I believe that a commitment to the epistemic analogue of consequentialist (or teleological) theories in ethics is pervasive in contemporary epistemology. These epistemic theories (i) take the good to be prior to the right in epistemology and (ii) explain epistemic rightness in terms of epistemic goodness via a promoting (or conducing) relation. Characterized thusly, epistemic consequentialism encompasses a wide array of currently popular theories, cutting across many of the familiar divides, including internalism vs. externalism, foundationalism vs. coherentism, and virtue- vs. belief-focused epistemology. But perhaps the most prominent instance of a consequentialist epistemic theory is process reliabilism, which I see as the epistemic analogue of satisficing hedonistic rule-utilitarianism (where instead of evaluating whether an action is based on a rule that does sufficiently well at promoting pleasure, we evaluate whether a belief is formed via a process that does sufficiently well at promoting true belief).

I also believe that this widespread commitment to consequentialism in epistemology is a mistake. My strategy for arguing against epistemic consequentialism has three stages (see Berker Reference Berker2013a, Reference Berker2013b). First, I consider certain counterexamples to extremely crude versions of epistemic consequentialism originally due to Roderick Firth (Reference Firth1981, Reference Firth and Troyer1998). Call these ‘Firth cases’. Second, I use the analogy with consequentialist theories in ethics to unearth the structural features that make Firth cases work the way they do. Third, I show how almost all forms of epistemic consequentialism sometimes allow a belief's epistemic status to depend on those same structural features, thereby giving me a recipe for generating counterexamples to even very sophisticated forms of epistemic consequentialism. Call the examples generated in this third stage ‘Firth-inspired cases’. They go beyond anything we find in Firth's work, but, if I am right, they demonstrate the same underlying insight that led Firth to propose his own cases.

Since most epistemic consequentialists develop their theories in a way that allows them to deflect Firth cases, it is the second and third stages of my strategy that are crucial. The basic idea is to argue that all interesting forms of epistemic consequentialism condone (or mandate) the epistemic analogue of cutting up one innocent person in order to use her organs to save the lives of five people.Footnote ¹ The difficult part is figuring out exactly what the epistemic analogue of cutting up the one to save the five consists in. This requires careful attention to the structure of Firth cases, so that we can use that structure to generate Firth-inspired counterexamples to the varieties of epistemic consequentialism one actually encounters in the literature.

In ‘Epistemic Teleology and the Separateness of Propositions’ (written in 2009), I postulated that the structural feature underlying Firth cases is this: the promotion of epistemic good with regard to one proposition is sacrificed in order to greatly benefit the promotion of epistemic good with regard to many other propositions. In slogan form: theories that run afoul of Firth's original examples do so because they ‘ignore the separateness of propositions’. In ‘The Rejection of Epistemic Consequentialism’ (written two years later), I changed my mind and decided that although the structural feature of ignoring the separateness of propositions suffices to generate problems for many consequentialist theories in epistemology, there is an even more general structural feature subsuming that structural feature which allows us to generate problem cases even for those consequentialist epistemic theories which adhere to the separateness of propositions. I am still struggling to figure out the best way to characterize this more general structural feature. But that won't matter for the purposes of this note: all of the theories I will be considering flout the separateness of propositions, so there will be no need to go beyond my initial diagnosis.

Alvin Goldman is the father of process reliabilism, so it is not surprising that he finds my criticism of epistemic consequentialism in general, and of process reliabilism in particular, unpersuasive. Goldman argues in ‘Reliabilism, Veritism, and Epistemic Teleology’ that the form of process reliabilism he favors can avoid my Firth-inspired attack. What exactly is process reliabilism? Here some terminology will help. Goldman (Reference Goldman and Pappas1979, Reference Goldman2012: 38) thinks of a process as ‘a functional operation or procedure, i.e., something that generates a mapping from certain states – “inputs” – into other states – “outputs” ’. In the case of cognitive processes, the output states are all beliefs.Footnote ² On the input end, we can distinguish between belief-independent processes, which take no belief-states as inputs, and belief-dependent processes, which take at least some belief-states as inputs. We can also distinguish between process tokens and process types in the usual manner. It is process types, not process tokens, which are assessed for reliability, and there are two ways of doing this, depending on whether the process is belief-independent or -dependent:

Process reliabilism may then be defined recursively as follows:Footnote ³

How do we determine the ‘relevant’ process type for a given process token? This, of course, is the dreaded generality problem for reliabilism.Footnote ⁴

Goldman's reply to my anti-teleological argument consists in arguing that three of my cases are not counterexamples to process reliabilism: the grant-seeking scientist case (a Firth case whose specific formulation I take from Fumerton Reference Fumerton and Steup2001), the sick mathematician case (another Firth case), and the prime numbers case (a Firth-inspired case customized to cause trouble for process reliabilism). Much of what Goldman says about these specific cases strikes me as correct. However, in restricting his reply to an assessment of these three cases, Goldman has failed to adequately address my anti-teleological strategy, for two reasons. First, I never claimed that Firth cases are counterexamples to process reliabilism. Second, Goldman's argument that the prime numbers case (and others like it) do not present problems for process reliabilists is uncompelling.

All Firth cases involve a seemingly unjustified belief being a causal means of promoting, sometime in the future, the acquisition of numerous true beliefs by the subject.Footnote ⁵ Firth cases are not counterexamples to process reliabilism, since in the definitions of unconditional and conditional reliability, ‘yield’ should be interpreted to mean ‘immediately yield’. All that matters for an assessment of either type of reliability is the truth-ratio of the token belief-states directly brought about by token instances of the process type being evaluated; any other token belief-states causally downstream from those token belief-states are irrelevant. So I agree with Goldman that the grant-seeking scientist and sick mathematician cases are not counterexamples to process reliabilism. In fact, I explicitly said as much in my two articles.Footnote ⁶

But process reliabilists are not so easily off the hook. Note the structure of Firth cases: they involve forgoing the promotion of true belief with regard to one proposition in order to greatly benefit the promotion of true belief with regard to many other propositions. In Firth's original cases, the item doing the forgoing and promoting is a belief token. However, that same underlying structure can also apply to cases in which something else, such as a process type, is sacrificing its performance with regard to one proposition in order to promote its performance with regard to many other propositions. For this reason, process reliabilism sometimes allows the structure at work in Firth cases to be normatively relevant, and when it does, I claim, it reaches an incorrect verdict. My prime numbers example is supposed to be one (but only one) case of this sort.

Here was my initial description of the case: ‘Suppose that whenever I consider whether a given natural number n is prime, I form the following belief: <n is not prime>’.Footnote ⁷ The general structure of the case – which I will call ‘structure S’ – was intended to be as follows:Footnote ⁸

Why are cases with structure S problematic for process reliabilism? Let us refer to the propositions in R₁ as ‘the few’ and the propositions in R₂ as ‘the many’. Intuitively,Footnote ⁹ when a subject uses X to form a belief in one of the few, that belief is not thereby justified: X's poor performance with regard to the few is not redeemed by its excellent performance with regard to the many. But process reliabilism appears to yield the verdict that such a belief is justified.

Goldman is not moved by this case. He presents a scattershot of possible responses. Two of these we can immediately set aside. First, Goldman (in effect) points out – quite correctly – that (c) only follows from (a) and (b) in S if we make an additional assumption, such as:

1. d. X tends to output beliefs in a given proposition in R₁ about as often as it tends to output beliefs in a given proposition in R₂.

Moreover, he doubts that (d) is true of the case as I describe it. Goldman (Reference Goldman2015: 141) writes, ‘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’. But this is at most an objection to my particular attempt to describe a case with features (a), (b), and (c), not to the very possibility of there existing a case with that structure. So let us stipulate that instead of the process being one that yields belief in propositions of the form <n is not prime> for arbitrary values of n, it only yields belief in propositions of that form for values of n that are at least three digits.Footnote ¹⁰ It is easy enough to imagine a case of that sort for which assumption (d) is true. Second, Goldman notes – quite correctly – that in many of his writings on process reliabilism he makes a distinction between belief-forming processes and belief-forming methods which my example appears to neglect. In Goldman's words:

Process reliabilism, Goldman insists, is only intended to be an account of the justifiedness of beliefs formed entirely on the basis of processes; for beliefs formed at least partially on the basis of methods, the account of justifiedness is allegedly more complicated, and in particular involves assessing whether the subject acquired the relevant methods in an appropriate manner.Footnote ¹² Moreover, the prime numbers case, as I described it, may seem to involve a consciously applied rule for forming beliefs about the primeness of natural numbers, and hence may seem to feature a belief-forming method, not a belief-forming process. But, again, this is at most an objection to my particular attempt to describe a case with structure S, not to the very possibility of there existing a case with that structure. We can easily change the case so that it features a subject with a ‘wired-in’ feature of her ‘native cognitive architecture’ that allows her to form beliefs about the primeness of natural numbers in the way specified. And when we do, I claim, that does not affect our judgments about the case: it is still implausible to hold that beliefs in the few formed on the basis of this process are justified.Footnote ¹³

Goldman has a third reply to the prime numbers example which is more promising than the other two.Footnote ¹⁴ This reply involves, in effect, denying that the process type specified in the example is an epistemically relevant one. Goldman notes that in his first article defending process reliabilism, he suggested that relevant process types should be restricted to those that are ‘content-neutral’ (Goldman Reference Goldman and Pappas1979, Reference Goldman2012: 39). This restriction was designed to rule out processes such as the following: inferring <p> from <The Pope asserted that p>.Footnote ¹⁵ And, Goldman insists, that same restriction rules out the process in my prime numbers case.

I am not convinced. I will now argue that depending on how we interpret the proposed restriction to content-neutral process types, either that restriction is extremely implausible (call this ‘the plausibility problem’), or it fails to rule out problem cases with the relevant structure (call this ‘the effectiveness problem’), or both.Footnote ¹⁶

It is very unclear what content-neutrality comes to. In his initial discussion of the content-neutral restriction, Goldman (Reference Goldman and Pappas1979, Reference Goldman2012: 39) defined content-neutrality as follows:

When content-neutrality is understood in this way, the resulting restriction on process types faces a plausibility problem. In particular, many of Goldman's own examples of relevant process types would be ruled out by this version of the content-neutral restriction, including ‘sight’, ‘hearing’, ‘ignoring contrary evidence’, ‘beliefs based on inference from observed samples’, and ‘accepting the reports of others’.Footnote ¹⁷ None of these examples involve input states with arbitrary contents. This first way of interpreting the content-neutral restriction also faces an effectiveness problem. Because it proposes no constraints on the content of output beliefs, the restriction fails to rule out the process type in the prime numbers case. (My description of that case is silent on the content of the inputs to the process described, if indeed those inputs even have content.) To avoid this last problem, we might switch to

This alteration helps with the effectiveness problem, but now we have a drastic plausibility problem. Pretty much the only process type that satisfies the content-neutral restriction under this interpretation is deductive inference.Footnote ¹⁸

Goldman (Reference Goldman2015: 141) has now changed his definition of content-neutrality to

Presumably ‘many different’ refers not just to the number of different contents – the process in my prime numbers case does very well by that count – but also to the diversity of those contents. Under this interpretation the content-neutrality restriction is less implausible (though perhaps misnamed: ‘content-flexible’ strikes me as a more accurate term than ‘content-neutral’ for the notion being glossed here). However, this version of the restriction is beset by an effectiveness problem. It is easy enough to tweak the prime numbers case so as to make process X content-neutral in the sense proposed: just keep adding extra propositions which are not of the form <n is not prime> into R₂ (the set of propositions that X always gets correct) until a desired level of diversity is achieved. This change does not, I claim, alter our verdicts about the case.

Thus I am not convinced by the three replies we have considered to my prime numbers example.Footnote ¹⁹ More generally, we can separate two issues with regard to that case. First, we may ask, ‘Does process reliabilism allow cases with structure S to exist?’ Second, we may ask, ‘If cases with structure S exist, would they be a problem for process reliabilism?’ Goldman seems to be conceding that the answer to the second of these questions is ‘Yes’. His strategy is to resist the example by attempting to answer ‘No’ to the first question. But this strikes me as a not-very-promising strategy. When R ranges over necessary propositions, cases with structure S are a dime a dozen – as we have seen, from the failure of Goldman's various replies to rule out cases with that structure. Moreover, if we let R range over contingent propositions and change ‘always’ to ‘almost always’ in clause (b), then cases with the resulting structure are also a dime a dozen. Finally, we can change ‘belief-independent’ to ‘belief-dependent’ in clause (a) and make accompanying changes in clause (b) to generate yet more cases with a problematic structure. So really structure S is a template for a vast number of ways in which a process type might forgo success with regard to certain propositional outputs in order to achieve great success with regard to many other propositional outputs. We can imagine cases of this sort involving not just mathematical processes, but also perceptual processes, introspective processes, inferential processes – the list goes on. I think it is striking how resistant we are to saying that such cases involve a ‘sacrifice for the greater epistemic good’, in the way in which an act-consequentialist might deem a cutting up of the one to save the five as a sacrifice for the greater moral good. This judgment of ours about the justificatory status of beliefs formed (or sustained) on the basis of processes with structures similar to S is not, I claim, merely a folk verdict, but rather is a persistent considered judgment that survives reflective scrutiny. And it is a judgment which Goldman seems to share. I conclude that Goldman has not successfully shown that process reliabilism avoids my charge that the view illicitly ignores the separateness of propositions.Footnote ²⁰