Chapter 2

The Connection to Truth

A salient point of our internalist resolution of the Gettier problem is the centrality of the concept of rationality. Indeed, an emphasis on the primacy of rationality in the hierarchy of epistemic concepts is one of the hallmarks of internalist metaepistemology. But this very point is the target of one of the most common complaints against internalism -- that in emphasizing rationality we have completely divorced epistemic rationality from knowledge.

Such a divorce is not unprecedented in the literature. Richard Foley, for example, argues that knowledge and epistemic rationality have so little to do with one another that one must simply choose which one to investigate at any given time. He therefore categorizes as "knowledge" cases where the subject has no reason whatsoever for holding a contingent belief (to the truth of which he has no direct access), but where the subject's inclination to believe is in fact reliably correlated with the truth of propositions of the kind in question. (1)

No internalist will be happy with such a proposal. But the externalist can press the question: does internalist justification have anything to do with truth? If not, then arguably it has no place in the analysis of knowledge. Knowledge comprises various conditions, but we do want these conditions to be related to one another rather than being an arbitrary and heterogenous collection. Furthermore, as Stewart Cohen has argued, a non-trivial connection to truth seems required for distinctively epistemic (as opposed to pragmatic or ethical) justification. (2) The J condition should have something to do with the T condition; if justification is to be an indispensable part of knowledge, we need to be able to defend its inclusion.

The Problem of the Connection to Truth

The connection to truth criticism has been a stumbling-block for traditional notions of justification since before the emergence of the contemporary internalist/externalist controversy. Richard Fumerton argues persuasively that Bertrand Russell anticipated externalism in 1948 by his reaction to Keynes's theory of probability. (3) More recently, so staunch an internalist as Laurence BonJour cedes ground to externalism insofar as he worries that even the "third condition" for knowledge (that is, justification aside from a Gettier-blocking condition) may have an external component. (4) Russell, BonJour, and many others have been motivated by the intuition that there must be a connection to truth in order for justification of any sort to obtain.

A survey of statements of this objection will not only illustrate the problem but also point towards the distinctions necessary to its solution. Fumerton, himself no externalist, states lucidly the motivation behind a rejection of internalist justification.

The fundamental idea behind Goldman's reliabilism is straightforward enough. When a belief is justified it has a virtue. There is something good about it. From the epistemic perspective, virtue has to do with truth. The reason epistemologists want epistemically justified beliefs ... is that having justified beliefs has something to do with having true beliefs. At the same time, we must understand justification in such a way that we allow the possibility of justified false belief. ... The answer is to focus on the processes that produce beliefs. (5)

Alvin Goldman makes a similar point in his own words in a rather roundabout way. He first states that the best "doxastic-decision-procedure" (DDP) which should guide people in accepting beliefs will be one that produces an "optimal" combination of true belief and error avoidance. This, it seems to him, follows directly from the plausible idea that the goals of cognition are believing truth and avoiding error. He then argues that these goals present a problem for internalism.

Unfortunately, the foregoing characterization of the right DDP ignores a crucial aspect of traditional epistemology. The foregoing conception rests on an "externalist" perspective: the perspective of a Godlike observer who, knowing all truths and falsehoods, can select the DDP that optimally conduces to true belief and error avoidance. Traditional epistemology has not adopted this externalist perspective. It has been predominantly internalist... On the latter perspective, epistemology's job is to construct a doxastic principle or procedure from the inside, from our own individual vantage point. ... The objective optimality of a DDP, on this view, does not make it right. A DDP counts as right only if it is "certifiable" from within. (6)

This internalist requirement that we certify any mode of arguing from within, says Goldman, cannot be satisfied. He therefore concludes that internalism is a "will-o-the-wisp" and that we should be satisfied with externalism, with its emphasis on the de facto optimal DDP. (7)

While Goldman stresses achievability, Bertrand Russell insists on what amounts to a form of externalism for more strictly analytical reasons.

If induction is to serve the purposes which we expect it to serve in science, "probability" must be so interpreted that a probability statement asserts a fact; this requires that the kind of probability involved should be derivative from truth and falsehood, not an indefinable; and this, in turn, makes the finite-frequency interpretation more or less inevitable. ... If the [inductive] principle is to serve its purpose, we must interpret "probable" as meaning "what in fact usually happens"; that is to say, we must interpret a probability as a frequency. (8)

While Russell does not use terminology exactly similar to that of the internalist-externalist controversy, the crucial concepts are here. Probability, according to Russell, must be "derivative from truth and falsehood." If it is related to truth and falsehood, it will not describe the rational credibility supported by a subject's evidence but rather will describe what in fact usually happens.

Michael Friedman also ties justification to truth in an externalist fashion when he addresses the possibility of an a priori defense of non-deductive inference forms in scientific methodology:

The impossibility of such a justification follows, it seems to me, from two simple and fundamental facts: (i) there has to be some kind of link between justification and truth; a justification of scientific method must say something about its propensity to lead to truth; (ii) scientific method is not logically guaranteed of reaching true conclusions; it is an incurably nondeductive method. There is no inductive method that is more reliable in every logically possible world than every other method; consequently, there is no method that is a priori best, there is no method that is a priori the most reliable. We have to know facts about the actual world if we are to know which method is best; and we have to know facts about the actual world to know even that any given method has any chance at all of leading to truth. (9)

Friedman puts the point more clearly than does Russell. If there is "some kind of link" between justification and truth, then non-deductive justification can only occur when one is using an actually reliable "method" of induction.

David Miller brings the point down to the level of practical application, expressing impatience with those who emphasize good reasons.

I often ask those who continue to find this ... insight unacceptable the following question. Which ticket would you prefer to draw in a sweepstake: the one bearing the favourite's name, or the one bearing the winner's? The answer I receive is almost always that the winning ticket is, of course, the best one to draw; ... But I am forcefully reminded after a semicolon's pause that, until the race is over, no one can know which ticket is the winning ticket, and as a matter of tactics the rational agent therefore prefers the ticket most likely to win ... It is the "therefore" here that takes my breath away. If the tactical preference for the most favoured ticket is not to be simply an underhand repudiation of the abstract preference for the winning ticket, then the agent must have conjectured that the ticket most likely to win actually will win. (10)

Miller concludes bracingly that "as far as rational thought is concerned, evaluation in terms of good reasons is a pure epiphenomenon" (11) -- a comment that recalls strongly Goldman's reference to the external perspective of a Godlike observer. Both emphasize the desire to have beliefs that are in fact true, regardless of whether one possesses good reasons for them.

Finally, and most interestingly, Laurence BonJour has been so greatly influenced by the connection to truth consideration that he nearly capitulates to externalism on the quintessential internalist notion of having good reasons. BonJour puts the argument into the mouth of Alvin Plantinga, although it nowhere occurs in just these terms in Plantinga's own work. But BonJour is clearly worried by it himself.

It seems apparent that an adequate third condition for knowledge must be one whose satisfaction yields an objectively good reason, not merely a subjectively good one, which is just to insist that justification or warrant must be objectively, and not merely subjectively, truth-conducive. The question then is whether ... oak tree experiences ... are objectively good reasons or merely subjectively good reasons for thinking that an oak tree is present. The answer to this question seems to hinge in large part on just how unusual this kind of case is in relation to the total class of actual and possible oak tree experiences. ... [I]f [this argument] is correct, then no purely internalist account of the third condition of knowledge will do, since some appeal to an external condition will be needed to guarantee the presence of an objective reason. (12)

BonJour says that he is not entirely convinced by this argument, because he has not yet given up on an a priori argument that (in the example given) "aberrantly caused oak tree experiences are sufficiently rare to make it objectively likely that such an experience will be accompanied by an oak tree," but he admits that "the prospects for success are extremely cloudy at best." (13)

Crucial Distinctions

To evaluate the externalist argument that justification must be "connected" to truth, we must make some essential distinctions. The first, frequently ignored in the philosophical literature, is the distinction between "sources of belief" or "belief-forming practices" and inference forms. In externalist writings one frequently encounters lists of belief-forming practices including everything from using induction to trusting our senses to trusting our friends. Alvin Plantinga groups together perception, the inclination to form beliefs about other minds, induction, memory, and the sensus divinitatis. (14) William Alston lists sense perception, memory, introspection, and deductive and inductive reasoning all as "basic sources" of belief. (15) Goldman speaks of "doxastic decision procedures" as if all ways of deciding what to believe or of forming beliefs were simply processes to be compared for their effectiveness. But prima facie, there is an important difference between induction and deduction, on the one hand, and such "sources" as memory and perception on the other. The former are contentless forms of reasoning, while the latter are not forms of argument at all but rather types of experience that typically move us to form beliefs.

The importance of the distinction between inference forms and other "sources" of belief becomes more evident when we consider another distinction -- the distinction between an intrinsic connection to truth and an extrinsic connection to truth, where the latter denotes some form of reliability in the world. In all of the quotations above where any determination can be made at all, it is clear that the authors are looking for an extrinsic connection to truth. Goldman speaks of the "objective optimality" of a DDP; Russell speaks of probability as a frequency; Friedman talks of a method's "propensity to lead to truth"; Miller is concerned with getting the "winning ticket" rather than getting the ticket one has the best reason to believe will win.

BonJour moves from a concern for an "objectively good reason" to the actual ratio of non-veridical oak-tree experiences to veridical ones. He seems convinced that reliability is an indispensible virtue for an argument form; otherwise, he fears, it provides only a "subjectively good reason." He recognizes that this requirement creates deep problems for epistemology; indeed, he suspects it will lead to skepticism of a particularly intractable type. (16) (See Chapter 4). Both he and Alston (17) have examined the possibilities for an a priori argument for reliability (e.g. for the reliability of our senses), but the obvious problem with this attempt -- that phenomenal experience does not entail realism -- makes its prospects not simply "cloudy" but nil. Whether sensation or any other "belief-producing mechanism" is actually reliable is not an a priori matter but an empirical one.

The terms "subjective" and "objective" in BonJour's argument require some scrutiny. In one of its meanings, "objectivity" is opposed to something like relativism. An objectively good reason would, on this interpretation, be one that is equally good for any reasoner who possesses it, as opposed to an argument that a reasoner merely happens to like or feels inclined to follow. Obviously, this is not what BonJour intends by "objectively good reason." He means one that (in some sense) tends to yield true conclusions in the real world.

But there are strong reasons to believe that the first concept of objectivity is the one that carries epistemic weight. We naturally want the epistemic resources to evaluate negatively the beliefs of lazy, inept, or biased reasoners; we do not want each person to be able to declare that something constitutes a "good reason for him." As an internalist, BonJour should be especially open to the possibility that there is a discernible intrinsic virtue which arguments objectively either do or do not have. The possession of this virtue would mean that the argument constitutes an "objectively good reason" for its conclusion in the sense of a non-relativistically good reason, regardless of whether that argument form, or the inference from those sorts of sensations to that sort of conclusion, has in some further sense an external-world "propensity" to yield true conclusions.

What might such an intrinsic virtue look like, and in what non-trivial sense could we call it a "connection to truth"? The answer to this question will depend in part upon whether the belief in question is foundational. Foundational beliefs will be evaluated solely in terms of generative epistemic principles, while an evaluation of the epistemic status of inferred beliefs will also require the use of transmissive epistemic principles. The epistemic principles will show that the foundational beliefs and the arguments based upon them have the relevant intrinsic virtue. It is therefore the epistemic principles themselves that must have -- or make evident -- the "connection to truth." There must be something the epistemic principles tell us about beliefs and arguments that shows those beliefs and arguments to be epistemically valuable.

For generative epistemic principles, not just anything will do. If the principle is to spell out an intrinsic connection between the foundational belief and truth, it cannot include the requirement that the belief be produced by a reliable mechanism or "proper function," for these would be extrinsic connections. It is in no sense a characteristic of a belief in itself that it is produced by a properly functioning or reliable mechanism. Moderate foundationalists, with their emphasis upon merely probable foundational beliefs, are especially likely to endorse extrinsic foundational connections. Robert Audi, for example, writes that grounding in experience "seems to explain why a belief so grounded may be expected to be true; for experience seems to connect the beliefs [it] ground[s] to the reality constituting their object, in such a way that what is believed about that reality tends to be the case." (18) If this is simply an empirical statement that certain types of experiences seem to be (perhaps causally) related to external reality, then it is not strictly speaking an epistemological claim, and it requires an empirical argument. If, on the other hand, Audi means to give an account of the source of justification (or some sort of epistemically positive status) for experiential foundations, he is describing only an extrinsic connection to truth, since in a different world such beliefs might not "tend" to be true. Other moderate foundationalists have argued that some beliefs really are intrinsically probable. (19) This concept of probability, whatever its further analysis, does aim at describing an intrinsic connection to truth.

It would take a different book to argue for incorrigibilist foundationalism of the sort we espouse. (20) But regardless of whether any sort of moderate foundationalism is defensible (and we believe that it is not), incorrigible foundations clearly possess an intrinsic connection to truth, for whenever they are believed, they are guaranteed to be true. (21)

It is more difficult to articulate what a connection to truth might mean for transmissive principles. One philosopher after another has slid down the reliabilist slope to the conclusion that any method of reasoning that purports to confer probability upon a proposition must itself produce a favorable proportion of true beliefs if consistently followed as a practice. If inference forms are separated from general "belief-forming practices," however, we can see that the reliability standard need not be applied to inference forms in order to grant them a connection to truth.

There is a helpful analogy here between deductive and non-deductive principles. In his polemic against a priori justifications of scientific method, Michael Friedman emphasizes that scientific method is "not guaranteed of producing true results" and is "incurably non-deductive." In contrasting deduction and induction, Friedman seems to imply that we know that deductive rules of inference are epistemically valuable because they are guaranteed to be reliable. The problem with non-deductive inferences on this view is that they possess no similar guarantee of external-world reliability.

It is true that when people start with true premises and reason correctly in accordance with the rules of deduction they invariably end up believing truths. But this does not mean that we need to look to empirical facts to investigate deduction's connection to truth. Indeed, the fact that correct deductions are defined by the rules of a deductive system enables us to investigate deduction non-empirically. The truth-preserving nature of deduction can be investigated a priori, both by our grasp of simple deductive rules and by way of proofs that display the consistency of a logical language in a way that can be clearly grasped. (22)

This fact about deductive logic points to the actual nature of its "connection to truth." Being transmissive, deduction merely preserves truth which must already be present in the premises. The truth-preserving nature of deduction is therefore the only "connection to truth" which it has in itself. Since its rules concern only connections between propositions, deduction does not possess a connection to external-world truth except insofar as it transmits the truth of external-world premises. But the fact that deduction preserves truth is not contingent but necessary. Hence, the distinctive deductive "connection to truth" is entirely a matter of the structure of deductive logic and owes nothing to the structure of the external world.

Correct non-deductive reasoning, like correct deductive reasoning, follows certain general and necessary rules. (23) The question of a connection to truth is interesting only in the case of reasoning that actually follows those rules, for we are no more interested in defending careless attempts at induction than in defending careless attempts at deductive reasoning. It is true that there is far more debate over the proper rules of non-deductive inference than over the parallel questions for deductive inference. But the correct rules of non-deductive reasoning -- whether direct inference, Bayesian, or yet more complex forms -- are neither equivalent to nor reducible to mere human "practices" or mechanisms generating readings in causal response to the environment, such as watches or thermostats. The premises of non-deductive arguments confer a certain degree of epistemic probability on their conclusions, and this is not a matter of what the external world is like but rather a necessary relation that can be determined a priori. As with deductive logic, the connection of probabilities to truth is intrinsic.

Non-deductive inference and the epistemic interpretation of deduction

An obvious objection to any analogy between deductive and non-deductive inference is that the latter does not preserve truth in the sense that deductive inference does. Deductive systems are checked by metatheoretic consistency proofs to demonstrate that they are truth-preserving; in deduction, there is no partial credit, no points for a system that usually yields true conclusions from true premises. In non-deductive inference there is an ineliminable slippage between premises and conclusion, so that even a conclusion inferred from premises known with certainty may turn out to be be false. Even if we say that non-deductive inference preserves or conveys probability, probability is not truth. It makes no sense to say that non-deductive inference preserves "probable truth," as if this were a species of truth, for truth does not come in degrees. (24)

What, then, is the point of referring to non-deductive inference as "partial entailment"? What does it mean to say that "probable" means "probably true"? And how can the analogy between non-deductive rules of inference and the rules of deduction provide the former with an answer to the problem of the connection to truth?

We can start to answer these questions by considering the connection between deductive logic and probability at the level of pure (formal) semantics. Instead of assigning the values "T" and "F" to the propositional letters, we can assign "1" and "0." The results of applying deductive rules to premises with these values can then be seen as limiting cases of the results of applying non-deductive rules, since non-deductive rules will in general yield values intermediate between 1 and 0. (25) For example, we may assign 1 to the statements

B: Nine out of ten Ss are Ps

and

P: a is an S

Here the principle of direct inference, which we shall explore in more detail in chapter seven, ascribes a valuation of .9 to the conclusion

C: a is a P

on the basis of B and P -- or, to put it a bit more precisely, it evaluates the strength of the connection between B and P on the one hand and C on the other at .9. This is an entirely formal concept of "partial entailment," in which both the rules of non-deductive inference and the valuations are given no ordinary semantic interpretation (such as "true," "false," "certain" or "probable") and in which even the propositional letters are not thought of as "propositions."

When we descend to the level of "depraved" semantics, deduction from premises given the values "T" or "F" admits of an entirely non-epistemic interpretation. Truth is non-epistemic; hence, we can define deductive entailments by speaking of which propositions must be true given the truth of other propositions. We need not make any reference to agents or epistemic categories such as rationality or justification in order to give a depraved semantics for deduction from true premises. In contrast, non-deductive inference always introduces valuations other than T or F for propositions and thus requires (at the depraved level) a semantics of probability. (26)

We notice, however, that deductive inference can be applied to premises labeled not "T" or "F," nor even "1" or "0" (which some might be tempted to treat as equivalent to truth or falsehood) but also to premises given numerical valuations between 1 and 0. For example, suppose that we assign .9 to

(P Y Q)

and 1 to

P

and that these are conditionally independent. The conditional probability of Q on these two sentences would not, in the formal semantics of probability, be 1 even though it is entailed by them. What is preserved by deduction in such cases cannot, therefore, be simply the truth of the premises. So the problem of characterizing probability at level of depraved semantics arises already with the shift to a continuum of valuations between 0 and 1, even before distinctively non-deductive rules of inference are introduced.

What does deduction preserve under such a valuation? What do the valuations themselves signify? An obvious answer is that, stretching the meaning of the term a bit, deduction "preserves" rational confidence about propositions. This suggestion raises the possibility that at the level of depraved semantics deduction itself is susceptible of an epistemic interpretation. The valuation "1" can be given the semantic interpretation "certainty" and the valuation "0" the interpretation "certainty of the negation of the proposition." When the premises of a deductive argument all are certain, a subject who reasons by correct deductive rules will give the conclusion the value 1. In these circumstances, it might seem that the use of such numbers is superfluous, yet when we provide a semantics for the numbers their relevance both to deduction with non-certain premises and to non-deductive inference becomes evident. The numerical values -- or intervals, where appropriate -- can be seen as credibility ascriptions representing the degree of confidence regarding premises or conclusions of a perfectly rational subject in a particular epistemic context. Certainty of the truth of a proposition and certainty of its falsehood are limiting cases of varying degrees of uncertainty. Non-deductive inference, like deductive inference, is able to convey probability, i.e. rational confidence regarding propositions.

One advantage of an epistemic interpretation of probability is that it makes evident the relation between probability and truth, since rational confidence is epistemically pertinent to the question of whether a proposition is true or false. Without specifying a particular "cut-off," we can say plausibly that a high rational confidence involves believing P, and that believing P means the same thing as "accepting P as true" or "taking P to be true." (27) Rational credibilities help us to determine whether a subject may rationally believe, disbelieve, doubt, or remain uncommitted regarding a proposition. It is in this non-trivial sense that "probable" means "probably true" -- if a proposition is highly probable on the basis of one's evidence, it is rational on the basis of that evidence to accept the proposition as true. Inferential relations which convey rational confidence are therefore intrinsically truth-directed, i.e. aimed toward truth, since they tell us when and how inference permits a subject rationally to believe a proposition -- that is, take it to be true. And this is the case even if the application of such inference forms is not truth-conducive, or successful, in the real world.

The asymmetry between inference forms and practices

The foregoing discussion highlights an asymmetry between inference forms and "practices" broadly conceived by such writers as Plantinga and Alston. It is true that inference forms can be treated as practices just in the sense that we might consider the human "practices" of following these inference forms correctly and might then investigate their successfulness in the real world. But to do so is to treat inference forms as belief-generating "black boxes," ignoring their claims to express necessary connections of rational support between propositions. Other "practices" differ from inference forms in that they make no such claims in the first place. It is therefore not possible to reverse the process and to treat other "practices" as inference forms. The practices of trusting the senses, trusting friends, trusting memory, and so forth are not intrinsically truth-conducive, nor do they describe rule-governed inferential connections between propositions which purport to convey rational confidence in propositions by their very nature.

The fact that the claims for inference forms are different from the claims for general practices indicates a further asymmetry. Any attempt to investigate the reliability of a belief-generating mechanism or the success of a belief-producing practice requires the use of inference forms; but it is not comparably true that any attempt to investigate the affidavits of inference forms requires dependence upon the reliability of some other practice. The first part of this statement is fairly obvious. Attempts to find out whether we usually get it right when we follow some practice are incurably empirical. To draw the conclusion that, by trusting perception, we form correct beliefs some favorable proportion of the time, we must gather evidence. And at a minimum, making the connection between evidence and conclusion requires using either deductive or non-deductive inference forms. (28)

It is the higher claim for inference forms that is likely to prove controversial. Alvin Plantinga frequently objects to any attempt to put the a priori on a different plane from other "sources" of belief, and he repeatedly quotes Thomas Reid to bolster the claim that all sources are in the same boat in that they must be taken more or less on trust.

Why, sir, should I believe the faculty of reason more than that of perception? They came both out of the same shop, and were made by the same artist; and if he puts one piece of false ware into my hands, what should hinder him from putting another? (29)

One can, on Plantinga's view, take some sources on trust and use those to investigate the others, but it will be an arbitrary exercise to decide which sources to take as given. Hence, he concludes, there is no privileged "original position" from which to evaluate claims which (on his view) are "properly basic," be they the deliverances of the senses, of memory, or of reason. (30)

In arguing for such complete parity among all sources of belief, Plantinga relies heavily on the assumption that our justification for believing what purports to be an a priori truth is simply that a certain belief has "the peculiar feel that a priori beliefs have -- that feel that somehow they just couldn't possibly be false. But of course," he continues, "Such a feel could be misleading." (31) Here Plantinga assumes without conspicuous argument that there can be no such experience as direct grasping of a priori truth. While he uses examples such as the Russell paradox to point out that people have been mistaken about what are (if anything is) a priori matters, he goes farther and implies that a subject's justification for believing even such a simple proposition as 2 + 1 = 3 is just a matter of "finding" himself believing it and "trusting" that the "feel" he has about it is not misleading in this particular case. Similarly, the best we can say, according to Plantinga, about the belief that the corresponding conditional of modus ponens is true is that we "just find ourselves with this powerful inclination to believe this proposition is true, and indeed couldn't be false." (32)

However, it is important to distinguish those a priori truths which are so simple that we can hold them in our minds all at once and see their truth clearly and distinctly from those which, because they require a concatenation of steps, many of us cannot grasp all at once. (33) In the latter case, the proofs must be broken down into shorter steps. This is not, however, to concede that in no case are we able to perceive the truth of an a priori claim in such a way that we could not be mistaken about it. In those cases what we have is not simply a subjectively strong feeling about a proposition but a direct grasping of its truth. It is for this reason that no one who is capable of understanding statements such as 2 + 1 = 3 can be mistaken about them, whereas for most people, more complicated propositions of logic or mathematics cannot be seen with this self-verifying clarity. (34)

It may seem that in making this distinction, we are making a fatal admission. For inductive logic is a difficult subject, as is epistemology. Even if it is in principle an a priori matter that certain rules of inference connect propositions in such a way as to confer probability upon conclusions, must we not nevertheless "take on trust" our own ability to see this truth connection, since much of probability theory is too complex to be grasped all at once? Even if we can give an a priori argument that induction confers rational credibility, must we not rely upon our memory, for example, to feel confident that we have not forgotten something or made some mistake when following the argument in its step-by-step form?

The answer depends on who "we" are. Some people can hold more steps of an a priori argument in their minds at once than others. Some can see clearly and distinctly a conclusion too complex to be seen all at once by those less well-trained. But far from making the evaluation of inference forms into an empirical discipline, these differences among people of varying degrees of talent and mental power show that partial reliance on memory is really beside the point. By their nature, a priori truths are the sort of thing that can be grasped in a self-evident fashion. Indeed, when contemplating 2 + 1 = 3 or the corresponding conditional for modus ponens, one has (pace Plantinga) exactly such an experience -- not simply a vague feeling which might or might not be correct, but a genuine experience of seeing the truth of a proposition by reason of its conceptual structure. It is, then, merely a matter of extrapolation of our own powers and our own experiences of grasping the truth to imagine more complicated a priori matters similarly being understood in a clear and distinct fashion. It is a contingent fact that this or that person must break down the mathematical rationale for, say, Bayes's theorem into several steps in order to grasp each part of the argument. Some other person may well be able to see it whole. The in principle nature of the subject matter is all that is required to distinguish inference forms from empirically based "practices." It is a necessary aspect of arguing for the reliability of practices that one use inference forms. It is no necessary aspect of arguing for the truth-directed nature of inference forms that one trust some other, contingently reliable, faculty or practice.

Reliability claims are all object level

There are several arguments for keeping empirical content at the object level and for keeping all claims about justification purely a priori. In Chapter 4 we argue that any attempt to allow empirical claims at the metalevel creates a vicious regress that precludes the very possibility of justification. Here we offer a different argument based on the distinction between inference forms and other "belief-forming practices."

If one permits, or even requires, reliability claims at the metalevel, there will be no clean and non-arbitrary analytical cut between the object level and the metalevel, since some ostensibly metalevel claims will now be empirical in nature. (35) More seriously, because of this muddying of the object/meta distinction a subject can, on many externalist models, receive an epistemic free ride for almost any belief. If the subject does not wish to give reasons for holding an empirical proposition which, if true, would render a belief of his reliably produced, he can arbitrarily state that the proposition in question is an epistemic principle and that, instead of reasoning from it, he is reasoning in accordance with it. For example, if he is using sensory experiences which incline him to believe propositions about the external world, he can say that it is an "epistemic principle" that sense perception is reliable. Since an epistemic principle appears at the next epistemic level, he can claim that he does not need to argue for it in order to be justified at the first level. He can simply trust it or rely on it without argument. This move effectively hijacks the traditional distinction between premises and epistemic principles to eliminate, at the whim of the subject, the need for defending even controversial empirical propositions.

This consequence will follow even if one says that object level beliefs must be defended with internal reasons down to the foundations, provided one has an expansive enough concept of what counts as "foundational." Plantinga's own list of "properly basic" belief-types provides an excellent example of what follows from this approach. Does S want to believe in God? He need only declare the sensus divinitatis to be a truth-conducive module of his properly-functioning faculties and instantly declare himself "warranted" for believing in God. Asked whether there is such a thing as the sensus divinitatis and what his reasons are for believing both that it is truth-conducive and that it is producing his belief in God, he can shrug and state that there are all sorts of "sources" for "properly basic" beliefs, that none of them can be validated from a privileged original position, and that we must trust our faculties when we "find ourselves" believing certain things. A similar move will apply to memory, perception, beliefs about other minds, and so forth. Indeed, given this externalist model, one can be "justified" without even stating that one's belief source is reliable, much less defending that claim. So long as the mechanism is in fact reliable, one can happily depend upon it and be de facto "justified."

One can construct a version of "externalist internalism" that avoids this unsavory consequence by stipulation. For example, one could require that foundational beliefs have an intrinsic connection to truth and that inferences from those beliefs necessarily convey rational confidence, adding the metalevel externalist requirement that the "practices" the subject follows must also produce a certain proportion of true beliefs either in the actual world or in some relevant set of possible worlds. This does not take care of the problem of the metaregress, but it does prevent a subject from declaring anything he does not want to defend to be a metaprinciple.

Even on a fairly stringent combination of traditional internalism with a truth-conduciveness requirement, however, the issue of classification is handled in an unsatisfying way. Suppose, for example, that S is depending upon his sensory experiences to conclude that an oak tree is present. According to BonJour, a fact about the actual proportion of veridical to non-veridical oak-tree experiences would then have to be a true metalevel proposition, in order to guarantee that S has an "objectively good" reason for believing in the existence of the oak tree before him. This would place at the metalevel many propositions about the presence of oak trees in various times and places (in order to support the empirical generalization about truth-conduciveness), although these empirical facts would have no place at the object level at all for this particular inference. Yet in some other context, where the subject is reasoning from the presence of oak trees to some further proposition (e.g. the statement that he will find an acorn), the very same empirical propositions would belong at the object level.

Matters are further complicated by any attempt to accommodate the internalist intuition that all empirical propositions relevant to justification must be justifiedly believed by the subject. If the reliability of S's watch is relevant to his being justified in believing that it is three o'clock, then it seems to a philosopher with internalist sympathies that S must have some reason for trusting his watch. If the reliability of our inference forms in the real world were a prerequisite for inferential justification, this principle would indicate that we must have reason to regard our inference forms as reliable. If reliability statements are present as empirical metalevel propositions, the internalist will want the subject to have justified empirical meta-beliefs in order to have justified object level beliefs. BonJour is explicit about this requirement in his discussion of an "objectively good reason" for believing in oak trees.

Moreover, since it is an internalist version of the third condition that is in question, it will not be enough for the experiences in question to be as a matter of external fact strongly correlated with oak trees; so long as the believer in question has no access to this fact, the experiences will still fail to constitute an objectively good reason for him. (36)

Fumerton also requires metalevel beliefs as a condition of (inferential) object level knowledge. His "principle of inferential justification" states that, in order to be justified in believing P on the basis of E, the subject must be "justified in believing that E makes P probable." (37) Fumerton distinguishes what he calls "primary epistemic principles" showing connections between E and P from "secondary epistemic principles." The latter category includes such connections as that between litmus paper's turning red and the acidity of a solution, while the former includes the probabilistic connection between the premises and the conclusion of an inductive argument and the relations of entailment in a deductive argument. (38)

Once again, it accords well with a strong internalist intuition to say that one cannot justifiedly infer on a mere whim that a solution is acidic from the color of litmus paper. Similarly, as Fumerton argues, if Mother Paula the astrologer informs us that she can tell from the positions of the stars that a year of prosperity lies ahead, we will be inclined to say that she is unjustified simpliciter (not simply unjustified in believing herself justified) if she has no reason whatsoever for thinking that stellar configurations are correlated with economic success. (39)

Must we require metalevel justification as a condition of object level justification in order to accommodate this intuition? Externalists will seize gleefully on the (also strong) intuition that anything like a JJ thesis or KK thesis is false to declare any such requirement excessive. If internalists treat empirical correlations as metalevel "relations of probability," externalists will use the implausibility of a metalevel belief requirement to reject the internalist requirement of reasons for empirical correlations.

Richard Feldman and Earl Conee point out that both accessibility internalism and mentalist internalism are logically distinguishable from a requirement for a priori metaprinciples. (40) But the denial of a priorism does place either the access internalist or the mentalist internalist in a dilemma: If he does not require the subject to have knowledge of metalevel propositions, the externalist can always "shelve" controversial empirical claims by moving them up to the metalevel and treating them as metaprinciples that he therefore does not have to defend. If the internalist wants to block this move but does permit new empirical information at the metalevel, he will not be able to distinguish in a principled way a requirement that the subject know that his watch is reliable from a requirement that the subject know probability theory. But this is a severely damaging problem, and it points to a priorism as the key component of any adequate internalist metaepistemology.

We can refrain from a metalevel requirement for object level knowledge while retaining the requirement of evidence for empirical connections if we draw the line between metalevel and object level precisely where Fumerton draws the line between "primary" and "secondary" principles. Both the rank arbitrariness of Plantinga's system and the conceptual confusion arising from BonJour's concept of an "objectively good reason" can be neatly avoided by following the simple requirement that we keep all reliability claims firmly at the object level. This will mean denying that statements of empirical correlation or reliability constitute epistemic principles in any sense whatsoever. Instead, they are empirical propositions that often must be used as ordinary, object level premises in order to draw conclusions from empirical data. The claim that red litmus paper is frequently correlated with acidity can easily be regarded as a needed premise of the argument for the conclusion that a solution is acidic. But the claim that the premises of an inductive argument probabilify the conclusion is not, of course, a premise of an argument for an inductively-drawn conclusion. (41)

The reason for this dissimilarity lies in the distinction we have drawn between an intrinsic and an extrinsic connection to truth. That distinction, in turn, arises directly out of the distinction between inference forms and other "practices." Statements about the rationality of inference forms belong at the metalevel. The subject reasons in accordance with the inference forms, and doing so correctly from justified foundational beliefs renders his inferences justificatory. Although a very interesting argument could be conducted on this point, there is at least prima facie plausibility to the position that the subject does not need to believe the a priori axioms of probability or the rules of deduction in order to be justified by following an inference that correctly exemplifies these inference forms. (42) On the other hand, if a subject draws conclusions about the future from tea leaves, it is perfectly legitimate for us to demand that he have some reason for thinking that tea leaves are predictive of the future. There is no such thing as "reasoning in accordance with a tea leaf inference form." Any connection between tea leaves and the future is a contingent, empirical matter; therefore, the subject's object level argument is not cogent unless it includes a justified belief that such a connection exists.

If all reliability claims are kept at the object level, the metalevel argument for the claim that a belief is justified will contain no new empirical information. The empirical statements in the object level tree of reasons will stand in a certain relationship to one another, the foundations will either have or not have the intrinsic truth connection, and these aspects of the object level tree will reveal the rationality or otherwise of an inferred belief. What of the claims that S does hold the belief in question and that he is basing his belief on such-and-such reasons? These are not premises of the object level argument, but they can be read off of the object level argument as it exists in S's mind; each of them has a counterpart in a feature at the object level, and in that sense their appearance at the metalevel is not the introduction of anything new. What will not appear at the metalevel are additional empirical statements about factors that have no counterpart at the object level, such as a particular physical cause for the belief or the reliability of a particular mechanism generating the belief.

In excluding reliability claims from the metalevel, we designate the metalevel by its one unique claim: the claim of rationality. The statement that a particular type of belief is justified means nothing if it does not at a minimum mean that the belief is rational. But this is not a matter of reliability at all. One can be entirely rational even in a world where, for reasons to which one has no access, all of one's "practices" turn out to be unreliable. If we make rationality not only a necessary condition for justification but also a sufficient condition, we cleanly separate the metalevel from the object level at the only principled place for that break to fall.

Replies to objections

It may be objected at this point that everything said so far applies only to our justification1, and that we have transferred externalist considerations to justification2 (which we are conveniently refusing to talk about), thereby obscuring an important concession to externalism. After all, the claim that a belief has justification2 seems obviously to be a metalevel statement, and it does involve the claim that the propositions the subject uses as reasons are true, which is not in itself an a priori matter.

To answer this objection we must emphasize again the "no new empirical information" requirement, which is satisfied by justification2 despite the T requirement. For justification2 (as for knowledge), the claim that a particular proposition is true can only be defended at the object level by way of the object level reasons for the proposition itself. There is no separate, uniquely metalevel, access to empirical truth. There is only our object level foundational access or our evidence for inferred beliefs. Therefore, the metalevel claim that P is true is not a new piece of empirical data introduced at the metalevel for purposes of underwriting justification at the object level. The T requirement for justification2, like the T requirement for knowledge, thus differs crucially from any reliability component for justification. A reliability component requires that a certain type of empirical information (namely, the information that a particular process produces a particular proportion of true beliefs) be introduced at the metalevel and be treated as specially relevant to the metalevel, regardless of whether it appears at the object level. A T condition, whether for knowledge or for justification2, involves no such requirement and therefore keeps all the empirical "action" squarely at the subject's own object level. It is perhaps for this reason that externalists have not rushed to embrace the Russellian solution to the Gettier problem and to declare an acceptance of it a victory for their position. (43) From a T condition there is no road to anything recognizable as externalism.

A potentially more serious concern has been raised by Richard Fumerton in his careful and largely positive consideration of a priori approaches to epistemology. Fumerton's point amounts more to a word of caution than to an objection as such, but it is a word of caution worth heeding. Fumerton points out the danger that, in availing ourselves of the claim that epistemological propositions have a priori status, we will become dogmatic and classify as a priori epistemic principles whatever propositions we find convenient for evaluating the conclusions we wish to dub "justified." This approach would be tantamount to a Keynesian version of Chisholm's epistemic principle lists. Using this method, one could help oneself to principles guaranteed to deliver a non-skeptical conclusion. (44)

Since this is obviously an unacceptable use of the a priori appeal, how can a priorists avoid it? As Fumerton emphasizes, one important way for a priorists to keep themselves honest is not to be driven by the fear of skepticism. One must be willing to develop principles independently of the conclusions one wishes to certify as justified and let the philosophical chips fall where they may. (45) It also "should go without saying," Fumerton points out, that in appealing to relations between propositions that one is able to "see" "one must be absolutely sure that one understands that about which one talks." (46) If there is supposed to be an a priori probability relation between propositions, one must really have a clear concept of it before appealing to it.

All of this is excellent advice, but there is more to be said. Fumerton takes the a priorist position regarding non-deductive inference to be that the probabilifying relationship between propositions is "unanalyzable." (47) He also asserts, in line with this assumption, that

the foundationalist who ... seeks to avoid both epistemic and conceptual regress concerning justified beliefs about probabilistic connections by embracing a Keynesian conception of epistemic probabilisty, will refuse to offer an argument for the various epistemic principles ... he endorses. It is, after all, part of the view that one can discover a priori primary epistemic principles. (48)

At this point worries about dogmatism understandably arise. We discuss the nature of argument in a priori contexts more fully in Chapter 6 in the context of the defense of deduction, contending that argument in a priori contexts is really explication. Given that this is the case, it does not follow either from foundationalism or from a priorism that one will refuse to give any sort of argument for one's claims of probabilistic relations. In fact, the argument will take the form of conceptual analysis. An indispensable requirement for the honest a priorist is that he analyze as far as he can, both to achieve entirely clear understanding for himself and to answer the objections of a skeptical audience. By way of conceptual analysis the a priorist should aim to present as convincing a case as possible for epistemic principles that have as wide an applicability as possible.

For example, some version or cousin of Ockham's razor may very well be an a priori epistemic principle. It is nevertheless quite unsatisfactory for an epistemologist to say, "Simple theories are more likely to be true than complicated ones." Even the addition of a ceteris paribus clause will scarcely suffice to render this bare assertion an example of a priori epistemological reasoning worth the name. What is the meaning of "comparative simplicity"? Can simplicity be quantified? Why are simpler theories more likely to be true than more complex ones? Can we give a rigorous proof of this claim using still more fundamental theorems of probability? All of these questions may have entirely adequate answers, but part of the epistemologist's job is to be able to answer them, even for the sake of his own clear vision of the principle he is claiming to know a priori. There will come a point where further analysis is impossible and where clear conceptual connections must be grasped directly, and the extent to which a principle needs to be analyzed and broken down may vary as a function of the capabilities of the subject in question. But probability theory is difficult enough that such an end of analysis will in many cases be a long time in coming. With this demanding version of an a priori approach on the table, the skeptic need not fear that he will receive simplistic answers based on an ad hoc attempt to guarantee common sense results.

The discussion so far will make it fairly clear that we take a priori epistemic principles to be analytic. In discussing analytic epistemic principles, Fumerton's general concerns about dogmatism become more acute, and he seems to believe that the only plausible version of a priorism is rationalist, taking epistemic principles to be synthetic a priori. The objection here is that analytic epistemic principles will trivialize the discussion.

It seems to me a little difficult to suppose that the many skeptics and those who took them seriously were all simply misusing language. However implausible we might view skepticism about the physical world, are we really to maintain that such skeptics were simply contradicting themselves? Can we really dismiss the skeptical challenge by exclaiming that we just understand rationality in such a way that it follows from the concept alone that sensations make it rational to believe propositions about the physical world? In short, the solution seems too easy. (49)

This objection is echoed by Michael Friedman, who expresses exasperation at

attempts to justify induction by appealing to the meaning of 'justified' or of 'rational' or of 'inference' ... It is not very comforting to be told that scientific method is justified or rational in virtue of the meaning of 'justified' or the meaning of 'rational'... (50)

Though we are defending an analytic version of the a priori approach Friedman finds "uncomforting," his impatience at an attempt to justify induction by linguistic fiat seems reasonable. A successful a priori defense of inference forms will have to do better than just smuggling 'rational' in as part of a stipulative definition of a form of reasoning. On the view we are advocating, something much more rigorous will be required. The epistemologist will have to show that certain forms of inference are intrinsically truth-directed, that it is rational to reason in accordance with them. For induction, this will require explicatory argument in terms of concepts such as reference class, sample, and randomness (see Chapter 7). For inference to the best explanation, the a priorist will need to parse out concepts such as simplicity and explanatory power, and for Bayesian inference, concepts such as comparative likelihood. The emphasis, then, is on concepts and their connections, with words playing a role only insofar as they are used to designate concepts.

In point of fact, conceptual analysis is less prone to produce unargued statements of "insight" than a position that depends upon the synthetic a priori. The advocate of analyticity bears the burden of showing that his positive evaluation of a form of inference really does bring together detailed, reflective understandings of basic concepts of rationality. While the concepts involved will ultimately be his own concepts, as he can have only defeasible access to the concepts of others, his position will only have plausibility for others (and, if he is honest, for himself) if it can be defended by careful analysis. The advocate of the synthetic a priori, on the other hand, has no such burden, since he does not claim to be analyzing anything at all. (51)

This version of analytic a priorism also avoids the danger of fragmentation, a charge leveled at Chisholm and Pollock by Stewart Cohen. Cohen points out that their principles "are nowhere united by a general theory that explains why those beliefs are justified under those conditions. In a sense, [their] theories do not tell us what justification is." (52) Fumerton raises the same issue with specific reference to analytic versions of a priorism.

One is sorely tempted to suppose that philosophers who take epistemic principles to be analytic do normative epistemology by simply listing their prephilosophical beliefs, deciding what they do infer the propositions believed from, and proclaiming the epistemic principle sanctioning such inferences to be analytic. But what exactly do all these inferences have in common that makes it plausible to claim that they fall under a single concept of rational inference? Is the concept of probabilistic or evidential connection simply a disjunctive concept?...It surely makes sense to ask: "In virtue of what do both sensations and memory experiences make probable, respectively, propositions about the physical world and propositions about the past?" (53)

It does, indeed, make sense to ask such a question. But far from dooming us to a list of unrelated "epistemic principles" that guarantee favorable status to our favorite empirical conclusions, the conceptual analysis of argument forms has the potential to yield a unified view of justification based on a detailed explication of rationality.

What of Fumerton's concerns regarding the skeptic? Does analytic a priorism imply that the skeptic's concerns are trivial and easily answered? Has the skeptic simply been "misusing language" by declaring ordinary beliefs unjustified if, in fact, they are justified as a matter of analytic, a priori truth? None of this follows if we work with a notion of analyticity that focuses on concepts rather than language. If the skeptic has been wrong, he has in fact been conceptually and therefore analytically wrong on our view (given certain assumptions about the skeptic's own concepts). But this implies no derogation to the skeptic. The concepts involved in probability theory are complex and difficult to grasp with complete clarity. Even in deductive logic, there are proofs so complex that people make mistakes about them. In mathematics, a provable theorem is declared to be "trivial," but it is only when "trivial" is thus used as a term of art that we will be able to declare the skeptic's concerns "trivial" -- and that, only after we have answered him decisively.

Friedman's complaint that a priori approaches are not "comforting," however, leads us back to the connection to truth issue and to the final objection we will consider. Many philosophers who have embraced the externalist version of a truth-connection requirement may simply feel that an intrinsic truth connection is not good enough. After all, one could reason in accordance with intrinsically impeccable inference forms from intrinsically truth-connected premises and still end up believing a set composed entirely of falsehoods (if the inference forms were non-deductive). It would require some ingenuity to dream up such a situation, but surely a sufficiently diabolical Deceiver can be found to do the trick. Many philosophers have a strong intuition that only an extrinsic connection will do. If our inference forms are not extrinsically truth-conducive -- if, in Miller's terms, we do not actually "have the winning ticket" -- then we are not justified in any sense that interests them.

Perhaps nothing said here will move those who have such an intuition. There are, however, two relevant responses to this externalist complaint. First, Fumerton has argued that any sophisticated version of externalism will be open to the same objection. (54) He points out that a belief-forming process type might operate only once in the actual world and might quite accidentally give a true conclusion in that instance. Since "accidental justification" bothers externalists a good deal (as witness their response to the Gettier problem), the sophisticated externalist will want to invoke counterfactuals or propensities in order to indicate a genuinely nomological connection between the operation of the belief-forming mechanism and truth. But once the externalist has shifted his analysis in that direction, it is logically possible that all of our own actual inferences are just those in which the process-type does not manifest its propensity to produce true beliefs. If this is in fact the case, then the externalist would be calling beliefs "justified" although they were produced by a process-type that was de facto unreliable (even if reliable counterfactually or nomologically). He would thus lose his supposed advantage over the internalist in bestowing the accolade "justified" only on beliefs with a real extrinsic connection to truth.

Second, the internalist can challenge directly any implication that he is insufficiently concerned about truth. The most die-hard internalist will agree that it would be a good and desirable thing for our inference forms to produce a hefty proportion of true beliefs in the real world. He may even believe that the world is stable and orderly, that no clever Deceiver exists who stirs up reality so as to defeat our inductions, and that therefore induction (for example) does usually produce true beliefs. But if so, he believes this as a result of an inference. No one who is not divine has direct divine access to the nature of the physical world. Even Plantinga's "properly basic beliefs" are "properly basic" only because a particular subject does not happen to infer them and because they are produced reliably by some non-belief-dependent, truth-conducive module of the subject's design plan. They are not the result of an infallible insight into the nature of the external world.

As we have already shown in describing the asymmetry between inference forms and practices, anyone who wishes to investigate claims about the stability of nature or the reliability of this or that process (rather than making such claims without argument and dubbing them "properly basic") will have to gather evidence and draw conclusions from that evidence. The argument for those conclusions will itself have a certain structure that either is or is not intrinsically truth-directed. If someone decides that nature is uniform, for example, and if he then ascends to the metalevel to investigate the affidavits of his inference form, there is no reason for him to feign ignorance by treating the inference as a structureless "mechanism" and engaging in a purely empirical investigation as to its reliability. (55) It is surely worth his while, if he has the capability, to look into the question of whether his inference form has the desirable quality -- and the quality it purports to have, in virtue of its premises and the inferential rules it follows -- of making its conclusion rationally believable. It is worth his while precisely because he is interested in the question of whether his object level conclusion about the stability of the world is true. Therefore, a concern about the truth-conduciveness of our practices leads directly to gathering evidence and making an empirical argument that (for example) the world is a relatively stable place. And a concern about the structure of that argument leads to an investigation of intrinsic truth connections.

Far from abandoning a concern for truth, internalists have taken the most direct route to discerning it, and to discerning that they have discerned it. The externalist approach offers only one empirical investigation after another, the final vindication of our beliefs ever receding before us, our hopeful spirits buoyed by promissory notes to the effect that we are justified if the process we have just used is reliable. Meanwhile, the externalist conflates inference forms and empirical practices and therefore ignores the special claims of the former, not bothering even to investigate the possibility of an intrinsic truth connection. It behoves us, precisely if and because we care about truth, to make that investigation.

1. Richard Foley, The Theory of Epistemic Rationality (Cambridge, MA: Harvard University Press, 1987), pp. 167-169. Working Without a Net: A Study of Egocentric Epistemology (Oxford: Oxford University Press, 1993), pp. 85-87.

2. Stewart Cohen, "Justification and Truth," Philosophical Studies 46 (1984), p. 279.

3. Richard Fumerton, Metaepistemology and Skepticism (Lanham, MD: Rowman and Littlefield, 1995), pp. 192, 200-202.

4. Laurence BonJour, "Plantinga on Knowledge and Proper Function," in Jonathan Kvanvig, ed., Warrant in Contemporary Epistemology: Essays in Honor of Plantinga's Theory of Knowledge (Lanham, MD: Rowman and Littlefield, 1996), pp. 53-55.

5. Fumerton, p. 97.

6. Alvin Goldman, "The Internalist Conception of Justification," in French, et. al., eds., Midwest Studies in Philosophy V 1980 (Minneapolis, MN: University of Minnesota Press, 1980), p. 32.

7. Ibid., pp. 38, 45.

8. Bertrand Russell, Human Knowledge: Its Scope and Limits (New York: Simon and Schuster, 1948), p. 401

9. Michael Friedman, "Truth and Confirmation," in Hilary Kornblith, ed., Naturalized Epistemology (Cambridge, MA: MIT Press, 1985), pp. 155-156.

10. David Miller, Critical Rationalism: A Restatement and Defense (Chicago, IL: Open Court, 1994), p. 66.

11. Ibid.

12. BonJour, "Plantinga on Knowledge and Proper Function," pp. 53-55.

13. Ibid., p. 54.

14. Alvin Plantinga, "Respondeo," in Kvanvig, ed., p. 342. Also Alvin Plantinga, Warranted Christian Belief (Oxford: Oxford University Press, 2000), p. 175.

15. William Alston, "Epistemic Circularity" in Epistemic Justification (Ithaca, NY: Cornell University Press, 1989), p. 326.

16. BonJour, "Plantinga on Knowledge and Proper Function," pp. 63-64.

17. William Alston, The Reliability of Sense Perception (Ithaca, NY: Cornell University Press, 1993), Chapter 3.

18. Robert Audi, "Contemporary Foundationalism," in Louis Pojman, ed., The Theory of Knowledge: Classical and Contemporary Readings (Belmont, CA: Wadsworth Publishing Co., 1999), p. 207. It is not clear exactly what significance Audi attaches to this connection to truth, since elsewhere he seems to disavow a need for actual reliability of belief sources. See The Structure of Justification (Cambridge: Cambridge University Press, 1993), pp. 314-322.

19. Evan Fales, A Defense of the Given (Lanham, MD: Rowman and Littlefield, 1996), pp. 174-176.

20. Timothy McGrew, The Foundations of Knowledge (Lanham, MD: Littlefield Adams Books, 1995).

21. In our view, this is because they possess a referential relation to experiential states which render them true necessarily.

22. In Chapter 6 we will respond to the externalist tu quoque argument that a priori consistency proofs in deductive logic have the same property of epistemic circularity that internalists find objectionable in externalist defenses of contingent epistemic principles.

23. In saying this we are consciously contradicting L. Jonathan Cohen's multicriterial theory of probability as he lays it out in The Probable and the Provable (Oxford: Clarendon Press, 1977), chapter 2. We agree with his contention that provability is a limiting case of probability, but unlike Cohen we do not believe that there is any contingent or singular relation of provability.

24. We are indebted to Richard Fumerton for emphasizing this point in private communication.

25. The idea is not idiosyncratic. See A. de Morgan, Formal Logic, or the Calculus of Inference, Necessary and Probable (1847) and particularly Keynes's Treatise on Probability (London: MacMillan, 1963), pp. 133 ff.

26. It would be possible to give such a semantics entirely in terms of class ratios, and frequentists will insist that this is the only correct interpretation of probability even for epistemic purposes. In chapter seven we will indicate more fully why we adopt an epistemic rather than a frequentist conception of probability.

27. This terminology is not intended to imply that subjects can only have beliefs if they possess the abstract concepts of truth and falsity.

28. See Lawrence BonJour, Epistemology: Classic Problems and Contemporary Responses (Lanham, MD: Rowman & Littlefield, 2002), pp. 236-7 for a similar point on the priority of the internalist approach and the metaregress created by externalism.

29. Alvin Plantinga, "Respondeo: Ad BonJour," in Kvanvig, ed., p. 342; Warranted Christian Belief, p. 130, n. 23 and p. 221; Warrant, the Current Debate, p. 100; cf. Warrant and Proper Function, pp. 236-7.

30. Alvin Plantinga, Warranted Christian Belief (Oxford: Oxford University Press, 2000), pp. 128-130.

31. Plantinga, "Respondeo: Ad BonJour," p. 341.

32. Ibid., p. 342.

33. Descartes appears to have had something like this distinction in mind in the second set of replies. See John Cottingham, Robert Stoothoff, and Dugald Murdoch, eds., The Philosophical Writings of Descartes (Cambridge: Cambridge University Press, 1984), vol. 2, p. 104. But the interpretation of Descartes on whether it is possible to be wrong about truths clearly and distinctly perceived is sufficiently controversial that we do not intend to take a position here on the historical question.

34. See Chapter 5 for further discussion of intuition of a priori truths and of Plantinga on phenomenology and a priori knowledge.

35. Note that we are here taking "reliability" to be by definition a contingent matter and statements of reliability to be contingent.

36. BonJour, "Plantinga on Knowledge and Proper Function," p. 54.

37. Fumerton, p. 36.

38. Ibid., p. 105. Fumerton has continued to maintain that metalevel justified belief is required for object level justified belief. But he has agreed in personal communication with the position expressed below that there are in fact no "secondary" (empirical) epistemic principles and that contingent connections should all be regarded as additional object level premises. The reference to contingent principles in the book was not intended to express his own position.

39. Fumerton, p. 86.

40. Richard Feldman and Earl Conee, "Internalism Defended" in Hilary Kornblith, ed., Epistemology: Internalism and Externalism (Cambridge, MA: Blackwell, 2001), pp. 233-6.

41. Ernest Sosa never considers the possibility of excluding empirical claims from the metalevel, and he therefore accuses internalism of generating a metaregress by its insistence that a subject possess evidence for reliability claims that are epistemically relevant. Knowledge in Perspective (Cambridge: Cambridge University Press, 1991), pp. 194-5.

42. Fumerton staunchly denies this possibility and maintains that a subject whose premises strictly entail his conclusion must believe that the entailment holds in order to be justified at the object level. Fumerton's position is that there is no justificatory inference at the object level (as opposed to a mere causal connection) if there is no justified belief that the premises make the conclusion probable. "Inferential Internalism and the Presuppositions of Skeptical Arguments," unpublished manuscript.

43. As we explained in Chapter 1, Plantinga rejects the Russellian solution summarily, moving on rapidly to his own "proper function" explanation based on factors inaccessible to the subject. It is clear that Plantinga does not consider the Russellian analysis of the Gettier problem to capture the externalist implications he sees in Gettier. Warrant and Proper Function (Oxford: Oxford University Press, 1993), pp. 32-36.

44. Fumerton, pp. 204, 216-218.

45. Ibid., p. 221.

46. Ibid., p. 218.

47. Ibid., p. 215.

48. Ibid., p. 204.

49. Fumerton, p. 194.

50. Friedman, p. 156.

51. It is worth noting that the position outlined here is very different from John Pollock's position, which Fumerton discusses briefly (p. 193). Pollock's "conceptual analysis" is closely related to verificationism. The crucial concepts on his position are the concepts of empirical objects themselves, understood in terms of their justification conditions. (John Pollock, Knowledge and Justification (Princeton: Princeton University Press, 1974). It is certainly impossible to avoid using the concepts which play a part in a particular hypothesis when one is evaluating the argument for that hypothesis. But the analytic approach we advocate concentrates on concepts involved in the form of the argument itself.

52. Stewart Cohen, "Justification and Truth," Philosophical Studies 46 (1984), p. 291.

53. Fumerton, pp. 194-195.

54. Fumerton, pp. 108-111, 203.

55. And in fact, such a procedure yields the metaregress we discuss in Chapter 4.