., base rate neglect), it is still unclear to what extent this version of the argument really supports the rather optimistic claims about human rationality generally attributed to advocates of the normative argument from bounded rationality. Specifically, this version of the argument suggests that reasoning according to the “standard picture of rationality” can be non-adaptive, as it may entail costs in terms of time and energies, and that heuristics might help save time and important energies. But from this it does not seem to follow that heuristics that violate such norms and allow us to save time and energies should be seen as rational, although optimism about human rationality is generally associated with the normative argument from bounded rationality. More precisely, if heuristics imply major losses in relevant goods other than speed and frugality, then the resulting behavior cannot be seen as optimal or as a clear instance of adaptive behavior and cognition. For example, saving time and energies can hardly be seen as an achievement when your goal is to get your prediction right.

A few important remarks are in order, though. Specifically, to strengthen the conclusions previously drawn about common justifications of heuristic reasoning, there are some other points that ought to be clarified. It is often claimed that heuristics only lead to saving time and energy, but it is also the case that some classes of heuristics perform better than “rational” algorithms when measured in terms of accuracy as well. In particular, whilst a classical justification for heuristics is that people save effort at the cost of accuracy, heuristics can be faster and more accurate than strategies that use more information and more computation. According to the framework of the “adaptive toolbox,” the algorithms that people use to make judgments and decisions do not use all the available information, but these strategies and heuristics lead nonetheless to adaptive behavior and cognition. In the literature less-is-more effects have been reported, where less information and computation lead to more accurate judgments (Gigerenzer & Brighton, 2009; Gigerenzer & Goldstein, 1996; Katsikopoulos, 2011; Martignon & Hoffrage, 2002). Whilst these so-called less-is-more effects might seem to be rather surprising, one explanation for them is that a strategy can make two kinds of errors, namely bias and variance. Specialized strategies have larger bias than more general ones, and general-purpose tools with many free parameters tend to generate more variance. A good cognitive system needs strategies that strike a balance between being too specialized and being too general. From this it follows that a mind with a general-purpose algorithm would not only be slow, but also perform inferior to a more specialized system if the error due to variance was larger than the error due to bias (for details, see Brighton & Gigerenzer, 2012; Gigerenzer & Brighton, 2009). This is a key message that Gigerenzer and other scholars researching the effectiveness of simple heuristics have tried to draw (Gigerenzer, Todd, & ABC Research Group, 1999). ²

This is not to say that fast-and-frugal heuristics obviously perform well in all environments, and the crucial task is to actually identify the structures of environments in which a given heuristic performs better (according to a specified criterion) than some other strategy, including complex methods. This is called the study of “ecological rationality,” and there is a huge body of research and answers to this question (Todd & Gigerenzer, 2012). Notably, in environments where frugal strategies do not lead to accurate predictions, the appeal to the normative value of frugal decision-making seems to lose cogency. After all, as it turns out, in some cases even small losses of accuracy might be rather important. To see this, consider that the importance of a goal seems to vary from context to context. For instance, the passage of time is certainly a pressing concern faced by an organism in a variety of dynamic environmental situations: organisms may have occasional speed-based encounters where the slower individual is placed at a serious disadvantage. In addition, the faster an organism can make decisions and act on them to accrue resources or reproductive opportunities, the greater advantage it will have over slower competitors. But at other times, accuracy is definitely a more pressing concern. In particular, there are some instances where making accurate decisions might be more important than saving time. In these latter contexts, even small losses in accuracy might result in maladaptive behavior on the side of the cognizer.

But the evidence about fast, frugal, and seemingly accurate heuristics concerns quite different heuristics and tasks from those typically investigated in bias research. Hence, the results about these particular classes of fast-and-frugal heuristics do not necessarily transfer to the heuristics and biases traditionally discussed in bias research in which adhering to principles of the “standard picture of rationality” seems to be particularly cognitively demanding. In particular, what advocates of the normative argument from bounded rationality have to do in order to challenge pessimistic assessments of human rationality is to show that, for the heuristics and biases reported in bias research, the benefits in terms of time and energies that these heuristics bring are not swamped by their costs when behavior is measured against other criteria and goals.

There are two serious hurdles with this task, though. First, not only are the heuristics in the heuristics and biases literature rather different in nature and definition from fast-and-frugal heuristics (for a review of different meanings of the term “heuristic” see Chow, 2015), but there are also striking differences in the precision with which heuristics are formalized. For instance, whilst it is often claimed that people make generalizations about frequency based on the availability of the instances in memory rather than by an accurate count of actual past experience, the relevant cognitive strategy is hardly spelled out in detail. Interestingly, Sedlmeier, Hertwig, and Gigerenzer (1998) defined the two most common meanings of availability. In one version, the availability heuristic rests on the actual frequencies of instances or occurrences retrieved, for instance, the occurrences of seizures among one’s acquaintances to assess the risk of seizures among middle-aged people. Another rests on the ease, that is, fluency with which the operation of retrieval of these instances and occurrences can be performed. Second, while in the case of the fast-and-frugal heuristics described by Gigerenzer and colleagues it seems quite easy to assess whether heuristics are successful or not (by considering, besides speed and frugality, the criterion of empirical accuracy), for many biases and heuristics described in bias research it is less straightforward to assess whether they are conducive to maladaptive behavior or not. In fact, most claims about heuristics being conducive to real-world disasters seem to be empirically ungrounded.

On the one hand, scholars in bias research often argue that such biases are conducive to poor life outcomes. I will cite one recent paper in which this link is explicitly emphasized. Milkman, Chugh, and Bazerman (2009) express the view that biases are conducive to maladaptive behavior. According to these authors:

Errors induced by biases in judgment lead decision makers to undersave for retirement, engage in needless conflict, marry the wrong partners, accept the wrong jobs, and wrongly invade countries. Given the massive costs that can result from suboptimal decision-making, it is critical for our field to focus increased effort on improving our knowledge about strategies that can lead to better decisions. (p. 379)

But this is not an exceptional statement (Dunning, Heath, & Suls, 2004, p. 70; Stanovich, 2011). Consider, for instance, what Johnson and Levin (2009) write:

In an ideal world, people would tackle major crises such as global climate change as rational actors, weighing the costs, benefits and probabilities of success of alternative policies accurately and impartially. Unfortunately, human brains are far from accurate and impartial. Mounting research in experimental psychology reveals that we are all subject to systematic biases in judgment and decision-making. … In today’s world of techno-logical sophistication, industrial power and mass societies, psychological biases can lead to disasters on an unprecedented scale. (p. 1593)

Clearly, if it were the case that heuristics make us save time and energies but, for example, make us fail to read medical prescriptions, this could be a non-trivial drawback of heuristic thinking and decision-making. The problem, however, is that there is little support for the claim that biases lead to maladaptive behavior (cf. Arkes, Gigerenzer, & Hertwig, in press; Polonioli, 2015). The only support generally presented for these claims comes from money pump arguments. Still, as some authors have pointed out, the very existence of money pumps is highly implausible, because no person would actually accept a series of bets unaware of the fact that he or she is being money pumped (e.g., Schick, 1986). But the threat of money pumps carries little weight if they are never instantiated, and some authors have actually emphasized that there is a corresponding “Czech book argument” which parallels the Dutch book argument, but with the conclusion that one ought to violate the probability calculus (Hájek, 2005). Even more importantly, the case can be made that in some scenarios biased behavior could actually be adaptive. This point is not completely new in the literature. For example, Nozick (1993) asks us to consider an individual who is strongly tempted to cheat on his spouse. If he does so, he will come to regard this act as a mistake for the rest of his life. The individual ultimately refrains, and does so in part because he reflects on how much he has invested in his marriage financially, emotionally, and temporally. These investments are sunk costs, but since he allows such allegedly irrelevant considerations to influence his decision-making, he is ultimately better off than he would be if he had ignored them and succumbed to temptation. What is important to stress, however, is that there is actually growing empirical research suggesting that behavior that departs from the tenets of the “standard picture of rationality” might be adaptive (e.g., Burns, 2001; Lenton, Penke, Todd, & Fasolo, 2013; Mercier & Sperber, 2011). For instance, when talking about overconfidence, Johnson, Weidmann, and Cederman (2011) point to the a “lottery effect”: even though overconfidence might lead to worse performance, overconfident people also engage more often in activities than people, therefore buying more lottery tickets in the competition for success.

Further, another reply available to advocates of the normative argument from bounded rationality to deny the link between biases and maladaptive behavior in the real world is to stress that the typical biases discussed in bias research are about the wrong kinds of problems and scenarios. After all, as we have already mentioned, advocates of this argument seem to have in mind problems that are different from those that are generally associated with bias research: paradigmatic examples of that research are about how people cope with tasks involving high uncertainty. So, the problems that these researchers have in mind seem to be more complex than those from bias research and, according to the advocates of this argument, more representative of the problems we face in the real world.

The abovementioned moves seem to be available to advocates of the normative argument from bounded rationality to challenge the conclusions from bias research. Still, what I have shown is that a proper assessment of the normative argument from bounded rationality requires and presupposes an assessment of other arguments and concerns expressed in the context of the “rationality debate.” More precisely, what needs to be assessed is whether the formats used in bias research are not representative of the real world and whether the oft-discussed biases described in the heuristics-and-biases literature can be conducive to good life outcomes, such as health, wealth, and happiness.

Conclusion

Different parties in the so-called “rationality debate” have sought to ground their views on considerations about people’s bounded rationality. Yet, this paper was specifically concerned with the appeal to considerations about people’s cognitive limitations to criticize the pessimistic views of human rationality that have been prominent for decades in the so-called “rationality debate.” More precisely, in this paper I took a fresh look at the normative argument from bounded rationality. I showed that commentators on the debate have generally interpreted this argument appealing to the ought-implies-can principle, and ultimately considered it unconvincing. But here I argued that there is an alternative way to interpret the argument, which appeals to the framework of adaptive rationality and better captures what scholars appealing to the normative relevance of people’s cognitive limitations had in mind when they were fleshing out their interest in people’s bounded rationality. Whilst this conceptualization looks more promising, the argument so construed no longer looks like an independent objections to pessimistic views on human rationality, but is instead a particular case of the more general concern that behavior violating rational norms can be adaptive. Further, it is unclear whether this version of the argument licenses the optimistic verdicts about rationality generally associated with scholars appealing to the normative relevance of bounded rationality. One main general message of this paper is that a great deal of literature on bounded rationality has been unduly dismissed as normatively and theoretically uninteresting only because scholars failed to appreciate the key premises of the framework of adaptive rationality and reasons behind it.