Reliabilist epistemology meets bounded rationality

Abstract

Epistemic reliabilism holds that a belief is justified if and only if it is produced by a reliable or truth-conducive process. I argue that reliabilism offers an epistemology for bounded rationality. This latter concept refers to normative and descriptive accounts of real-world reasoning instead of some ideal reasoning. However, as initially formulated, reliabilism involves an absolute, context-independent assessment of rationality that does not do justice to the fact that several processes are reliable in some reasoning environments but not in others, as is widely reported in the cognitive sciences literature. I consider possible solutions to this problem. Resorting to ‘normality reliabilism’, a variant of the theory, is one; but I find it insufficient. Therefore, in addition, I propose to relativise the reliability assessment to reasoning environments. This novel version of reliabilism fits bounded rationality better than the original one does.

1 Introduction

According to reliabilism, the epistemic theory first introduced by Alvin Goldman, a belief is justified if and only if it is produced by a reliable belief-forming process, which is one that usually generates true beliefs rather than false ones. In this essay, I argue that reliabilism offers an epistemology for bounded rationality, a term used to refer to a broad spectrum of normative and descriptive accounts of real-world reasoning instead of some ideal reasoning. However, my argument continues, reliabilism can only provide this service as long as it is subject to changes that lend a more contextual character to the judgement of reliability. Indeed, no version of the theory, neither the original one nor those proposed later, can account for reasoning processes that are reliable in certain types of reasoning environments and unreliable in others, while that this does occur is an empirical fact well-known in cognitive sciences.

In my discussion, I will assume that there is a close connection between epistemological investigation and empirical sciences. This naturalistic approach to epistemology has been adopted in several philosophical inquiries into justification and rationality conducted taking into consideration the findings of the psychology of reasoning (for instance, Bishop & Trout, 2005; Goldman, 1986, 2008; Kornblith, 2014; Sturm, 2020). The idea is that the psychological sciences offer empirical results about how we actually think and reason that might be useful or essential to make progress in addressing epistemic questions (cf. Feldman’s [2012] ‘cooperative naturalism’).

It should be noted that reliabilist epistemology and bounded rationality have already met. The venue was a colloquium on the foundation of the norms of rationality between leading philosophers and psychologists.¹ On that occasion, Alvin Goldman gave a talk on epistemological and psychological perspectives on human rationality. In his article which followed the colloquium, Goldman acknowledged that: “In fact, each of the other contributors to this symposium, Gerd Gigerenzer and Michael Bishop, has made roughly this proposal; each regards truth, or accuracy, as (at least) one type of consequence that is pertinent to a rationality appraisal.” (2008, p. 240). Gerd Gigerenzer is a cognitive psychologist who has propounded an original approach to bounded rationality called ‘ecological rationality’. His reply to Goldman arrived later in a 2019 article, where he claimed that “ecological rationality can… be evaluated by goals beyond coherence, such as predictive accuracy, frugality, and efficiency” (p. 3547) and that: “The extension of goals of rationality from coherence to performance… has much in common with existing approaches, such as Kitcher’s (1992) naturalism and Goldman’s (1999) epistemological reliabilism.” (p. 3556).

The take-home message from this exchange over the years seems to be that process reliabilism (PR, henceforth) fits Gigerenzer’s ecological approach to bounded rationality (EBR, from now on) pretty well. Assessing this match is my primary goal here. To make this assessment possible, I will need to go into the details of both standpoints. Thus, in Sect. 2, I introduce bounded rationality, focusing on EBR. I then (Sect. 3) present reliabilism and Goldman’s paradigmatic examples of reliable and unreliable processes. A relevant difference between PR and EBR will manifest itself at that stage. In Sect. 4, I will show how Goldman’s original, simple version of reliabilism involves an absolute assessment of rationality, which does not do justice to reasoning processes that are reliable in some contexts but not in others. This is problematic since the psychological sciences have shown how the performance of several reasoning strategies changes from one environment to another. In Sect. 5, I start searching for solutions to this problem which preserve the match between PR and EBR by appealing to a variant of PR called normality reliabilism or normal conditions reliabilism. Despite specifying ‘normal’ and ‘abnormal’ domains for the reliability assessment, I will claim that normality reliabilism fails to solve the problem of absolute reliability. Even so, the normality condition proves helpful to block some of the standard objections to reliabilism, so we should keep it rather than drop it. In the following section, I advance my proposal to relativise the reliability assessment to reasoning environments. The novel version of PR that will emerge can account for differences in the performance of one and the same process between one reasoning environment and another. Thus, my reliabilist epistemology offers a theory of rational belief that fits bounded rationality better than the original version does. In Sect. 7, I draw some conclusions.

2 Bounded rationality

The conditions in which human and non-human agents find themselves when drawing inferences or making decisions are often less than ideal. For instance, agents could evaluate alternative options having only scarce information about the alternatives. Similarly, they could need to forecast uncertain future events that are beyond easy prediction or lack the time needed to consider all relevant factors and be forced to come to an answer quickly. Moreover, both animal and machine computational capabilities have intrinsic limitations. For example, the storage capacity of human working memory is restricted to a small number of meaningful pieces of information. For these and similar reasons, the term ‘bounded rationality’ was introduced to qualify the reasoning capacities of agents who draw inferences and make decisions under environmental and cognitive constraints. Economist Herbert Simon initially proposed it as a departure from the conception of rationality common in neoclassical economics (Simon, 1957). Since then, it has been used in numerous other fields, including computer science, decision sciences, cognitive psychology, neuropsychology, biology, and philosophy.

Research on bounded rationality encompasses several accounts of behaviour which deviates from exhibiting ideal rationality. To begin with, it is possible to distinguish between descriptive and normative theories. Moreover, specific instances of these two kinds of theories are sometimes combined in distinct views on human reasoning. In the following two subsections, I briefly present the main accounts of such behaviour and then focus on EBR, one specific view.

2.1 Normative and descriptive accounts

Part of the inquiry concerning bounded rationality addresses the normative standards against which judgements and decisions are evaluated. These standards are those people are expected to reason by. Broadly speaking, there are two main views. The first one, the coherence-based view, takes compliance with rules of logic, probability, and decision theory as the rationality criterion. Dubbed by Stein (1996) the ‘standard picture of rationality’, this view considers reasoning to be at fault whenever it does not comply with these rules (cf., for instance, Piattelli-Palmarini, 1994).

Alternatively, the process-based view understands rationality as the use of appropriate reasoning processes (cf., for instance, Simon, 1976). The appropriateness of processes is often evaluated in consequentialist ways, such as promoting one or more epistemic values. Accuracy—the production of true beliefs while not producing false ones—is the value most often mentioned in this regard. Process-based theories of rationality that prioritise accuracy hold that reasoning is rational if it employs processes that allow it to arrive at the correct output in a cognitive task. When the rationality criterion is that it applies accurate processes, reasoning is evaluated on its performance.

Regarding empirical research, cognitive psychology is the discipline that has incited most debate about the bounds of reasoning. Here, it is possible to recognise two dominant lines of thought. The first trend is associated with the Heuristics and Biases programme (H&B, henceforth; see Kahneman et al., 1982; Gilovich et al., 2002). According to this view, reasoners’ cognitive limitations make it challenging to reason according to the rules of logic, probability, and decision theory. Thus, people frequently use heuristics to reduce complex tasks, such as assessing probabilities and predicting values, to simpler judgemental operations. However, heuristics often depart from the standard norms and tend to make people fall into systematic reasoning mistakes called biases. For example, people are prone to commit the conjunction fallacy when they evaluate the probability of two co-occurring events (Tversky & Kahneman, 1983), or they tend to neglect relevant base-rate information when predicting uncertain events (Kahneman & Tversky, 1973).

The Ecological Rationality programme promotes the second prominent view in this debate (see Todd & Gigerenzer, 2012; Gigerenzer et al., 1999). Supporters of this view have drawn attention to the capacity of animal and artificial agents to take advantage of environmental structures to draw intelligent inferences and make smart decisions. People can adapt their cognitive strategies to the context in which reasoning occurs, using those strategies that fit the features of the environment and result in good reasoning. Most notably, it has been empirically demonstrated that simple heuristics can enable people to draw correct inferences, and sometimes even more correct inferences than those produced by complex strategies that obey logical and probabilistic norms (see Gigerenzer et al., 2011). For instance, simple recognition heuristic can outperform Wimbledon experts and ATP rankings in predicting the outcomes of Wimbledon matches (Serwe & Frings, 2006); and fast and frugal decision trees can be more accurate than logistic regressions and other complex statistical models in several tasks, such as predicting bank failure (Aikman et al., 2014) or in medical diagnoses such as detecting depressed mood (Jenny et al., 2013).

Some combinations of normative and descriptive accounts might have already become evident to the reader. On the one hand, H&B subscribes to the coherence-based theory of rationality and evaluates reasoning on its respect for standard rules. On the other hand, EBR assesses as rational those inferences and decisions which fulfil the epistemic goals, embracing the process-based and consequentialist standpoint (cf., for instance, Schurz & Hertwig, 2019).

The differences between the two schools of thought are substantial. They represent two cultures of research on bounded rationality, telling two different stories about rationality (Katsikopoulos, 2014). These differences fostered debate within the psychology domain in the 1990s, sometimes leading to harsh clashes.² The two psychology programmes portray images of human rationality in different shades of colour: H&B shows numerous ways humans fail to conform to the standard rules of reasoning, while EBR points out many instances of adaptive and successful behaviour.

So far, I have distinguished between normative and descriptive accounts of bounded rationality and illustrated how they can be combined in existing research programmes. I will now focus on EBR, given its compatibility with epistemic reliabilism.

2.2 The ecological approach to bounded rationality

EBR is committed to explaining how animal and artificial agents cope with the bounds of reasoning. The empirical research conducted within this programme encompasses cases of both theoretical and practical reasoning. For example, it studies how reasoners form beliefs when deciding between two alternatives (e.g., which city has more inhabitants, Milan or Brescia?) and how agents choose the course of action to take (e.g., how should money be invested in a certain number of assets?). However, here I will only look at cases of theoretical reasoning and consider bounded rationality only as a form of epistemic rationality.³

According to the ecological approach, every agent facing an inferential or decisional task has recourse to a plurality of reasoning strategies for addressing the task. These are algorithmic procedures of different kinds. Some are complex and make extensive calculations; others, the heuristics, are fast and frugal processes that produce outcomes in little time and using few pieces of information. Altogether, they constitute the ‘cognitive toolbox’, the repertoire of tools for drawing inferences and making decisions. It is crucial that the reasoner selects the strategy that best suits the reasoning task and the context in which the cognitive performance takes place. Indeed, using the appropriate reasoning method substantially contributes to reasoning success (cf. Gigerenzer & Sturm, 2012, pp. 251–252; Hertwig et al., 2021, pp. 17–18).

Ecological rationality builds on Herbert Simon’s adaptive view of rational behaviour. As Simon puts it, “Human rational behavior… is shaped by a scissors [sic] whose two blades are the structure of the task environments and the computational capabilities of the actor” (Simon, 1990, p. 7). This claim does not have to be understood in descriptive terms only. Indeed, it contains an important normative message: what is to be considered rational behaviour depends on the relationship between the reasoning environment and the cognitive capacities of the reasoner. Ecological rationality emerges when the structure of reasoning mechanisms matches the structure of the environment and the cognitive setup of the reasoner. Therefore, no judgement of rationality can be made without reference to the context in which the reasoning is performed.

Ecological rationality, normatively understood, is about the success of reasoning strategies in the world, assessed by the following principle (Brighton & Todd, 2009, p. 337):

ECOLOGICAL RATIONALITY

A mechanism M is ecologically rational in environment E in comparison to some other mechanism M’ when M outperforms M’ on some criterion, or currency, of comparison.

The criterion of comparison most often used is the accuracy of judgement. However, other criteria are also sometimes adopted, such as speed (how fast an inference can be drawn or a decision made), frugality (how many pieces of information need to be searched before a judgement can be made), etc.

It might be helpful to see an example from the EBR literature to understand how this principle works. Consider this inferential task:

Creditworthiness

of two companies, you have to choose the more creditworthy.

To draw the appropriate inference,⁴ agents usually use information gathered from different cues such as the company’s financial flexibility, its efficiency, the qualifications of the employees, etc. Psychological research has shown that these cues can be processed in somewhat different ways (Gigerenzer & Goldstein, 1996). Suppose all the cues are binary (0 and 1), where 1 is the higher criterion value, and they can be ranked according to their validity. By cue validity, we mean the probability of drawing a correct inference on the condition that the cue used in drawing that inference discriminates between the options. The higher the cue validity, the higher the probability of drawing a correct inference when relying on that cue. In our example, this is the probability of correctly inferring which of two companies is more creditworthy relying on a cue whose value is positive (1) for one company and non-positive (0 or unknown) for the other.⁵ The first strategy people might choose from their cognitive toolbox is the linear weighted additive strategy (WADD). For each company, WADD calculates the sum of all cue values multiplied by the cue’s validity and then selects the alternative with the largest sum. Alternatively, there is evidence that people rely on a simpler method known as the take-the-best heuristic (TTB). TTB is a one-reason decision-making method that decides which of two objects scores higher on a criterion solely on the basis of the most valid cue that discriminates between the two objects. This strategy consists of three steps (Gigerenzer, 2008, p. 32):

• 1. Search rule: Search through cues in order of their validity. Look up the cue values of the cue with the highest validity first.

• 2. Stopping rule: If one object has a positive cue value and the other does not (or is unknown), stop the search and proceed to Step 3. Otherwise, ignore this cue and return to Step 1. If no more cues are found, guess.

• 3. Decision rule: Predict that the object with the positive cue value has the higher value on the criterion.

How each strategy performs is a matter of empirical research, and the evidence gathered by proponents of EBR shows something quite surprising. Despite its simplicity, TTB matches alternative, more complex and cognitively demanding strategies, including WADD, in inferential accuracy. This happens when specific circumstances obtain, for example, when the information available to draw the inference is non-compensatory. This concept refers to the fact that each cue available for choosing between the two alternatives cannot be outweighed by any combination of less valid cues. To see what this means, consider the following explanation from Gigerenzer (2008, p. 37). Assume an environment with M binary cues ordered according to their weights W_j, with 1 ≤ j ≤ M. A set of cue weights is non-compensatory if the following property holds for each cue’s weight:

$${W}_{j}>\sum_{k>j}{W}_{k}$$

As an example, take a set of weights that decreases exponentially, such as 1, 1/2, 1/4, 1/8, and so on. The sum of the cue weights to the right of a cue can never be larger than this cue’s own weight. In environments with this feature, namely, non-compensatory environments, no linear strategy can outperform the faster and more frugal TTB, achieving at most the same level of accuracy as this simple heuristic⁶ (Martignon & Hoffrage, 2002).

Furthermore, TTB can also outperform WADD when the non-compensatory condition holds (Brighton & Gigerenzer, 2015). This phenomenon is called the less-can-be-more effect and is one of the most astonishing findings of EBR research.⁷ Under non-compensatory conditions, using TTB is ecologically rational. Alternatively, if this does not hold, and for instance we have compensatory conditions, then TTB performs poorly and is usually outperformed by WADD. TTB is not ecologically rational in compensatory environments, and people would be better off relying on other strategies, such as WADD.

This example shows how rationality is assessed in the ecological framework. Each of the two strategies for solving Creditworthiness is more accurate than the other in certain environments. Therefore, given the definition of ecological rationality, it is rational to use it in such circumstances. This rationality assessment in terms of accuracy reveals the consequentialist nature of EBR. This is where bounded rationality meets epistemic reliabilism.

3 Epistemic reliabilism

In the late 1970s, Alvin Goldman put forth his first reliabilist theory. Since then, he and other authors have kept developing that original proposal, which has been amended by numerous alternative accounts.⁸ Initially introduced as a normative theory of justification, it has also been presented as a theory of rationality (Goldman, 2008, p. 240). A common feature of the original ‘simple’ theory and more sophisticated versions is that the rationality of a belief depends on the reliability of the processes which give rise to the belief in question. In this sense, PR is a historical or genetic theory that makes the rationality of a belief depend on the belief’s provenance, namely, the process that generates it. This also makes it an indirect form of epistemic consequentialism: the rightness of belief is determined by the reliability of the process which generates the belief and not by its consequences, as maintained instead by direct consequentialism (cf. Ahlstrom-Vij & Dunn, 2018, p. 6).

Epistemic reliabilism rests on process reliability. In the words of Goldman, “a reasoning process is rational if and only if it is reliable—i.e., usually generates true beliefs rather than false ones” (Goldman, 2008, p. 240). This property characterises all rational processes and distinguishes them from irrational ones. More fundamentally, reliabilism assumes that accuracy, or truth-conduciveness, is the primary epistemic goal the reasoner ought to pursue. Beliefs owe their epistemic goodness to the rationality of their belief-forming processes. Thus, simple PR consists of the following claim:

A belief is rational if and only if it is the output of a belief-forming process that is reliable—i.e., one that usually generates true beliefs rather than false ones.

Belief-forming processes confer rationality if they have a high truth ratio, that is, most of the beliefs they produce are true. Precisely how high the truth ratio must be is left vague. The truth-ratio threshold need not be as high as 1, but it must be greater (presumably much greater) than 0.50.

PR distinguishes between justified and unjustified beliefs according to the reliability of the processes that generate them. In his writings, Goldman provides the following examples, we can assume without claiming completeness for them:

Unreliable processes: Confused reasoning, wishful thinking, reliance on emotional attachment, mere hunch or guesswork, hasty generalisation (Goldman, 1979, p. 9), and failure to take account of all one’s relevant evidence (or failure to take account of obviously relevant evidence) (Goldman, 1986, p. 104).

Reliable processes: Standard perceptual processes, remembering, good reasoning, introspection (Goldman, 1979, p. 10), and certain patterns of deductive and inductive reasoning (Goldman, 1986, pp. 103–104).

By using these and similar reliable processes and avoiding the use of unreliable ones, the reasoner satisfies the necessary condition for producing rational beliefs. By doing so, reasoners will draw more accurate inferences than they would have done by relying on non-truth-conducive tools.

These are the essential features of simple PR. Epistemologists have raised objections to it, and more elaborated versions have been proposed to evade them. To assess the match of PR with bounded rationality, we can, for the moment, stick to Goldman’s simple version. Later, I will introduce normality reliabilism, one of its developments, and show how it copes with some objections to the original proposal.

Now, a tempting inference is that every agent reasoning along the lines prescribed by this theory also fulfils the principle of ecological rationality presented above. Indeed, we should expect reliable processes to perform better than unreliable ones, scoring higher on accuracy. However, the distance between simple PR and EBR is not as short as one might imagine or hope. Some findings of the psychological sciences regarding the working and effectiveness of inferential methods are hard to square with the prototypical instances of reliable and unreliable processes. In particular, inferential processes that do not consider all available evidence can produce true beliefs contrary to what is assumed by simple reliabilists. This points to a significant normative difference between simple PR and EBR.

4 Absolute versus contextual assessment of rationality

In the previous sections, I described the main characteristics of EBR and PR. To sum up: (i) both programmes advance a consequentialist account of epistemic rationality, understood as the achievement of some epistemic values; and (ii) among these values, accuracy, or truth-conduciveness, is the most relevant. At this point, someone might object that accuracy is not all that matters for bounded rationality theorists. This is indeed the case: they also measure the success of cognitive strategies in the world by other criteria, such as speed and frugality. This point requires clarification.

EBR theorists make rationality assessments using a variety of criteria. Looking at these assessments, one notices two recurring points. First, some criteria, such as speed and frugality, are used only occasionally. Moreover, one criterion is never absent in evaluations of rationality: accuracy. Hence, it seems plausible to infer that accuracy is a necessary condition for ecological epistemic rationality, whereas speed and frugality are not. If this is true, we should conclude that accuracy is the most relevant rationality criterion for EBR.⁹

We can also add summary point (iii): both theories of rationality are based upon the use of appropriate reasoning methods, referred to as ‘processes’ by reliabilists and ‘strategies’ by psychologists of bounded rationality. These methods are algorithmic procedures for deriving inferences and making decisions. This is to say, the methods arrive at outcomes by following rule-based patterns of reasoning. This is the case for complex, highly resource-consuming methods and also for fast and frugal heuristics.¹⁰

These are the relevant similarities between PR and EBR; but we can now ask: Does PR perfectly fit EBR? In response, it is possible to point out at least two dissimilarities between the two. If we recall the definitions of ecological rationality and process reliability, we will notice the first difference. The former is a comparative notion of rationality: a strategy is ecologically rational compared to some other strategy when the former outperforms the latter on some evaluation criterion. For example, when used in non-compensatory environments, TTB is ecologically rational compared to other inference models, such as WADD. By contrast, process reliability does not require comparison: a reasoning process is rational if and only if it is reliable.¹¹ Therefore, strictly speaking, a reliable process is not ecologically rational unless there is evidence that it outperforms rival processes. But what about the reverse: Is an ecologically rational process also reliable? The definition of ecological rationality does not allow us to say that it is: ecological rationality poses no requirement regarding the truth-conduciveness of a process. However, I think we must say that ecologically rational processes are also truth conducive, otherwise ecological rationality would get into trouble. Let me illustrate this claim with an example.

Imagine two reasoning strategies, S and S’, whose truth ratios are 0.30 and 0.20, respectively. If ecological rationality were merely a matter of outperforming, then we should conclude that S is (ecologically) rational, for S produces more true beliefs than S’. However, nobody would evaluate S as rational, given its low level of accuracy. Thus, there seems to be a hidden condition that ecological rationality requires a minimum level of accuracy, and plausibly this must be much greater than 0.50, as it is for PR. Further support comes from the study of simple heuristics. By scrutinising the EBR literature, one can see that no heuristic with an average accuracy of less than 0.50 (or probably less than 0.60 or 0.70) has been put forward. When assessed as ecologically rational, heuristics are always remarkably accurate (for a recent systematic review of the literature, see Katsikopoulos et al., 2018).

So, if ecologically rational processes must reach a high truth-ratio threshold, they are reliable. However, the opposite cannot automatically be said to be the case: not all reliable processes are ecologically rational. This discrepancy does not have troublesome consequences for the match between PR and EBR; and bearing this asymmetry in mind, we can safely defend a strong compatibility between the two programmes.

There is a second difference between PR and EBR. Recall the original definition of reliability again: it is the tendency of a process to generate true beliefs rather than false ones. Where this tendency needs to be shown or in which reasoning context, is a piece of information that the definition of simple reliability does not require. Simple reliabilism is based on a non-located notion of rationality. Now, recall how ecological rationality is defined: a reasoning strategy is ecologically rational in environment E compared to some other mechanism when it outperforms this other mechanism on some criterion of comparison. Here, the rationality assessment is relativised to a specific context or environment of reasoning; whereas, as we have just recalled, this relativisation is absent in Goldman’s reliabilism.

This discrepancy creates trouble. To see why the non-locatedness of PR is a problem for its match with EBR, consider once again Creditworthiness. When inferring which of the two companies is the more creditworthy, people can either use WADD, which processes all the relevant cues about the two companies, or TTB, which only relies on a fraction of the available data. Evaluating these processes according to simple reliabilism, one should conclude that TTB is irrational. Indeed, as we saw before, not taking into account all one’s relevant evidence is considered a paradigmatic example of an unreliable process. However, empirical research in cognitive psychology has shown that TTB can predict which company is more creditworthy as accurately as WADD can, and sometimes even more in non-compensatory environments, thus proving to be ecologically rational. Therefore, PR considers irrational a reasoning process that is considered rational in a specific context by EBR, although the two accounts start from a very similar concept of rationality.¹²

This dissimilarity highlights how simple PR adopts a relatively rigid notion of reliability. According to this, a reasoning process is rational if and only if it tends to produce true beliefs rather than false ones. Apparently, it is tacitly assumed that the process ought to do so wherever it is used. I will call this an absolute understanding of process reliability. By not specifying where a process needs to produce true beliefs to be assessed as rational and assuming that this is in every place it is used, simple PR precludes the possibility that a method is truth-conducive in specific reasoning contexts and not truth-conducive in others. However, the reliability of a process can change from one context to another, as shown by Creditworthiness. Therefore, simple PR is not equipped to account for such changes in the performance of reasoning processes from one environment to another. This is a substantive theoretical issue which deserves attention.

Although PR does not perfectly fit EBR, I believe that something can be done to remedy this; something that is worth doing. Indeed, there are good reasons for preserving their match. On the one hand, PR can strengthen the theoretical underpinnings of EBR, securing it from puzzling rationality assessments. As I have shown, the purely comparative evaluation EBR offers allows for cases of ecologically rational but non-truth-conducive processes. To avoid this pitfall, EBR should also require a high truth ratio. This leads to a hybrid definition of ecological rationality according to which a reasoning mechanism M is ecologically rational in environment E in comparison to some other mechanism M’ only if M is more accurate than M’, and M exceeds a relatively high accuracy threshold.

On the other hand, the psychology research pursued by the ecological project might be seen as a practical implementation of reliabilism. Moreover, EBR can also help to improve some conceptual aspects of the epistemic theory proposed by Goldman and other reliabilists, and contribute to overcoming the problem of absolute reliability, as we will see in Sect. 6.

So, to preserve the match, simple PR should be modified. It seems clear that it needs to acquire the capacity to relativise reliability assessment to the contexts in which the belief-forming processes are used. I will now consider a couple of ways to do this. The first possible solution I contemplate is using a tool the reliabilist family already possesses.

5 Reliability under normal conditions

One of the most disputed questions facing reliabilism concerns the domain in which a process is assessed for reliability. Much of the dispute originated from the so-called ‘new evil demon problem’ (Cohen, 1984). According to this problem, it is possible to imagine a world where an evil demon creates non-veridical perceptions of physical objects in everybody’s mind, which are qualitatively identical to our own, but false in the imagined world. Therefore, the perceptual processes of the inhabitants of this world are unreliable, and their beliefs so caused are unjustified. However, given that their perceptual experiences are qualitatively identical to ours, those very beliefs in the demon world should be justified. Recently, some authors proposed a solution to this problem by taking the domain in which reliability is assessed as the ‘normal conditions’ for using a given process. Since these conditions are free from evil demons, perceptual and other processes are indeed reliable in these domains.

Normality-based versions of reliabilism as a theory of justified belief can be found, for instance, in Leplin (2007, 2009) and Graham (2012, 2017). Meanwhile, Beddor and Pavese (2020) introduced a ‘normal conditions’ variant of the reliabilist account of knowledge. This account offers a simple and effective tool to handle those cases in which knowledge is generated in contexts we would be hard pushed to consider suitable for its production. With some adjustment, this tool can also be used in a theory of rational belief. Moreover, it might offer a solution to the problem of absolute reliability.

Consider the following example by Beddor and Pavese, adapted to a theory of rationality. Temp forms beliefs about the room temperature by consulting a broken thermometer. Unbeknownst to him, Temp has a guardian angel in the room who manipulates the thermostat, ensuring that the room’s temperature matches the reading displayed on the thermometer. Any belief Temp forms about the temperature will be true. However, intuitively, Temp’s thermal beliefs are not rational.

This case represents an objection to the sufficiency of a reliability condition for rational belief. This objection can be blocked by restricting the domain of reliability assessment to normal conditions, understood by Beddor and Pavese as those we would consider ‘fair’ for performing and assessing the task.¹³ So, to defend reliabilism, we might want to modify simple reliability by adding the normality condition:

NORMAL CONDITIONS RELIABILITY

A reasoning process is rational if and only if it is reliable in normal conditions—i.e., it usually generates true beliefs rather than false ones in such conditions.

Accordingly, Normal Conditions Reliabilism (NCR, from now on) holds that a belief is rational if and only if produced by a process which is reliable in normal conditions. Now, let us consider Temp once more. The task being performed is forming beliefs about the temperature on the basis of a thermometer. The reasoning process used is consulting a broken thermometer. Intuitively, having a hidden helper manipulating the thermostat is a highly abnormal circumstance for Temp’s task. However, in normal conditions, namely, in helper-free worlds, consulting a broken thermometer leads to false beliefs. So, Temp’s reasoning process is unreliable in normal conditions. Therefore, NCR deems Temp’s beliefs to be irrational.

Unlike Temp’s case, the new evil demon problem challenges the necessity of the reliability condition for rational belief. Can NCR block this criticism as well? Consider Danny, a subject deceived by the evil demon, who possesses false perceptions of physical objects identical to ours (Danny could alternatively be a subject in a vat whose brain is stimulated by a supercomputer which generates non-veridical representations of the world). The task Danny is performing is forming beliefs on the basis of visual stimuli or simply forming visual beliefs. To do so, the perceptions of physical objects are relied upon. Intuitively, the normal conditions for Danny’s task are free from the demon’s interventions. Thus, Danny is in highly abnormal conditions for performing and assessing the task. In the demon world, Danny’s reasoning process is unreliable, and the resultant beliefs are false. In contrast, Danny’s reasoning process is reliable in normal conditions. Hence, NCR deems Danny’s beliefs in the demon world to be rational (as also for the brain-in-vat case).

Normality-based reliabilism is a variant of the traditional theory that Goldman has welcomed (Goldman & Beddor, 2021, pp. 10–11). Its novelty consists in relativising the assessment of process reliability to normal conditions, understood as those we would consider fair for performing and assessing a given task. Now, the questions for us are: Can normality reliabilism account for the change in performance of a reasoning method from one environment to another? Can it explain why using TTB in non-compensatory environments is rational but it is irrational in compensatory ones? And: Does NCR succeed where simple reliabilism fails?

Empirical and analytical research has shown that the performance of many reasoning strategies is influenced by features of the reasoning context and the cognitive setup of the reasoner. Non-compensatoriness, for instance, makes TTB as accurate as WADD and sometimes even more accurate. Should we conclude that non-compensatory environments are normal conditions for TTB? These conditions are those we consider fair for performing and assessing a reasoning task. For instance, under these conditions, no manipulation has occurred to make an unreliable process truth-conducive (as occurred in the Temp case) or to make a reliable process non-truth-conducive (as in the evil demon case). However, the normality condition does not seem to solve the TTB case. Indeed, the change in performance of TTB from one environment to another is not due to any manipulation of the reliability of the process. Instead, we would need to say that the performance of TTB is rational in an environment having a specific feature, namely, non-compensatoriness. Fairness for performance and assessment seems to be irrelevant here. Hence, NCR fails to offer a solution to the problem of absolute reliability.

Notice that this failure does not mandate abandoning the normality condition, which blocks the new evil demon and the guardian angel objections to simple PR. We should not renounce this achievement, given that we propose PR as an epistemology for bounded rationality. Therefore, we should embed the normality condition in our definition of process reliability and continue the search for a way to complete the account and solve the problem of absolute reliability.

6 A contextual approach to reliability

In light of the foregoing considerations, relativising reliability assessment is the way to go. Here, in addition to normality reliabilism, inspiration can be found in a recent dispute in the reliabilist camp. The so-called ‘temporality problem’ for reliabilism (Frise, 2018; Tolly, 2019) consists of the fact that a process can be more reliable at one time than another. For instance, weather forecasting has improved over time. Thus, the process type ‘forming a belief based on the forecast’ is more reliable now than it was twenty years ago. This raises a question about the temporal parameters we should use when evaluating reliability. Imagine we are evaluating whether belief B is justified at time t. Should we focus on whether the belief-forming process responsible for B is reliable at t? Or should we consider whether it has always been reliable up until t; or at all times that are temporally close to t? Or something else? According to Frise (2018), the temporality problem generates insurmountable difficulties for reliabilism, whereas Tolly (2019) rejects this conclusion and shows that there are reasonable temporal parameters for the reliabilist to adopt.

I do not need to enter into the details of the issue. Nevertheless, it provides helpful insight when framing a solution to the problem of absolute reliability. First, it seems to support the idea that an absolute conception of reliability is implausible: it may be the case that most reasoning processes are more or less reliable depending on some contextual features, such as the time or the place of their use. Hence, it encourages the relativisation of reliability. Moreover, the issue suggests that, by analogy with the temporality problem, reliabilism could also have a spatiality problem: a process can be more reliable in one environment than another. Thus, a possible way to tackle the question of absolute reliability might be to introduce a spatial parameter E which relativises the reliability assessment to specific environments. Reliabilism could borrow this parameter from ecological rationality, which already possesses one. Hence, as a first approximation, an environment is generated by the interplay between features of the reasoning scenario (for instance, available information, time pressure, etc.) and features of the cognitive setup of the reasoner (for example, memory states, computational faculties, etc.).

I propose that the assessment of process reliability should be a function of the process truth ratio, the conditions for performing and assessing the reasoning task (i.e., whether they are normal or abnormal), and the environment in which the process is used. Therefore, the following principle should replace simple reliability:

NORMAL ENVIRONMENT RELIABILITY

A reasoning process is rational in environment E if and only if it is reliable in environment E under normal conditions—i.e., it usually generates true beliefs rather than false ones in this environment under normal conditions.

To illustrate this principle, recall the task of assessing the creditworthiness of two companies. The cognitive goal is to make the most accurate judgement based on the available cues. Two reasoning strategies are available to achieve this: TTB and WADD. Cognitive psychology has shown that the former performs as well as the latter and sometimes even better in environments characterised by non-compensatory information. In contrast, TTB performs poorly and is usually outperformed by WADD in environments characterised by compensatory information. The normative conclusion is that TTB, an inferential process that ignores part of the available evidence, is rational in non-compensatory environments since it is reliable there—i.e., it tends to produce true beliefs rather than false ones in these environments. In contrast, it is irrational in compensatory environments since it is unreliable there—i.e., it tends to produce false beliefs rather than true ones in these environments. Thus, normal environment reliability can account for the change in the performance of this reasoning process from some environments to others.

With this new notion of reliability, we can now introduce the corresponding epistemic theory called Normal Environments Reliabilism (NER, henceforth), according to which a belief is rational in environment E if and only if it is the output of a belief-forming process that is reliable in environment E under normal conditions—i.e., it usually generates true beliefs rather than false ones in this environment under normal conditions. This is a candidate for an epistemology for EBR. Notice that EBR only offers a theory of rational reasoning strategy: an account of what makes a belief-forming process rational. In this sense, ecological rationality parallels process reliability. However, EBR provides no fully fledged epistemology intended as a theory of rational belief. A further step is needed, and reliabilism seems helpful here. The beliefs an agent produces under environmental and cognitive constraints are rational in the context of production if and only if they are generated by belief-forming processes that, in normal conditions, are reliable in that context.

Let us consider the main features of the new account. First, it should be clear that normal environment reliability is not more demanding than absolute reliability, as it does not introduce any further rationality criterion. Accuracy, or truth-conduciveness, remains the only standard for evaluating the performance of a reasoning process.

Second, normal environment reliability is a located notion of rationality: it refers to the tendency of a reasoning process to be truth-conducive in specific contexts. Thus, reference must be made to environments when assessing this kind of reliability. To do this, the assessor should use knowledge about where a strategy performs well and where it does not.

Moreover, the parameter E and the normality condition play different functions, and we need both. E individuates reasoning contexts with features that play out in favour of or against the performance of a reasoning process, such as non-compensatoriness and compensatoriness for TTB. Meanwhile, normality discriminates between normal and abnormal conditions for the performance and assessment of a task. Conditions are normal when nothing has altered the reliability of a process that can be applied to tackle the task. In Creditworthiness, conditions are normal if evil-demon-like factors have not modified the reliability of TTB and WADD. Imagine instead that an agent is assisted by her guardian angel who, whenever and wherever she uses TTB, ensures that her prediction matches the actual relative creditworthiness of the two companies. Suppose also that the agent is in a compensatory environment. The reasoner will draw many true inferences thanks to the aid of the guardian angel. Will they be rational? Under normal, guardian-angel-free conditions, TTB is unreliable in compensatory environments. Hence, the agent’s inferences produced by TTB will not be rational. Suppose now that the agent is instead in a non-compensatory environment. In this case, the agent’s inferences produced by TTB will be rational because TTB is reliable in non-compensatory environments.

Finally, NER is more descriptively plausible than the simple, absolute version. Indeed, it allows us to distinguish where a process is reliable and where it is not, and it does not oblige us to consider a process reliable or unreliable wherever it is used. This brings justice to the empirical evidence gathered by EBR.

Reliability can now be relativised to environments where reasoning methods are used or usable, making the environment a central notion in the new epistemic framework. So, at this point, it might be opportune to clarify how this notion should be understood.

6.1 Reasoning environments

Reasoning environments are contexts within which and upon which the reasoner acts. Their features affect the agent’s thoughts and actions in various ways. To begin with, they provide the input processed by reasoners and shape the tools available for reaching their goals. Moreover, reasoning environments influence process reliability. As we saw in Creditworthiness, the reliability of processes such as TTB and WADD sensibly varies depending on the environment in which the reasoning takes place. Therefore, the reliability assessment cannot neglect environmental factors.

Offering a complete characterisation of reasoning environments might be difficult, but it is possible to highlight several important structures. Gigerenzer and Sturm (2012) isolated three: the degree of uncertainty, the number of alternatives, and the learning sample size. This list is not meant to be complete; other structural features might be added. Time pressure is a candidate here. For instance, Marewski and Schooler (2011) show that the inferential accuracy of heuristics varies depending on whether people can take all the time they want to draw the inference or are given a short timeframe. But I will now briefly address the three items already on the list.

Uncertainty. The degree of uncertainty refers to the extent to which available cues can predict a criterion. Some predictions are more uncertain than others, depending on what is predicted. For instance, the future performance of stocks and funds is highly unpredictable; heart attacks are slightly more predictable; and tomorrow’s weather is even more predictable. Uncertainty characterises most reasoning environments in which humans can find themselves. According to the definition introduced by economist Frank Knight (1921), it applies to all those situations where we cannot know all the information about a future event and, therefore, cannot calculate its probability accurately. It must be distinguished from risk, which is the feature of those situations where, although we do not know the outcome of a future event, we can accurately measure its probability. A pure game of chance in a casino is an example of a risky situation.

Number of alternatives. When drawing inferences or making decisions, people can deal with different numbers of alternatives. These can refer to individual objects (such as companies) or sequences (such as moves or pathways). This environmental structure becomes particularly relevant when the alternatives are many. Algorithmic procedures for estimating a criterion might run into the problem of computational intractability, the impossibility of estimating the criterion optimally, given a massive number of alternatives. Consider, for instance, the game of chess. Although an optimal, best sequence of moves does, in theory, exist, no mind or computer can determine it. This is one of the reasons why both machines and humans rely on simpler, non-optimisation techniques, including heuristics.

Sample size. This structural feature of the environment constitutes the number of sampling units included in the sample. Statistical methods for predicting future events estimate parameter values from past data. Sample size directly influences the performance of predictive strategies. For example, consider two models, one complex with many free parameters and a heuristic with few parameters, predicting an uncertain future event using a small sample of data. In this circumstance, the resulting error due to ‘variance’ committed by the complex model may exceed the error due to the heuristic’s ‘bias’, making the latter strategy preferable (Gigerenzer & Brighton, 2009). Thus, different sample sizes can call for different predictive methods.

Notice that ‘environment’ is not used here to indicate only the physical environment. For instance, the degree of uncertainty reflects the environment as well as the limited understanding of the mind. Hence, the degree of uncertainty is located in the mind–environment system (Gigerenzer & Sturm, 2012, pp. 256–258). More generally, the environment here is to be understood as the totality of the relevant physical features of the world that, combined with those of the reasoner’s cognitive system, favour the performance of some reasoning strategies over that of others.

7 Conclusion

I have explored the relationship between reliabilist epistemology and the ecological approach to bounded rationality. The two accounts have been recognised as largely compatible by the authors who have contributed most to their development. However, I first show that the comparative definition of ecological rationality allows an unreliable process to be assessed as rational. To avoid this problem, I suggest that EBR might adopt a reliability condition. Furthermore, I highlight how the absolute character of Goldman’s simple reliabilism does not square with the located notion of rationality adopted by EBR. Simple PR cannot account for processes that are reliable in one reasoning context and unreliable in others, the existence of which is an empirical fact widely reported in the psychology literature.

A first attempt to save the relationship between the two theories resorted to normality reliabilism, which relativises process reliability to normal conditions, understood as ‘fair’ domains for performing and assessing a reasoning task. Although the normality condition solves the new evil demon problem for reliabilism, it does not capture those environmental features that play out in favour of or against a reasoning strategy. Thus, I proposed solving the problem of absolute reliability by relativising process reliability to reasoning environments. Normal environment reliability adds the normality condition and the environmental parameter E to simple reliability. The former blocks some standard objections to simple PR¹⁴; the latter individuates reasoning environments whose features influence process reliability. A new theory of rational beliefs emerges, NER, which, as opposed to the original reliabilist view, does justice to those reasoning processes that are truth-conducive in some environments and non-truth-conducive in others, such as TTB in non-compensatory and compensatory environments, respectively.

We can see an advantage of NER when it comes to heuristic reasoning. As we saw at the beginning, there is disagreement about the pros and cons of heuristics. We can interpret the ongoing debate as a controversy about the reliability of heuristics, since scholars disagree about the tendency of heuristics to produce true beliefs rather than false ones. For instance, Goldman expressed his sympathy for the line of thought according to which heuristics are “highly error-prone, or indeed biased toward error”, the “quick and dirty” view often associated with Tversky and Kahneman’s H&B approach (Goldman, 2017, p. 24). However, this view clashes with the empirical findings of Gigerenzer and his colleagues of the EBR, which have instead shown when heuristics perform well and when they do not. Adopting a contextual approach to epistemic rationality might contribute to solving the dispute about heuristics since relativising rationality assessment to reasoning environments leads to more balanced conclusions. Indeed, using normal environment reliability, it is possible to express that a particular heuristic is reliable in some environments and unreliable in others. Accordingly, agents are rational if they use such a heuristic in the former environments and irrational if they do so in the latter.

On the one hand, I hope to have contributed to strengthening the philosophical underpinnings of EBR by exploring and critically assessing its relationship with PR. On the other hand, I believe that reliabilism can now account for empirical facts that are well-known in cognitive sciences. If this is indeed the case, this work can represent a starting point for future research across philosophy and psychology of reasoning.