The Evaluation of Cognitive Systems [i]

I. Introduction

Cognitive system s structure thought. They encompass our patterns of inference as well as the types of evidence we require before forming a belief. Since patterns of inference and standards of justification influence how we interpret the world around us, cognitive processes reach into the core of how we think and perceive. We do not use one global system that guides our thinking in every situation; instead, we use multiple, localized cognitive systems as we are confronted with different tasks and needs. As this use is largely unconscious, we likely cannot choose to use a particular cognitive system and then simply begin using it. Cognitive systems underlie decision-making; they are not the results of decisions. We can examine the systems we use, but we cannot try out various systems to see which we prefer.

Richard Nisbett and Lee Ross undertook such an examination when they experimented with subjects to assess common inferential practices. The results provide insight into our typical patterns of inference, which can in turn be used to describe our cognitive systems. They found that we typically make basic errors in the detection of covariation, a basic relationship between variables, and we "often rely on poorly justified causal theories of questionable origin and place too much emphasis in even those explanations prompted by causal theories held with good justification" (Nisbett and Ross 1980: 137). People also "perform many prediction tasks quite poorly both in the laboratory and in everyday life" because of a lack of understanding of "fundamental statistical principles" (Nisbett and Ross 1980: 165). We are especially poor at theory maintenance and change: we will force evidence to fit currently held theories and will persevere in beliefs that new evidence plainly renders unwarranted. At first glance, these inferential practices are strikingly bad, but they raise important questions about what it is for such a practice to be good:

But the perseverance tendencies of subjects in these experiments were so extreme as to force consideration of the possibility that the traditional scientific standards may not apply. In particular, it seems possible that the behavior of subjects, inappropriate as it is from the standpoint of rationality in the inferential contexts studied, may arise from the pursuit of important, higher order epistemic goals. (Nisbett and Ross 1980: 191)

Even though the inferential patterns discovered in these experiments could be better at producing true beliefs, common reasoning is nonetheless effective in daily life. Perhaps there is some reason for these unexpected patterns of inference; they may, for instance, be more economical. Yet our initial reaction is to write them off as bad reasoning. How ought we to evaluate such practices? Do they represent bad or faulty cognitive systems? Even supposing we could make the patterns of inference more rational, we first need to know if we should do this. Are more rational cognitive systems better than the ones used by these subjects?

In The Fragmentation of Reason, Stephen Stich advocates treating cognitive systems pragmatically as tools and judging them with regard to their efficacy in achieving what we value. To reach this conclusion, he denies that true belief is valuable. When something is valued intrinsically, it is valued for itself: it is valued because it is what it is and not because of any effects it might have. Something's being instrumentally valuable, however, means that it is valuable as a means to some end. Stich insists that when we come to a proper understanding of the nature of truth, we cease to find it either intrinsically or instrumentally valuable.

I aim to show that he does not adequately support this conclusion about truth's lack of value and that his arguments for his standard of cognitive evaluation rely upon this conclusion. I first reconstruct Stich's account of truth before exploring its implications and his objections to the value of true belief. I then explain how all three objections rely upon faulty assumptions and how his core theory of truth conflicts with two of his objections. Because this unfavorable evaluation of his arguments casts doubt on his proposed standard for cognitive evaluation, I propose a new standard that relies upon a different argument for treating cognitive systems as tools.

II. Reconstructing Stich's Position

i. A theory of truth

Stich begins by explaining the nature of belief. He thinks that "beliefs are real psychological states," and he also "embraces the so-called token-identity hypothesis" (1990: 103). In contrast to theories that beliefs are explanatory constructs, Stich says they are real phenomena. But what type of phenomena are they? The token-identity hypothesis identifies beliefs with brain-states: these patterns of neurological firings are the real phenomena which we call beliefs. A brain-state cannot be identified with more than one particular belief, and the occurrence of a type of brain-state in a particular brain signals the possession of its equivalent belief. Because particular beliefs are identified with brain-states, one brain-state type cannot represent different beliefs in different situations.

Stich thus thinks that beliefs are real, physically-based phenomena, with each particular belief being identical to some particular neurophysiological state. A belief can therefore be thought of as a particular brain-state, and a different brain-state would represent either a different belief or no belief. This is a physicalist conception of mind that takes some states of the brain to be identical with beliefs, so that the mental is bonded to the physical.

The difference between those brain-states that are also beliefs and those other brain-states that are not beliefs is that "beliefs have semantic properties. They are true or false" (Stich 1990: 103-104). Yet how can a brain-state, a mere firing of neurons, be true or false? Stich answers this question by linking brain-states with truth conditions, which specify the circumstances under which a belief is true.[ii] Brain-states obtain truth-values through their connection to truth conditions, and brain-states that are not beliefs lack truth-values because they lack this connection.

The choice of truth conditions determines which beliefs are true and false. This important role opens up another puzzle: Why is a particular belief connected to one truth condition and not to another? How are the circumstances specified for a brain-state's being true? Stich cannot appeal here to the relation of truth conditions to truth since they determine whether a belief is true. Imagine a person having brain-state Belief1 this past Sunday. This brain-state is a belief because it is true or false, and its truth-value is determined by its connection to truth conditions. If it is connected to truth conditions that require that the current day be Sunday for the belief to be true, the belief would be true. If it is instead linked to truth conditions that require that the day be Wednesday, the belief would be false. Because these truth conditions define truth for the belief, one cannot choose between them by referring to truth. So why use one truth condition and not the other?

To resolve this issue, Stich posits an "interpretation function" that maps beliefs to truth conditions. A belief's truth-value is determined by certain truth conditions because these are assigned to it by the interpretation function; this mapping interprets brain-states into truth conditions. The interpretation function links brain-state Belief1 with some truth condition, and if this truth condition is that today is Sunday, Belief1 is true.

A function connects every element of a domain (beliefs) with some one element of a range (truth conditions), but each element of the range can be connected with more than one element of the domain. The function could link the truth condition that today is Sunday with Belief1 and with other beliefs, but Belief1 could only be linked to that one truth condition. The use of function implies that every belief is associated with one truth condition, but any truth condition may be associated with any number of beliefs.

A function may link Belief1 to the truth condition that today is Sunday, but the particular linkage still seems arbitrary. To clarify the connection between beliefs and truth conditions, Stich must explain the particular function we use. He calls his theory of our interpretation function the causal/functional theory, and he thinks it is "a familiar and justifiably popular one" (1990: 106). To provide truth conditions for all beliefs, it relies upon Tarski's theory of truth, by which the statement "Tom is standing on a beach" is true if and only if Tom is standing on a beach.[iii] A natural language sentence is one actually uttered (e.g. "Tom is standing on a beach"), and a meta-language sentence is one in which truth conditions are ascribed to a natural language sentence (e.g. Tom is standing on a beach). The meta-language statement specifying truth conditions for "Tom is standing on the beach" only works if there are axioms stating, for instance, that "Tom" refers to Tom. Such axioms are necessary for the relationship between the natural language sentence and the meta-language sentence to be clear, but unless there is a means of verifying or deriving them, these axioms become arbitrary connections. The Tarskian theory depends upon a list of axioms without offering an account of "what it is to get these axioms right" (Stich 1990: 108).

Stich turns to a Kripkean causal theory of reference to avoid this arbitrariness. [iv] He explains the references of words in terms of causal chains that link words to their meaning-bestowing initial baptisms. A word initially acquires meaning by, for instance, someone explicitly pointing it out as referring to a particular object. This meaning is then passed on through people's use of the word, so my use is connected to the referent in virtue of the causal chain stretching from my current use back to the initial bestowal of meaning. For example, a baby is born and her parents dub her Alice. My use of "Alice" is connected to this initial baptism through chains of communication: I may have seen the word on a birth certificate, or somebody may have told me the baby's name. Yet there are many causal chains that end in my use of "Alice." There is a causal chain connecting "Alice" to a picture in a book by Lewis Carroll, and there is another chain connecting the word to a phone directory in which I saw it printed. Because there are many causal chains linking words back to all sorts of objects and processes that we do not identify as their referents, only causal chains of the right sort fix the referents of our words. These causal chains can justify the axioms required by Tarski's theory. So if "Tom" refers to Tom via the right sort of causal chains, and if the components of "is standing on a beach" refer in their way to is standing on a beach, then "Tom is standing on a beach" is true if and only if Tom is standing on a beach.

Stich completes his account of the interpretation function by "putting the language inside the head" (1990: 109). By the Tarskian/Kripkean elements of the theory, words have referents and sentences have truth conditions. If the sentence is a belief, then, by the aforementioned token-identity hypothesis, the sentence is a brain-state. Imagine a computer designed to make truth-claims and that uses electron-states as its storage medium. Now imagine a string of electron-states, illustrating a sentence that is a belief and is stored as a brain-state. The computer first interprets the electron-states as meaningful data in a manner analogous to the assignation of reference through causal chains. The computer, in its Tarskian equivalent, next runs the interpreted data through a program to assign it truth conditions, and it then checks on these truth conditions to determine truth-values. The computer can therefore verify that it only encodes true beliefs, and it encodes them in an electron script that does not directly have semantic value. Natural language sentences are associated with beliefs, and because Stich identifies beliefs with brain-states, we can be said to store these sentences in some neurological script in the brain. The interpretation function's role is to specify truth conditions for these "cerebral inscriptions" (Stich 1990: 109). Belief1 is interpreted by identifying the causal chains behind its components, and this meaning of "Today is Sunday" connects it with the truth condition that Belief1 is true if and only if today is Sunday. The interpretation function fixes the reference of a belief through causal chains and then connects the belief with the appropriate truth conditions.

ii. The first implication – Limited domain

The first important implication of Stich's account of truth is that the causal/functional interpretation function does not map all possible brain-states on to truth conditions. Returning to the computer analogy, imagine electron-states that are unreadable: they are similar to the electron-states that acquire truth conditions, but these cannot be interpreted into data and so do not have truth conditions. There can also be brain-states which are similar to beliefs, but which lack truth conditions -- both these brain-states and beliefs are "belieflike," but only the latter have truth conditions. The causal/functional interpretation function does not operate over all of the belieflike brain-states, for "the belieflike mental states [which Stich uses interchangeably with brain-states] for which it provides a specification of truth conditions constitute a small subset of the possible belieflike mental states that a human or other organism might have" (Stich 1990: 110-111). While this interpretation function does offer truth conditions for brain-states that are beliefs, there are similar brain-states for which it fails to provide such conditions.

Stich says this limitation is due to its causal theory of reference and to its formal requirements. By the Kripkean theory of reference, a given word is connected to its referent via a causal chain stretching from the word's initial baptism to the word as it stands in the lexicon of the current speaker. Because there are many different causal chains connecting a word to different things, the reference-fixing causal chains must be of a particular type. So, Stich's argument runs, because a word is tied to the world in many different ways, a word could end up in the speaker's lexicon while lacking the type of causal chain that the causal/functional interpretation function takes as fixing reference. For example, the word "Bish" may have entered my lexicon by my seeing a newspaper's misprint of the caption under a picture of a world leader. My use of the word can be accounted for by the causal chain connecting it to the misprint, but this causal chain is not of the Kripkean type, since it lacks an appropriate initial baptism at its foundation. Because "Bish" does not have any Kripkean causal chain by which it gains a referent, "Bish" does not refer when I use it. Stich thinks that Kripkean causal chains are not necessary for a word to end up in a person's lexicon. Because they lack the appropriate sorts of causal ropes, these hypothetical non-Kripkean words do not refer, and the causal/functional interpretation function will not assign truth conditions to sentences which contain these words. Any belief of mine that uses "Bish" will be neither true nor false since it will be meaningless. The domain of this interpretation function is restricted to those words that are tied to the world via the proper sort of causal chains.

The other limitation to this interpretation function stems from its formal requirements. This function only assigns truth conditions to certain types of syntactical relations. Relations such as conjunctions and counterfactuals have "patterns of inference" which are "intuitively logically permissible" (Stich 1990: 113): a conjunction is a syntactical relationship that leads to predictable inferences which we intuitively understand. Stich says there are other syntactical relationships among sentences that are not so easily understood:

There are indefinitely many possible patterns of formally specifiable causal interactions among mental sentences and thus indefinitely many possible mental sentence constructions, which admit of no intuitively plausible semantic interpretation at all. Most purely formal, syntactically characterizable patterns of interaction among sentences or well-formed formulas have no intuitively plausible semantics. (1990: 113)

Many relations can be derived from syntactical rules, but they may not be explainable as true or false. Stich offers the example of the construction 'p {} q', for which two premises 'p' and 'q' are related by 'p {} q' if and only if the first contains more conjuncts than the second, the second contains fewer disjuncts that the first, or both begin with certain abstruse symbols (1990: 113). We know how to derive 'p {} q' from a pair of premises, but it is not clear what it means for 'p {} q' to be true or false. Given the two premises "The dog ate my homework or I forgot it" and "I swim well," we can derive "The dog ate my homework or I forgot it {} I swim well," because "I swim well" contains fewer disjuncts. But how can we interpret this derivation in terms of truth conditions? Because the axioms necessary for the meta-language sentence to make sense of the connective {} are not as easily created as for a conjunction, it does not make much sense to say, "'The dog ate my homework or I forgot it {} I swim well' is true if and only if the dog ate my homework or I forgot it {} I swim well." The truth-value of this syntactic construction cannot be explained in terms of the causal/functional interpretation function. This function's Tarskian basis limits its domain to connectives whose semantics are intuitively understandable.

The causal/functional interpretation function therefore has a limited domain: many belieflike brain-states are not assigned truth conditions because of the function's restrictions on how a word refers and on which connectives are allowable. This limited domain implies that there are many computational systems whose components lack truth conditions. Stich observes that these systems may be more helpful than ones we currently use:

No doubt many of the systems in this semantics-free space are useless and chaotic. But there is certainly no reason to suppose that they all are. A much more likely possibility is that in this huge space there are systems that would vastly increase their user's power or happiness or biological fitness, systems that would lead to substantial reductions in the amount of suffering in the universe, and systems that would significantly reduce the probability that we will bomb ourselves into oblivion along with much of the biosphere. (1990: 118-119)

Truth-based cognitive systems do not utilize many belieflike brain-states, but cognitive systems that are not so tightly linked to truth can utilize them. Some of these cognitive systems may be substantially more beneficial than the truth-based ones we use.

iii. The intrinsic value of truth – Objection 1

Because these other cognitive systems are not concerned with truth, valuing truth creates a reticence in people to consider them. Regarding our pursuit of these other cognitive systems, Stich says, "Those who would accord intrinsic value to the holding of true beliefs may well be reluctant to explore that vast space and will resist adopting what may be found, since we know in advance that it contains no true beliefs. But theirs is a profoundly conservative normative stand" (1990: 119). Even though these other systems may have effects that better accord with our values such as happiness, those who value true belief for its own sake will not want to explore these semantically-free systems, because they by definition will not lead to true beliefs. The belieflike states these systems emphasize are neither true nor false.

Calling this "a profoundly conservative normative stand," Stich admits that some extreme traditionalists may actually find their valuation reinforced by the discovery of truth's conservative nature (1990: 120). He also, however, believes that "there are many people, and I am among them, who are not much inclined to value what is traditional and familiar for its own sake in matters epistemic" (1990: 120). Because clinging to the value of truth precludes the possibility of advancing other values through non-semantic cognitive systems, Stich thinks the only reason to continue to value truth is an excessive valuation of tradition.

iv. The second implication – Idiosyncrasy

In addition to having a limited domain, the causal/functional interpretation function is also "highly idiosyncratic" (Stich 1990: 114). This function maps beliefs on to truth conditions, so Belief1 is mapped on to the truth condition that today is Sunday and Belief2 is mapped on to the truth condition that today is Wednesday. This mapping can be switched, however, so that Belief1 is true if and only if today is Wednesday and Belief2 is true if and only if today is Sunday. Mapping beliefs to truth conditions in one way means that they can be mapped in many other ways (Stich 1990: 114). If every belief is connected to a truth condition, these connections can be switched around to create different mappings. The causal/functional interpretation function does not provide the only possible mapping.

One way to differentiate between these alternative mappings is to focus on the causal chains used to fix reference. The causal/functional interpretation function upon which we rely only respects Kripkean causal chains as reference-fixing, yet "what ties all these [Kripkean] causal chains together is not any substantive property that they all share. Rather, what ties them together is that commonsense intuition counts them all as reference-fixing chains" (Stich 1990: 115). Aside from my using them to fix reference, the chains connecting my use of "Alice" with the naming of the baby and my use of "Bush" with the naming of the president do not have a common property not shared by the chains connecting "Alice" and "Bish" with the pictures in the Carroll book and the newspaper. The only difference between them is that we do indeed use the first two chains to fix reference. Our only reason for using these is that our intuition identifies them as the proper types of causal chains for fixing reference.

Many different non-Kripkean causal chains may be used by other reference schemes, "and the only obvious complaint to lodge against many of these alternative schemes is that they do not happen to be the scheme sanctioned by our commonsense intuition" (Stich 1990: 115). The only feature that distinguishes causal chains sanctioned by the Kripkean account of reference from other chains is that this version of reference happens to be the one we use. This, in Stich's view, makes valuing the results of the causal/functional interpretation function idiosyncratic.

It is quite difficult to construct a non-Kripkean reference relation that we might plausibly employ. Perhaps this lack of plausibility indicates that Kripkean causal chains have a greater family resemblance than Stich allows them. One example of an alternative reference relation emphasizes the causal chains connecting proper names with the people who introduced them to us. So, for instance, some parents name their baby Alice, and when I meet the family a few years later, the mother introduces her child to me as Alice. Under the Kripkean reference scheme, my use of "Alice" will refer to the child, but under this alternative reference scheme, my use of "Alice" will refer to the mother. If, in the future, I believe that Alice eats cereal, the truth-value of the belief may depend on which reference relation I use. The truth condition arrived at via the Kripkean reference relation will make my belief true if and only if the child eats cereal, but the truth condition led to by this alternate reference relation will make my belief true if and only if the mother eats cereal.

I will call such alternate reference relations reference*, and they explain how words refer*. To differentiate the causal/functional theory's reference relation from reference*, I will designate it by reference^. Reference* leads to truth* conditions, and reference^ leads to truth^ conditions. For Stich, the result of the causal/functional interpretation function is the assignation of truth^-values to beliefs to make them true^ or false^, and the results of other interpretation functions than the one we use are truth*-values for beliefs to make them true* or false*. I will henceforth use "truth," "falsity," "reference," and their cognates as purposefully vague notions, in contrast to the marked words that indicate particular interpretation functions and their results.

With reference* as defined in the above example, should we use the causal/functional interpretation function, or should we use the interpretation function through which words refer*? Stich does not think either function has any distinctive features that should lead us to regard one or the other more highly. In fact, he believes that the only significant difference is that reference^ is the relation we happen to use. Truth^ and truth* differ only in their reference relations, and because there is no significant difference between the relations, a preference for truth^ must be idiosyncratic. One who values truth^ would do so solely because the causal/functional interpretation function is the one tradition leads us to use.

v. The intrinsic value of truth – Objection 2

This idiosyncrasy leads Stich to again designate intrinsically valuing truth^ as a conservative position. He says that our use of the Kripkean reference relation is not the result of an informed decision:

Whatever the explanation [for why we sanction this particular interpretation function], it is clear that our intuitions do not result from a systematic and critical assessment of the many alternative interpretation functions and the various virtues that each may have. One way or another, we have simply inherited our intuitions; we have not made a reflective choice to have them. (1990: 120)

Most people who currently value true^ belief never consider the other options. In a world with many alternatives to truth^, we ought to decide which is best for us rather than simply accepting the one we inherit. Yet people who reflect on the options and still value truth are not much better, for "they are making a profoundly conservative choice; they are letting tradition determine their cognitive values without any attempt at critical evaluation of that tradition" (Stich 1990: 120). The existence of equally commendable alternatives reduces the attractiveness of sticking by the intrinsic value of truth^. Were we to reflect on all the options for interpretation functions and to choose one, we would not have an obvious means of choosing one because of their lack of distinguishing features. Stich faults those who currently value truth^ for doing so unreflectively, and the value of truth^ is lessened for those who do reflect on it because the function that provides truth^ conditions is merely one pedestrian function among many.

vi. The instrumental value of truth

In attempting to overturn the value of true belief, Stich must also consider the claim that it has instrumental value. We use the causal/functional interpretation function to arrive at truth^-values, but there are many possible interpretation functions, through which we aim at true* beliefs, true** beliefs, and the like. Stich uses this insight to raise the bar for the advocates of the instrumental value of truth^. Not only must they show that true^ beliefs are more instrumentally valuable than false^ ones, but these beliefs must also be measured against true* beliefs, true** beliefs, and their kin. A simple comparison with falsity^ will not suffice, because there are other options that are neither strictly true^ nor false^. Some true* beliefs may also be true^, others may be false^, and still others may have no truth^ conditions and so will be "neither true[^] nor false[^]" (Stich 1990: 121). To be instrumentally valuable, truth^ must be better than these other options at promoting valuable ends.

Stich says that true^ beliefs do not always represent "the best way to achieve our more fundamental goals" (1990: 122). As an example of a case in which true* belief would have been more instrumentally valuable, he describes a person dying in a plane crash. If he had possessed a true* belief about the departure time [where true* belief agrees with true^ belief entirely except that in this one case the true* belief is false^ (Stich 1990: 123)], this person would have avoided the crash and so saved his life. From this story, Stich concludes that "true beliefs are not always optimal in the pursuit of happiness or pleasure or desire satisfaction, nor are they always the best beliefs to have if what we want is peace or power or love, or some weighted mix of all of these" (1990: 123, italics added). Sometimes it would be better to have false^ beliefs which are true* than to have true^ beliefs. Stich knows of no argument that attempts to show that true^ beliefs do a better job than these other options "in general, or in the long run" (1990: 123). Since we know that truth^ is not always preferable, he thinks it a difficult task to show that it is even generally preferable to all other options. Since some other options may be more efficacious, Stich thinks it doubtful that truth^ has instrumental value.

vii. Universalizing the attack on truth

Stich freely admits that the above objections to the value of true belief have been presented against "the background assumption that the causal/functional interpretation function is the right one" (1990: 124). He must now attempt to demonstrate the applicability of his arguments to any plausible account of truth, so that the strength of his position does not depend upon the adequacy of his particular description of the casual/functional interpretation function. He wants to show that any alternative account must share the key features that overturn the value of truth for the causal/functional theory. His two arguments against the intrinsic value of truth follow from the limited domain of the causal/functional interpretation function and from the idiosyncrasy of valuing it, and his attack on the instrumental value of truth relies upon the efficacy of alternatives to truth.

For an account of truth^ to be plausible, Stich requires that it meet our commonsense intuitions about truth; any theory that does not meet these intuitions is deemed implausible. Yet our commonsense intuitions say nothing about many possible cognitive systems which depart radically from anything yet imagined. Also, there are infinitely many syntactically possible connectives about which our intuition is totally silent. Since a plausible account of truth is restricted to our intuitions, Stich concludes that it must share the limited domain of the causal/functional interpretation function. This limited domain leads to his accusations of conservatism and thus overturns the intrinsic value of truth for any version of the interpretation function.

The idiosyncrasy of valuing the results of the causal/functional interpretation function derives from two factors: that cultural influences play a significant role in determining which function we use, and that there are many alternatives to it. Because the culturally inherited notions that lead us towards a given interpretation function are only sanctioned by "tradition" (Stich 1990: 126), the function chosen is also only sanctioned by tradition. He thinks it unlikely that it has any distinguishing feature or that it was chosen on the basis of such a feature. Because there will always be alternatives to any interpretation function, and because the retention of our inherited function would only be guided by culturally inherited notions regardless of the specific description of the function, the idiosyncrasy implied by the causal/functional theory is generalizable to any plausible account. The resultant attack on the intrinsic value of truth is therefore also generalizable to any plausible account of truth.

Regarding his objection to truth's instrumental value, the first key point is that true^ belief is instrumentally valuable only if it is more efficacious than false^ belief and all versions of true* belief. Truth* and its ilk are developed through alternative accounts of reference, and there will be many alternative functions providing alternative versions of truth regardless of whichever particular function best accords with intuition. Any account of the interpretation function would still require one to compare truth with truth*, and this leads to the downfall of true belief.

Stich's other key point is that there are circumstances in which true^ beliefs are not the most efficacious: true* beliefs sometimes promote survival better. Because "for any interpretation function to be in the running, it must match up with commonsense intuition, at least most of the time," the example already described of the person catching the doomed plane would hold under any plausible interpretation function (Stich 1990: 125). However the interpretation function that we use is explained, the traveler's belief must be true, for he believed that the plane would leave at the time it did in fact leave. A plausible account of the interpretation function that we use could not stray far enough from our intuitions to preclude the construction of such counterexamples that provide instances of truth serving us more poorly than would truth*.

viii. Epistemic pragmatism

The success of Stich's objections to the value of truth is crucial to his proposal for the evaluation of cognitive systems. He uses truth's lack of value to reject other proposals, and this in turn enables him to present his own. The first proposal is that of analytic epistemology. In brief, for Stich, analytic epistemology would seek to analyze commonsense epistemic notions in order to elicit "criteria of rightness" that establish right "justification rules" (1990: 91-92), and these justification rules would determine which cognitive systems are good. He first argues against the intrinsic value of such a "right" cognitive system on the basis of the cultural nature and variability of our commonsense notions. Stich then must overturn such a system's instrumental value. To do this, he feels that he must reject the value of true belief, so he can thereby counter any appeal to the ability of these cognitive processes to generate true beliefs.

From the alleged collapse of the solution of analytic epistemology, Stich concludes that cognitive systems should be evaluated on whether "the consequences of employing one alternative or the other will lead to something we value" (1990: 130). The "obvious" position suggests that cognitive systems are good insofar as they produce true beliefs and avoid falsehoods (Stich 1990: 130). Yet because "a consequentialist account must take as the relevant consequences something that people actually value" (Stich 1990: 131), true belief's lack of value means that it cannot be the ultimate criterion.

Because Stich is concerned with consequences, he proposes to view cognitive systems pragmatically as tools. We ought to evaluate these systems in terms of the agreement of their consequences with our values. As an admittedly problematic example, he describes a cost-benefit analysis in which we assign values to the systems' outcomes and choose the ones that maximizes the expected value (1990: 133-134). The evaluation of a cognitive system is relative to the user (Stich 1990: 136); it is relative to "the environment in which the system is operating" (Stich 1990: 136); and it is relative to "the purpose of inquiry" (Stich 1990: 155). The judgment of cognitive systems depends upon what the prospective cognizer values, the circumstances in which she will be using the systems, and her objective in making the decision.

Stich labels his position "epistemic pragmatism." He does not see cognitive systems as valuable in themselves; instead, they are valuable as means to ends. To be better than another option, a cognitive system must be more efficacious in promoting desirable ends. Cognitive systems are tools whose normative status depends upon their ability to produce effects that cohere with our values.

III. Evaluating Stich's Position

i. Evaluating the first objection to truth's intrinsic value

The limited domain of the causal/functional interpretation function gives rise to Stich's first objection to the intrinsic value of truth. Many cognitive systems that would benefit us may not be concerned with truth. Because valuing truth makes us disinclined to explore these other cognitive systems, it is a conservative position. This argument relies upon a correlation between valuing and consideration: Stich assumes that valuing truth means that we will only consider cognitive systems that employ truth. He leaves no room for the possibility that one can consider the other benefits of a non-semantically-based cognitive system while valuing truth. This move requires the assumption that if truth is valuable, it must be of highest value in all circumstances. In this case, the possible benefits of these other cognitive systems would never be of sufficient weight to distract us from the pursuit of true belief. If the value of true belief is not of this highest type, however, supporting it is not incompatible with favorably evaluating non-truth-based cognitive systems. We could value truth while approving of cognitive systems that do not use truth. Only if we value truth in this highest manner will we demand that a good cognitive system utilize truth regardless of whatever other benefits it may bring.[v]

Stich believes that most people would cease to value truth if they understood its nature. In this first attack, he argues that people's recognition that they are ignoring beneficial cognitive systems would cause people to cease to value truth intrinsically. Stich's argument is unsound, however, because of the key assumption. Rejecting the assumption, and so denying that true belief must be of overriding value, would not result in the conservative stance described by Stich. There is no reason why the value of true belief could not be occasionally outweighed by the value of survival, of health, or of pleasure. These represent important motivations that may also be ingrained in the human condition. If true belief can be comparable to some other values, then one who values it need not be so reluctant to consider non-semantic cognitive systems.[vi] Stich requires a false disjunction between granting truth highest value or no value, and the failure of this disjunction leads to the failure of his argument.

Stich could reply that valuing true belief does not prevent us from considering non-truth-based cognitive systems, but it precludes us from such consideration. As truth-valuers, we want to have true beliefs about cognitive systems when we evaluate them. Working from within the context of a truth-based cognitive system, however, we know that we can never acquire true belief about these systems. They are too foreign to our own thinking to be accurately conceptualized. Valuing truth therefore prevents us from ever seriously evaluating these cognitive systems because we can never feel up to the task.

This difficulty in conceptualizing these systems also may indicate their illusory nature, however. It is indeed quite difficult to imagine what such a cognitive system might involve, and as will be argued later, true belief is necessary for proper evaluation. Ceasing to value true belief altogether might cause us to have the false belief that a non-truth-based system causes unhappiness, and this false belief might change the outcome of our evaluation. Furthermore, even if these cognitive systems exist, it is doubtful that we could intentionally adopt one, so the issue of whether we would consider such systems is likely of little practical importance.

ii. Evaluating the second objection to truth's intrinsic value

Stich's second objection relies upon the idiosyncrasy of valuing truth^. Because the reference relation used by truth^ is only option among many, truth^ is also one option among many. Since truth^ lacks any distinguishing feature other than our traditional use of it, valuing it is an idiosyncratic activity. If we reflect on our preference for truth^, we will cease to find it valuable because we will recognize it as one equally commendable option among many.

Stich claims that truth^ and each of its alternatives lack any special feature, whether positive or negative. If the value of truth is to be lessened by this knowledge, it must be due to truth^'s being one undistinguished option amid several. Yet something can be valuable in the midst of many alternatives; the mere presence of equally commendable alternatives does not detract from the value of any one alternative, unless its value was based on its being the only option. [vii] As an example, suppose I could buy many different brands of bottled water with nothing special or differentiating about any one. Their price, volume, and ratio of hydrogen to oxygen are all equal. I value the brand of water on my desk because it quenches my thirst and tastes much better than any other accessible water source, but I did not consider all the other options before valuing it. Does this lessen the value of this one bottle? And suppose that I consider the other options and find nothing special about this one. It is just as good and just as bad as any other brand. Does the existence of other equal options cause this bottle to lose value in my eyes? Not necessarily. Unless my valuation was solely based on the apparent uniqueness of this brand of water, my appreciation of it will not change, and because Stich claims that our valuation of truth is unreflective, he must not think that it is based on such a similar misperception.

The presence of other options is not a sufficient reason to lessen the value of one option, unless that value was based on a false belief about the option's uniqueness. Stich might respond that the value of truth would be of a different sort than the value of a water bottle. This may be right, but the effect of other options on its intrinsic value should be analogous to the water bottle situation. Stich must show that our valuation of truth is predicated on its being unique. Idiosyncrasy does not alone imply a lack of value. Even if we grant him the idea that other options exist, Stich still does not succeed in arguing for truth's lack of intrinsic value.

iii. Evaluating the objection to truth's instrumental value

This argument also relies upon the presence of alternatives to truth^. Stich says that to establish truth's instrumental value, one must show that truth^ is more effective than all other options at promoting that which we value. Since it is easy to construct examples in which true* beliefs are more beneficial than true^ beliefs, true^ beliefs are not always better, and the burden of proof is on the defender of truth^ to show that they are more effective in general.

By insisting that we compare truth^ with all of the other options such as truth*, Stich assumes that to be instrumentally valuable is to be optimally instrumental. He demands that truth^ not only be more effective in achieving our ends, but that truth^ be more effective than any alternative. [viii] Without this assumption, he would not be able to object to truth^ by showing an alternative to be more effective. He would instead have to show that truth^ is not effective at achieving our ends, but Stich only compares truth^ with other options and under special circumstances.

Stich's assumption about instrumental value means that if Harry's microwave is instrumentally valuable to him as a means to cook popcorn, and if he finds that an air popper that he also owns is better suited for this task, then he will cease to value the microwave instrumentally as a means to cook popcorn. Plainly, however, the microwave need not be the best option to be valuable in this manner; it need only be a good option. Contrary to Stich's assumption, Harry would not be unreasonable in still valuing the microwave as a tool to cook popcorn, even though he no longer uses it for this task. It worked rather well while he used it, and finding a better option does not erase all its value. Again, while the value of a microwave is not that similar to the value of truth, the effect of alternatives on its instrumental value should be analogous.

If to be instrumentally valuable is to be optimally instrumental, the burden of proof is on the truth^-valuer, who must show that truth^ is better than any other option. If such optimality is not required, the burden of proof is on Stich, who must show that valuing truth^ is not a good option, irrespective of other options. While we may value something more highly if it is more effective than all other options, we do not require this greater efficacy for it to have any instrumental value. Because Stich attempts to show only that truth^ is not the most effective option, he fails to raise significant doubts about truth^'s having any instrumental value.

iv. Problems stemming from the base account of truth

I will now argue that Stich's basic account of truth causes problems for two of his objections. My goal is to show that even if we grant Stich that the counter-arguments just advanced against his particular objections do not work, he still does not establish truth's lack of value, because the implications of his underlying account conflict with two of his arguments.

Stich's universalization of his attack on truth does not extend to radically different accounts of truth that rely on different base assumptions about the nature of belief; instead, it extends only to different accounts of the specific nature of the interpretation function. He wants to show that even if the causal/functional theory does not describe the interpretation function we use, his objections to truth still stand. Throughout, he preserves the realist claim about beliefs, the token-identity hypothesis, and the need for an interpretation function. He argues "that any plausible alternative story about the interpretation function is likely to share those features" that lead him to deny value to truth (Stich 1990: 124, emphasis added). He specifically generalizes his argument to any plausible story about the interpretation function, for such a function seems to be necessary for a plausible story about truth. He never considers accounts of truth that do not utilize such a function.

For Stich to have a basis for his objections to truth's value, every plausible account of truth must preserve two key points: the existence of other functions and the limited domain of the function. The existence of alternative mappings and truth conditions leads both to the need to compare truth^ with truth*, which is crucial to his objection to truth's instrumental value, and to the idiosyncrasy of valuing truth, which results in his second objection to the intrinsic value of truth. The limited domain of truth is necessary for Stich's first objection to its intrinsic value. These two points are critical for the success of his objections, but I will argue that his own constant base account provides reasons to doubt the first of the twin tenets.

Stich's theory creates a divide between truth conditions and beliefs, and he uses the interpretation function to bridge it. The interpretation function determines the meaning of beliefs and then supplies the proper truth conditions. Intuitively, this makes sense, since the truth of a belief obviously depends upon its meaning, which, in turn, depends upon the referents of its constituent terms. Changing the reference function used for elements of a belief may change the meaning of the belief, and the truth conditions for the belief will also change appropriately.

Let us return to Belief1, an uninterpreted brain-state. Suppose functionA, an interpretation function, interprets it as meaning "Jane is reading a book," and suppose functionB interprets it as "The cat is sleeping." Belief1 remains the same brain-state in each scenario, but with functionA it is true if and only if Jane is reading a book and with functionB it is true if and only if the cat is sleeping. The truth conditions differ depending on the reference relation used. Yet these are not, properly speaking, conditions of truth^ versus truth*. Because the Tarskian condition that "'x' is true if and only if x" does not change, the relations expressed in the truth conditions do not change. The truth conditions change because the input changes, but the same type of relation is used for the input in both cases. These conditions are not conditions of alternate versions of truth. Unless there is a radical change in the Tarskian relation used to establish truth, there is no reason to associate the products of these different truth conditions with anything but truth. If the reference of "cat" in "My cat ate her dinner" changes from my cat to my cousin under different interpretation functions, the truth conditions of the sentence will be different. They are still conditions of truth, however, since if both interpretation functions provide the same meaning for a belief, they will also provide it with the same truth conditions. They differ only in their assignations of meanings, not in their notions of truth; Stich never gives an example in which the basic idea of truth varies. He uses terms like truth* to describe alternate versions of truth that are really only alternate versions of reference.

Because Stich's own underlying theory counters the justification for the comparison of truth^ with truth*, the alleged fateful idiosyncrasy of truth vanishes. Valuing the particular reference relation of an interpretation function may well be idiosyncratic, but valuing truth is not. Because all of these interpretation functions result in truth, they no longer divide it into many options. One need also not worry about comparing the instrumental efficacy of truth^ with that of truth* and all other options in order to assess instrumental value. Ttruth* and its kin are only misleading names for different reference relations that all help determine the truth-value of a belief. One can compare the efficacy of a reference relation to that of other reference relations, but they all lead to truth or falsity, not to some variant. The effectiveness of truth need only be compared with that of falsity, and as will be seen, this makes the defense of the instrumental value of truth a much easier task.

Stich's first objection to truth's intrinsic value does cohere with his underlying account. Beliefs contain implicit truth-claims, and there may be brain-states that are not beliefs and that do not make truth-claims. A cognitive system that utilizes only this non-semantic type of brain-state would indeed, as Stich claims, not involve truth. While the feasibility of such a cognitive system is unknown, and while the moves he makes from this point are questionable, the objection does not rely on anything that contradicts his core theory of truth.

Stich attempts to universalize his attacks so that they hold for other accounts of the interpretation function, but he does not abandon the account of truth that postulates the function. I have shown that the core relation established between beliefs and truth conditions implies that his distinction between truth^ and truth* is spurious. This distinction is crucial to his explanation of the idiosyncrasy of valuing truth^ and of the need to compare the efficacy of truth^ with that of truth*. We can now reject both the idiosyncrasy and the need for total comparison. Since this second argument against the intrinsic value of truth and his only argument against the instrumental value of truth require these claims, both attacks fail. The underlying account does not, however, provide a means of countering his claim of a limited domain for interpretation functions. The counter-arguments given in the above evaluation of Stich's first objection to the intrinsic value of truth must hold if this objection is to be overcome.

v. Implications for his larger project

Stich attempts to persuade the reader to forswear the intrinsic and instrumental value of truth, but his arguments rely on faulty assumptions. These assumptions concern the importance of truth's value relative to other values; the effect of alternatives upon intrinsic value; and the necessity of optimality for instrumental value. Furthermore, his core theory contradicts two of his objections to the value of truth. The counter-arguments presented do not establish the value of truth or the lack of such value, but they do show that Stich's analysis is off-track.

The failure of his attack on the value of truth means that his larger argument for epistemic pragmatism also fails. As already described, this argument twice relies upon the conclusion about truth. The first place is in his argument against analytic epistemology: without his conclusion about truth's value, Stich does not reject the instrumental value of the cognitive systems that analytic epistemology would allegedly endorse. The second place is when Stich considers which consequences are relevant to the evaluation of cognitive systems. Because "truth, or true beliefs, is the obvious answer" (Stich 1990: 130), he only moves on to develop his version of epistemic pragmatism after appealing to his attack on the value of truth to overthrow this truth-linked position. His argument for epistemic pragmatism revolves around the dismissal of the solutions offered by analytic epistemology and by truth-linked accounts. His dismissal fails, however, because it relies upon his rejection of the value of truth. This failure to eliminate other alternatives destroys the means by which he arrives at his final position. Stich offers the reader little reason to prefer his solution to those of analytic epistemology or truth-linked accounts.

IV. The Evaluation of Cognitive Systems

i. Cognitive systems as tools

As described in the introduction, cognitive systems structure thought, encompassing patterns of inference, standards for justification, and even differences in information processing and retrieval. They are necessary for orderly thought and hence for conscious action and decision-making. They are even essential for the regular drawing of conclusions from data: without utilizing inferential strategies, we cannot systematically arrive at beliefs about our world. As conditions of orderly thought, cognitive systems help achieve the goals of thought. They are valuable for their effects in thinking.

We utilize cognitive systems because we think, and indeed we can utilize cognitive systems only if we think. We think for many reasons, some of which are practical. Cognitive systems, the facilitators of thought, share these reasons for use. These systems are instrumentally valuable because they enable thought and because they help achieve the motivating causes for thought. They can therefore be thought of as tools and judged according to their utility.

Stich agrees that cognitive systems "should be thought of as something akin to tools or technologies or practices that can be used more or less successfully in achieving a variety of goals" (1990: 131). This is the point from which he launches his epistemic pragmatism, but we have arrived here via a path which does not depend upon the rejection of other positions. This argument avoids the pitfalls of Stich's argument and its requirement of a lack of value for truth. We ought not yet, however, follow him in rejecting truth-linked accounts that view cognitive systems as tools specifically for the achievement of truth.

ii. The standard for evaluation

If cognitive systems are tools, what are they measured against? Is there some end common to all cognitive systems? Many tools have a specific function and are judged with respect to that function. A shovel, for instance, may be used to dig holes and so can be compared with other shovels on how well it performs this task. Similarly, a cognitive system's main function may be a specific task. Yet they are only being called tools analogously in order to portray their role and value in human life. They are not developed for a specific task as a shovel might be; they play a crucial role in all human life. This special difference leads to the alternate possibility that they are used to promote that which we value generally. So cognitive systems are either tools for specific values, or they are tools for our general values.

As a representative example, let us suppose that the promotion of truth and the avoidance of falsity are the ends towards which cognitive systems are specifically geared. Under this scenario, cognitive systems are good insofar as they reliably produce true belief. This is a truth-linked standard of evaluation, and it is one to which Stich is hostile. The danger here is that in viewing cognitive systems as processes that we utilize primarily to develop true beliefs, we are converting the aims of certain practices into a generalized standard. We use cognitive systems in much of our daily trivia, and many of these uses aim at something other than truth. Truth is not the only goal of thought. In deciding what to eat, I am primarily using my cognitive system to appease my hunger and to promote my health, not to generate true beliefs. In making inferences while playing sports, I primarily aim at winning and enjoyment: having true beliefs may facilitate these outcomes, but the truth of my beliefs is not my main goal. Similar examples could be constructed with deciding whether to bike or walk and with reading fiction. This type of argument could be used against any value that is proposed as the specific end of cognitive systems. To evaluate cognitive systems solely with regard to any one function, such as the generation of true belief, is to fail to recognize these systems' versatility and multiplicity of functions. Cognitive systems are geared towards many different endeavors, not to any one end.

As stated above, cognitive systems are either tools for the promotion of some specific values or they are tools for the promotion of what we value generally. Because we do not use them mainly for the promotion of one value such as truth, cognitive systems are only analogous to other tools. Their main role is to promote that which we value generally; there is no one value by which they can be evaluated in all their applications.

To engage in proper evaluation, one requires accurate information. If I am deciding which type of bread to buy and I value my health and money, I might decide to buy the cheapest whole wheat bread I can find. If, however, I have false beliefs about its impact on my health or about its price, I may choose a loaf that does not really accord with my values. Despite my intentions, my lack of true belief will have caused me to fail with regard to the aims of my decision. I should therefore value true belief to increase the effectiveness of my evaluations.

The project of evaluating cognitive systems likewise contains an implicit valuation of true belief. Hilary Kornblith claims that "Stich's view fails" because "the presuppositions of cost-benefit calculations undermine Stich's attempt to turn this into an account of cognitive evaluation" (1993: 369). Cost-benefit calculations assign values to outcomes and then choose the outcome with the highest expected value. Kornblith argues that truth plays a valuable role here: without knowing the truth about what we value, cost-benefit analyses will not arrive at the desired outcome. If the goal of such a cost-benefit analysis is to promote our values, having false beliefs could easily lead one to fail in the task by feeding the cost-benefit analysis incorrect data. Kornblith concludes that true belief is valuable for pragmatic reasons because of its role in decision-making:

Allowing our cognitive systems to be determined by the totality of our interests exclusive of truth thus undermines our ability to make choices, outside the cognitive realm, which are conducive to those very interests . . . . It seems that someone who cares about acting in a way which furthers the things he cares about, and that includes all of us, has pragmatic reasons to favor a cognitive system which is effective in generating truths, whether he otherwise cares about truth or not. (1993: 371)

A cognitive system that is not effective at generating true belief will not be adept at promoting our other values. True beliefs aid us in making choices; false beliefs may lead us to make choices that produce effects which do not accord with our values. Stich fails to notice these pragmatic reasons for valuing truth; he should not advocate pragmatic evaluation while denying truth's importance.

Hilary Putnam also convincingly argues for the value of true belief, but he does so on the basis of our survival. He begins by noting that even the opponents of a realist conception of truth such as Kuhn and van Fraasen agree that theories must be observationally adequate in order to be accepted by science (1981: 39). In these philosophers' views, beliefs are accepted because they can predict events or explain our sensations with some degree of success. They think this success does not imply that the beliefs are true, but instead that these beliefs are arrived at through trial-and-error. Yet why does trial-and-error work? Putnam says we can only explain this by examining "the environment-human interaction" at its heart, for "to posit that the interaction produces in our minds false theories which just happen to have successful predictions as consequences is to posit a totally inexplicable series of coincidences" (1981: 39). Our beliefs undoubtedly possess explanatory power with regard to the world, and it would be quite remarkable for trial-and-error to result in false beliefs that have such power. The simplest explanation is that these beliefs are at least approximately true.

This does not by itself mean that truth is valuable, however. The establishment of this conclusion requires the further linking of truth and survival. Beliefs that directly lead to action are called directive beliefs: one has a goal and believes that doing x will lead to that goal, and this directive belief then leads one to do x. Regarding these beliefs, Putnam says, "If too many of our directive beliefs are false, we will perform too many unsuccessful actions; so truth of (sufficiently many of) our directive beliefs is necessary for survival" (1981: 39). If one has a false directive belief about what it takes to achieve a goal, one is less likely to achieve the goal. If the goal is important enough, one's survival may be negatively impacted. Imagine that I have a false directive belief about where I must go to find water. If I do not replace this with a true directive belief, my health may suffer. Or imagine that I have a false directive belief about what I must do to drink water: I believe that I must add a certain powdery substance to it before drinking it. If this powdery substance is poison, I will not survive very long. If I had possessed a true directive belief about what I needed to do before drinking water, I might have survived longer.

The effects of my false belief about what I must do before drinking water could also have been avoided by my having a true belief about the effects of the powdery substance, because this would have caused me to generate some other directive belief. The role played by this other belief demonstrates the reliance of directive beliefs upon a network of other beliefs. We form our directive beliefs under the influence of our other beliefs, and "if these beliefs were mainly false, would it not be a mere coincidence if they nonetheless led to true prediction of experience and to true directive beliefs" (Putnam 1981: 40)? Our survival leads one to suspect the approximate truth of a substantial number of our directive beliefs, and this in turn increases the probability that many of our other beliefs are approximately true. True belief is conducive to survival, and anybody who values survival therefore has good reason to value truth.

We value truth in the context of belief-formation, but we also have other important values; indeed, this valuation of truth is in one respect tied to our valuation of survival. When engaged in cognitive evaluation, we should be neither wholly unconcerned nor wholly concerned with the generation of true beliefs. We ought to consider truth in its actual context amid the range of human values. We use cognitive systems for their benefits, and we should not pretend that true belief is not a significant benefit any more than we should pretend that it is the only benefit.

V. Conclusion

We ought to evaluate cognitive systems on the basis of their effects: They are good insofar as their effects accord with our values. Such evaluation implies that true belief is valuable, and their effectiveness at generating true beliefs is one of the criteria against which cognitive systems are judged. The common inferential strategies described by Nisbett and Ross ought to be compared with more rational alternatives on the basis of their effects relative to the subjects' values. Such evaluation would require greater exploration of the effects of various cognitive systems and the possible advantages that commonly used systems may hold over more rational systems. Without this knowledge, proper cognitive evaluation is impossible.

The standard proposed here points towards further research regarding values. More work needs to be done on precisely what values are and how they are related to cognitive systems. Thought especially needs to be given to standards of evaluation for values and to the different types of values. The value of true belief, for example, may not be comparable with the value of a microwave, and it would be worthwhile to describe how they differ. Furthermore, if, as Putnam suggests regarding true belief, certain values ought to be held simply because one is a living creature, then some cognitive systems may be better or worse for all humans.

If the subjects described by Nisbett and Ross are in fact reasoning poorly, the next question is how to improve the situation. Since we use cognitive systems unconsciously, the practicality of altering them through education needs to be explored. The possibility of such improvement increases the importance of developing a standard for cognitive evaluation. Further exploration of the cognitive systems we use in different situations could provide important data that would aid evaluation and improvement. The discovery of poor patterns of inference in common reasoning would be troublesome, but we should not judge these common cognitive systems without considering whether they promote some values other than true belief. As we cannot avoid using cognitive systems, it would behoove us to gain a better appreciation of what they are and how we can shape them.

Works Cited

Jacobson, Stephen. 1992. In Defense of Truth and Rationality. Pacific Philosophical Quarterly, 73: 335-346.

Kornblith, Hilary. 1993. Epistemic Normativity. Synthese, 94: 357-376.

Nisbett, Richard, and Lee Ross. 1980. Human Inference: Strategies and Shortcomings of Social Judgment. Englewood Cliffs, New Jersey: Prentice-Hall.

Putnam, Hilary. 1981. Reason, Truth and History. New York: Cambridge University Press.

Sanford, David H. 1992. The Anastylosis of Reason: Fitting Together Stich's Fragments. Inquiry, 35.1: 113-137.

Stich, Stephen. 1990. The Fragmentation of Reason: Preface to a Pragmatic Theory of Cognitive Evaluation. Cambridge, Massachusetts: The MIT Press.

Taylor, James E. 1999. The Value of Epistemology: A Defense. Philosophical Papers, 28.3: 169-185.