1. THE ETHICAL AND EPISTEMOLOGICAL BACKGROUND

At some point in the distant past, ethics and epistemology were positioned in opposite corners of the philosophical establishment. One was the epitome of “theoretical” philosophy and the other the epitome of “practical” philosophy. Relatively little contact was made between them. Today, however, analogies abound between ethics and epistemology. It is widely accepted, for example, that epistemology has a strong strand of normativity and should therefore be viewed (at least in certain respects) as a legitimate cousin of ethics. Ethics has been the chief battleground of powerful debates between consequentialism and deontology, but epistemology has also witnessed an analogous set of disputes.

Selim Berker is immersed in both of these disputes. In two recent articles (Berker Reference Berker2013a, Reference Berker2013b) he highlights the pervasiveness of consequentialism, or teleology, in contemporary epistemology. The main form of epistemic consequentialism regards truth – or true belief plus error avoidance – as the “final” value or end in the epistemological domain. Epistemic teleologists take other epistemic goods to have derivative value in virtue of their instrumental relationship to the aforementioned final end or ends. Berker himself rejects all forms of epistemic consequentialism, particularly insofar as they purport to explain justification, reasonability, rationality, etc. in terms of those truth-linked values or ends. He regards process reliabilism as “the paradigm example of a teleological epistemic theory” (2013a: 350), and it is the central target of his critique. This paper responds to Berker's critique. As the originator and long-time proponent of process reliabilism (see especially Goldman Reference Goldman and Pappas1979, Reference Goldman1986, Reference Goldman1992, Reference Goldman2012), I ask the reader's indulgence in my focusing attention on this particular theory of justification in the present response to Berker's considerably broader critique.Footnote ¹ During the course of the paper I shall also address the connection between process reliabilism and a related, but distinct, epistemic theory, i.e., “veritism,” which is also a target of Berker's critique.

Let us set the stage for Berker's arguments by considering some of the scholarly background he provides for the debate between consequentialism and deontology within the ethics of the last century. Berker clearly sees the contemporary issues for epistemology through the lens of the ethics literature. This is important partly because the parallel with ethics leads him, in my opinion, to some misconstruals of reliabilism, which might not have been made if he were not so consumed with ethics parallels. Here are some of the particularly salient passages from ethicists that Berker cites in the run-up to his discussion of epistemic consequentialism. These quotations can help us understand where Berker is “coming from” in his interpretation of epistemic consequentialism.

The continuing debate between Kantianism and utilitarianism (the most prominent form of moral consequentialism) is vividly depicted by Christine Korsgaard in the following passage:

A similar passage by Philippa Foot attests to the depth and power of the dispute:

Turning from moral to epistemic consequentialism, Berker himself writes:

I see Berker's invocation of Korsgaard's and Foot's longstanding battle with utilitarianism as providing assistance in conjuring up a suitably fierce and persistent opponent for his battle with epistemic consequentialism. If, switching metaphors a bit, he can portray reliabilism as a suitably fierce dragon, slaying that opponent will be all the more worthy and creditable. I don't mind his portraying reliabilism as a battle-worthy dragon. If it is going to be portrayed as a dragon, however, I would like him to be straight on the sort of fire it breathes. A large part of the argument of the present paper is that Berker brings the wrong kind of fire-retardant to the task because he misconstrues the dragon's type of fire. Or, returning to Berker's own metaphor, he proposes misguided formulas to utter in trying to exorcise the specter because he misunderstands the nature of the spook.

Reliabilism, says Berker, might be better called “truth-conducivism.” The guiding thought is that what makes a (token) belief epistemically justified is its connection to a process (or indicator, or method) that conduces toward the promotion of true belief and the avoidance of false belief. This is analogous to various kinds of ethical consequentialism or teleology, in which an action, rule, motive, etc. qualifies as morally right if and only if it conduces to good consequences, or better consequences than its alternatives. In the first and longer of the two papers Berker suggests that a principal shortcoming of this kind of theory is that it ignores the “separateness of propositions.” This is a good place to begin, especially in light of the fact that the phrase “the separateness of propositions” appears in the title of this paper. What does this mean?

Consider Judith Thomson's (Reference Thomson1976) example of a transplant surgeon who encounters a patient during a routine check-up and decides to chop him up, take his organs, and transplant them in five other patients, each in need of one of those organs. From a teleological point of view, this action is good, perhaps even right. Greater overall happiness will be promoted by this action than by simply attending to the original patient and letting those in need of organ transplants die. Following many moral theorists, however, Berker argues that trade-offs between persons should in general be avoided. This point was initially introduced by John Rawls (Reference Rawls1971), who rejected any principle of utility maximization for this kind of reason.

2. ALLEGED EPISTEMIC ANALOGS OF ETHICAL TRADE-OFF PROBLEMS

Now consider epistemological analogs of such a trade-off case. Berker presents a series of such examples, starting with one in the style of Roderick Firth (Reference Firth1981), though in a form shaped by Richard Fumerton (Reference Fumerton and Steup2001). In this example Berker (Reference Berker2013a: 364) imagines himself to be a scientist seeking a grant from a religious organization that only gives grants to believers. He is a longstanding atheist, but knows he couldn't lie about this matter. Driven by a keen desire to obtain many true scientific beliefs by means of winning the grant, he manages to form a belief in God's existence. Is this belief justified? Intuitively, the correct answer is “No.” What verdict does epistemic, or veritistic, consequentialism deliver? Here is what Berker says:

The clear implicature is that the veritist – including, according to Berker, the process reliabilist – is committed to a verdict about this case that is mistaken. In other words, the grant-seeking scientist case is a counterexample to veritism in general and process reliabilism in particular. Surprisingly, the following sentence appears immediately after the above-quoted passage: “My aim here is not to argue by counterexample” (365). This remark is very difficult to understand (or take literally) because it is hard to see how Berker's critique of epistemic teleology in this paper and its companion could proceed or succeed unless Berker's examples – at least some of them – are intended to be counterexamples. What other argumentative philosophical form could his two essays assume if not that of refutation (or challenge) by counterexample?Footnote ² In any case, Berker proceeds to write as follows:

According to Berker, veritistic epistemic consequentialism is committed to the verdict that the grant-seeking scientist's belief in God is justified. The belief is justified because it generates a greater balance of true beliefs over errors, most of the true beliefs being effects of the grant. Although the belief in God, we may suppose, is false, the numerous true scientific beliefs that are its products outweigh this false belief, making the belief a good choice from a “veritistic” point of view. Berker writes: “It is difficult to see how veritism, at least as far as we have defined it so far, can avoid saying that such a trade-off would maximize overall epistemic value” (Berker Reference Berker2013b: 364).

3. WHY RELIABILISM IS INNOCENT OF ALL CHARGES

We see, then, where Berker is coming from, that is, what feature of reliabilism, veritism, or epistemic teleology supposedly gets it into trouble. But is it guilty of the charges leveled by Berker? First of all, is it correct that reliabilism would deem the grant-seeking scientist's belief to be justified? Not at all, I contend. I agree with Berker that classifying this or similarly generated beliefs as justified would be “beyond the pale” (2013a: 363). What I deny is Berker's claim that reliabilism is committed to such a verdict, or comes anywhere close to implying it.

With the help of Figure 1, let us review the core ideas of process reliabilism and explain why it would deliver the correct verdict, i.e., “unjustified.” Simplifying somewhat, the core conditions of process reliabilism can be stated as follows (drawing on “What Is Justified Belief?,” the original statement of process reliabilism).

In the grant-seeking scientist case the target belief is a belief in God's existence. This belief is produced by a desire-driven process, something like wishful thinking. A desire or hope to bring about many true scientific beliefs (of his own or other people) generates or outputs a belief in God's existence. Thus, the operative belief-forming process is an instance of the process type, wishful thinking, which is clearly an unreliable process. This much is clearly implied by the example, though not explicitly stated in the description of the case. Hence, not only is the scientist's belief intuitively unjustified, but process reliabilism would definitely so classify it. End of story. The fact that the target belief token, Bel(P), also leads to many subsequent true beliefs has no bearing on its justificational status (J-status) according to process reliabilism.

Fig. 1.

Producing many subsequent true beliefs does not help a target belief qualify as justified.

As can readily be inferred from what we have said, process reliabilism is a backward-looking theory, not a forward-looking one. It is backward in the sense that the J-statuses assigned by this theory are always a function of events upstream from the target belief, not downstream from it. This is contrary to what Berker claims about consequentialist theories in general and process reliabilism in particular. Process reliabilism is a historical theory (Goldman Reference Goldman and Pappas1979, Reference Goldman1986: 78, 83; Reference Goldman1999, Reference Goldman2009). In other words, it makes the J-status of a belief depend on the belief's provenance, not its subsequent effects. Thus, the J-status of the grant-seeking scientist's theistic belief depends on the belief-forming process that immediately outputs it (as shown in Figure 1 by the arrow immediately to the left of the oval shape labeled “Bel(God exists),” and perhaps by the desire itself. Nothing else depicted in the diagram is relevant to this belief's J-status. (Had other beliefs been causes of the target one – for example, by being early links in a long chain of inferences – they too could have been justificationally relevant; but there are no such beliefs in this scenario.)

If the foregoing is correct, Berker's interpretation of process reliabilism is very wide of the mark. How might he have gotten the process reliabilist story so wrong? One possible problem is an early passage he cites (but misinterprets) from “What Is Justified Belief?” He writes:

This, however, is an inaccurate characterization of what I was saying in the relevant passage. The referenced passage concerned a meta-level question of how exactly theories of justification should be formulated. Should they be formulated using other normative terms or only substantive (factual) terms, as would be suitable for an account that aims to be a “reductive” analysis of its analysandum? After posing this question, I announced my own intention to take the latter route, and I compared this course to act utilitarianism in normative ethics. Under act utilitarianism, only such descriptive terms as happiness, unhappiness, and causality play central roles. No normative terms (especially no morally normative terms) are employed. A parallel strategy, I urged, should be adopted in constructing a theory of justifiedness.Footnote ⁵ It was only this “formal,” meta-theoretic respect of similarity between reliabilism and act utilitarianism that I meant to endorse. Nothing I said there implied that process reliabilism is a species of consequentialism or instrumentalism in which beliefs downstream from a target belief (token) might be relevant to the latter's J-status.Footnote ⁶

A separate point is worth making at this juncture, a point concerning the alleged principle that one should not “flout the separateness of propositions.” The implication seems to be that each belief token should be epistemically evaluated “on its own merits.” It J-status should not be beholden to any other beliefs or their J-statuses. (This is a more definite statement of what a no-flouting principle would hold than any that Berker offers; but it seems to be in his spirit.) But such a principle would be misguided. When one infers a given proposition from other propositions that one antecedently believes, then the J-status of one's conclusion belief does depend upon – or is beholden to – the J-statuses of one's antecedent premise beliefs. In this very straightforward sense, the epistemic statuses of propositions are not always “separate from” – i.e., independent of – the epistemic statuses of other propositions. Process reliabilism is right on track in incorporating this sort of epistemological connectivity, and is in no way imperiled by any alleged principle of the separateness of propositions; because no legitimate principle of this sort has been demonstrated.Footnote ⁷

4. RELIABILISM AND FORWARD- VERSUS BACKWARD-LOOKINGNESS

However, it must be acknowledged that there are treatments of reliabilism, including some of my own treatments, that support a certain chunk of what Berker says in drawing parallels between reliabilism and rule utilitarianism. Indeed, I drew such a parallel myself in Epistemology and Cognition (Goldman Reference Goldman1986). There I advanced a rule consequentialist framework for process reliabilism, and classified it as a form of “verific consequentialism” (Reference Goldman1986: 103). So Berker's interpretation is not wholly ungrounded. But does this framework imply that ascriptions of justificational status to (token) beliefs have a “forward-looking” as opposed to a “backward-looking” character, as Berker says? Not so, I reply.

First of all, despite my rule framework, I still introduced the same sort of historical theory that was endorsed in “What Is Justified Belief?” In presenting the rule framework, I write as follows:

Clearly, according to this principle, it is states and state-transitions leading to the target belief that confer justification on that belief, not what follows (temporally) from that belief. Moreover, two pages later, I make it clear that even within a rule-framework approach belief-forming processes are to be viewed as critical:

With the process element re-introduced into the picture, the old (Reference Goldman and Pappas1979) story about the determinants of justifiedness is firmly back in place. The J-status of a given belief depends on what happens in the “run-up” to the belief, not what happens afterwards. And once again, crucial to this run-up are the psychological processes that generate the belief in question. So even if one chooses Epistemology and Cognition as the preferred formulation of reliabilism, one still finds no story to the effect that the J-status of a belief is determined by that belief's consequences (“down the road”). Yet this is what Berker claims in his interpretation of reliabilism, notably in the grant-seeking scientist case.

The “central slogan” of epistemic consequentialism, he writes, is that “Epistemic normativity is a matter of bringing about states of affairs with a distinctive sort of value” (Berker Reference Berker2013b: 377). As we see, however, process reliabilism does not hold that the determination of a belief's J-status is a matter of what it brings about – certainly it isn't directly a matter of such instrumental relations.

Berker is correct, however, in stressing the fact that under process reliabilism justificational normativity derives from the consequences of belief-forming process types, or what their outputs are.Footnote ⁸ But is this reliabilist claim about justificational sources mistaken? Recall here some of the principal points made in the original proposal of process reliabilism (Goldman Reference Goldman and Pappas1979). It was first pointed out that the cause(s) of a token belief is critical to its justificational status. Examples were given in which beliefs satisfied some orthodox account of justifiedness – for example, their being well-supported by the agent's evidence – but were intuitively unjustified because they were caused in strange, quirky, or unacceptable fashions. For example, they were caused by wishful thinking, or the accidental movement of a brain surgeon's hand, or the blinding aura of the Presidency. The argument continued as follows:

So far, however, this only shows that a token belief's provenance is a key to its J-status. But what features of its provenance are crucial? Reflection suggests that the belief-producing process must be one that, generally speaking (not necessarily on the occasion in question), is prone to be reliable, i.e., tends to output many more true beliefs than false beliefs.

Thus, process reliability seems exactly on track in pinpointing the propensity of a process type to output more true beliefs than false beliefs as the key to the determination of justificational statuses. (It is not all downstream effects of a process type that figure in the formula; only its immediate outputs.) Although this formula does not use the terminology of epistemic “value,” it is no stretch to connect up with that language. The important point is that these ideas seem right, not mistaken, as Berker claims. This is the best explanation of why process reliabilism rapidly became a serious contender among theories of justification, soon after “What Is Justified Belief?” was published. It offers a parsimonious account of the intuitions about cases reviewed above.

What we have acknowledged is a covert forward-lookingness embedded in the background of process reliabilism. Types of processes have justification-conferring or unjustification-conferring properties as a function of their truth ratios, where truth-ratios are a matter of the pattern of outputs that the different process types have. Such forward-lookingness is relevant, however, only for process types, not to the (target) belief tokens themselves. Belief tokens are what get evaluated as justified or unjustified, reasonable or unreasonable, knowledge-worthy or not knowledge-worthy. These bearers of J-status do not enjoy their respective J-statuses as a function of their consequences, contrary to what Berker seems to imply in his discussion of the grant-seeking scientist.

Berker has his own way of distinguishing different components of the stories offered by process reliabilists and other so-called epistemic teleologists. He distinguishes between different “deontic focal points” such as beliefs, on the one hand, and belief-forming processes on the other. However, despite acknowledging such differences, he makes inadequate use of them by my lights.

5. OTHER POSSIBLE COUNTEREXAMPLES

It is time to examine two other counterexamples discussed by Berker: Roderick Firth's (Reference Firth1981) John Doe case and Berker's own prime number case. Here is the John Doe case:

Berker says that Firth's John Doe example and the others like it “point the way to a more serious problem that afflicts almost all varieties of consequentialism with respect to epistemic matters” (2013b: 369):

This passage runs together two things that should be kept apart: (A) a theory of epistemic value called “veritism” and (B) a theory of belief justifiedness called “process reliabilism.” It is correct that in Knowledge in a Social World (Goldman Reference Goldman1999) I advocated a form of epistemic consequentialism called “veritism.” And it is correct that this epistemic consequentialist scheme was used to evaluate (epistemically) a wide variety of social practices such as interpersonal argumentation and judicial rules for the admitting or excluding evidence in a trial. Such rules and practices, it was argued, can be assessed in veritistic terms, and this is a reasonable thing for theorists to do, since patterns of truth-value attainment are of central importance in many social (as well as individual) domains. But Knowledge in a Social World never proposed that the indicated veritistic evaluations are justification-directed evaluations. Specifically, the justification of beliefs simply wasn't on the table. Veritistic evaluation, as understood there, was indeed a form of direct consequentialist assessment. Practices, rules, and institutional arrangements, it was urged, should be evaluated in terms of “downstream” effects such as the truth-ratio of jury verdicts. But it was nowhere suggested that belief justifiedness was to be so evaluated.

Contrary to Berker, therefore, it is not really “difficult” to see how a veritist can avoid the (unwanted) conclusion that John Doe's recovery belief is epistemically justified. As we saw earlier, process reliabilism asserts that it is the causes, not the effects, of a belief that are relevant to its J-status. So it should be clear that if John Doe's recovery belief is to be evaluated justificationally, it must be evaluated in terms of its causes, not its downstream effects. Veritism of the sort advocated in Knowledge in a Social World simply has nothing to say about (token) belief evaluation (in justificational terms). Hence it seems straightforward for veritism to avoid taking a position about John Doe's belief. The veritistic program advanced in Knowledge in a Social World isn't in the business of making justification evaluations.

Despite the passage quoted above, Berker's subsequent discussion of the John Doe case indicates that he does not, in the end, intend his original John Doe case to be a counterexample to process reliabilism. What he does claim, however, is that just a little tweaking of the case will yield such a counterexample:

The reasoning here is fairly straightforward, as I would reconstruct it. According to process reliabilism, a process is justification-conferring just in case it has a high truth-ratio. So a process of “wishful recovery belief-formation” is justification-conferring just in case it has a high truth ratio. By tweaking the original John Doe example so that Doe's wishful recovery belief-forming process not only generates many true beliefs of his own, but this same pattern recurs frequently in Doe's world. Hence the process type is reliable (Berker implies) and process reliabilism must declare Doe's output belief(s) to be justified.

Berker thus seems to assume that the domain in which a process type should be evaluated for reliability is the actual world of the epistemic subject, in this case, John Doe. Truth ratio in the actual world as a whole, of course, comprehends far more than the truth ratio of a single subject's beliefs. Berker's assumption, however, does not comport well with the unresolved, continuing debate on this issue among reliabilists themselves. My own publications (including Goldman Reference Goldman and Pappas1979, Reference Goldman1986, Reference Goldman2012) have presented several alternative possibilities. These include the following three alternatives.

First, the domain of evaluation might be expanded beyond the actual world of the agent to other nearby possible worlds. Second, the domain of evaluation might be neither the actual world nor neighboring worlds but the “normal world,” where “normality” is specified by what ordinary people presume to be true of the actual world (cf. Goldman Reference Goldman1986: 108). Third, reliabilism about justification might be revised more extensively by introducing an interpretation I originally called “two-stage” reliabilism (Goldman Reference Goldman1992), but am currently inclined (following a suggestion of Elizabeth Fricker) to label “approved-list” reliabilism.

The idea behind the third of these options is that people pass through two stages in acquiring and deploying the cognitive tools for making justification evaluations. In the first stage they acquire beliefs about which belief-forming process types are reliable or unreliable. This is based on their own and/or other people's “observations” of which belief-forming processes have high or low truth-ratios when used in the actual world. Second, after acquiring and mentally storing lists of such “approved” and “unapproved” processes, they proceed to apply their favored and disfavored process types to scenarios in which rather different truth-ratios for processes may be operative (e.g., in worlds ruled by an evil demon). When making judgments at the evaluation stage, they do not judge a process's reliability by what is stipulated to occur in the hypothetical scenario, but rather by the attributors' previously entrenched lists of approved and unapproved process types. Thus, in the case of John Doe, we evaluate his recovery belief as unjustified because we previously stored wishful thinking in our minds as an unreliable process. This unapproved (indeed, disapproved) process will continue to exert its influence on our evaluation of Doe's belief even if we are told, as in Berker's tweaking, that recovery beliefs are generally quite reliable in the actual world Doe inhabits.

My present inclination is to favor the approved-list form of reliabilism (it will be applied to the next putative counterexample too). But it isn't necessary at this juncture to register a decided preference for any one of these alternative forms of reliabilism. The main point is that it isn't easy to refute reliabilism by a simple example such as the “tweaked” version of the John Doe story, because reliabilism has several options for ironing out this wrinkle. Berker needs to give us a broader argument to demonstrate reliabilism's alleged futility. Perhaps such a broader argument is what he intended with his appeal to the separateness of propositions. Unfortunately, for reasons provided earlier, that “broad” line of argument fails. Here I am saying that his critique by counterexample seems at best inconclusive. And to the extent that many others have offered similar types of counterexamples, it is also not very novel.

There remains a final counterexample that Berker lodges against process reliabilism, a more novel and interesting one, in my opinion. It is his prime number case, which runs as follows. Whenever S considers whether a particular natural number n is prime, S forms a belief that n isn't prime. This process tends toward a truth-ratio of 1.0 as n increases, so process reliabilism, says Berker, must declare that all of these beliefs are justified. Intuitively, however, they are unjustified. How might reliabilism reply to this apparent counterexample? There are several possible responses.

First, suppose that the domain we select for determining a process-type's reliability is the actual world (plus, perhaps, a class of a relevantly nearby possible world). I now make the following observation. In all likelihood, people who use this process type apply it exclusively or primarily to relatively small integers. Within that class of numbers, however, the truth-ratio of the process will not be very high, because many small integers are prime (1, 2, 3, 5, 7, 11, etc.). So it is questionable whether the counterexample even gets off the ground.

A second reply would invoke a rarely discussed constraint I proposed on belief-forming processes in “What Is Justified Belief?,” a constraint requiring eligible process types not to be content-specific. Instead, they must be content-neutral. For epistemological purposes, suitable process types should be general processes, which can presumably take many different contents as inputs (and as outputs). No process should be confined to a single, specific subject-matter. Thus, what Berker calls a “process” in his prime number example cannot really be accepted by process reliabilism. Agent S may be said to use a certain rule for deciding whether a particular natural number is prime (always saying “no” for every natural number specified). But this rule should not be equated with a process. Processes will presumably be used, of course, in following the rule, but that does not make the rule itself a process. A distinction between “processes” and “methods” was drawn in Epistemology and Cognition (Goldman Reference Goldman1986: 93–5), proposing that processes be understood as basic psychological processes and methods are understood as learnable algorithms, heuristics, or methodologies (not part of one's native cognitive equipment). The rule in Berker's prime number example clearly falls on the methods side of this distinction. So it is reasonable to impose a criterion of justifiedness for methods that is suitable to methods rather than processes. Use of a reliable method does not automatically or invariably confer justifiedness on an output belief, because a perfectly reliable method might have been acquired in a defective – e.g., haphazard though lucky – fashion. If that is how an agent acquired a given method, its reliability per se does not guarantee justifiedness of a belief so generated. Unless an agent acquires Berker's prime-number rule appropriately, a belief generated by it won't be justified. If it is acquired appropriately, though, its output beliefs would be justified. Hence, no counterexample would emerge.

Finally, the approved-list version of reliabilism looks like a very promising way to handle the counterexample – even if we do not press the distinction between processes and methods. It is part of the approved-list version of reliabilism (see Goldman Reference Goldman1992) that when a novel process type fails to match any type already on one's approved list or unapproved list, then an epistemic evaluator will find the most similar type on one of the lists and classify the target process type accordingly. A likely point of comparison for purposes of Berker's prime number example is some process type that is entirely arbitrary and unmotivated. That's the character of the prime number rule as Berker presents it. But arbitrary and unmotivated process types can be expected to belong on an agent's unapproved list, not on her approved list. So Berker's rule would be regarded as unreliable and hence incapable of generating justified beliefs – the correct verdict by everybody's lights.

In sum, once Berker's misinterpretation of process reliabilism is straightened out, his putative counterexamples turn out to be either totally unproblematic for reliabilism or at best questionable cases that reliabilism has pretty good prospects for accommodating.