Contemporary Epistemology III: Analysis

UC Davis Philosophy 102

Theory of Knowledge

Instructor: G. J. Mattey, Senior Lecturer

Version 3.1.1, May 20, 2015

The Analytic Project

We will examine the analytic project as it is carried out according to the method described in the introductory module. We will understand an analysis of the schema ‘S knows in circumstances c that p’ (or ‘Ks,c p’) as the proposal of conditions which are either necessary or sufficient for S to know in circumstances c that p. The most common analyses propose a set of necessary conditions, which, taken together, are proposed as being sufficient. The bulk of this module will be devoted to a discussion of proposed necessary conditions NC, of the form: ‘If S knows in c that p, then NC.’ Toward the end we will take a look at some attempts to provide necessary and sufficient conditions for knowledge.

Knowledge and Truth

Most epistemologists maintain that p’s being true is a necessary condition for S’s knowing in c that p. In the standard format, where the arrow indicates a relation of analytic implication, the condition is formulated this way:

Truth: If S knows in circumstances c that p, then it is true that p (or Ks,cp → Tp).
Epistemologists for the most part have accepted this condition without question. One of the primary tasks of ancient epistemology was to distinguish between knowledge and opinion, and the fact that one can have a false opinion was taken to be one of the main differences between the knowledge and opinion. In Plato’s dialogue Theatetus, which contains the first attempts at a kind of an analysis of knowledge in the philosophical literature, truth is taken to be a necessary condition given for knowledge. Early in the dialogue, Socrates describes knowledge as “unerring” (152c). Later, Socrates asks rhetorically, “can he who misses the truth of anything, have a knowledge of that thing?” and his interlocutor Theaetetus answers in the negative (186c). A proposed analysis that knowledge is simply opinion is quickly rejected: “I cannot say, Socrates, that all opinion is knowledge, because there may be a false opinion” (187b).

What the truth condition for knowledge means depends on how truth is understood. On the standard view, the truth of a proposition p is the correspondence between the state of affairs that p represents and reality. It is generally accepted that the same proposition cannot be both true and false, because the truth-value of the proposition is constrained by the world, which exists in only one way. If the proposition “corresponds to” the single way the world is, it is true, and if it does not so correspond, it is false. On this view, knowledge that p requires that p correspond to the unique way that the world is. If the proposition p does not so correspond, it cannot be known that p. So the truth condition is in the end a requirement that what is known “matches” what is real.

Some people think that truth is “relative,” perhaps to a person and a time. So a relativist might say that the proposition p is true for S1 in circumstances c, but the same proposition p is false for S2 in simultaneous circumstances c′. This position, when scrutinized, usually turns out to be about belief rather than truth. What it might express is that S1 takes it to be true in c that p, while S2 takes it to be false that p at that same time. It might be thought as well that there is no “fact of the matter” about the way the world is (or even that there is no “world” at all, as suggested by Nietzsche and Richard Rorty), so that the only use of attributions of truth or falsehood is to express a view of the world, or simply a view.

This kind of relativistic position was first espoused by the ancient Sophist Protagoras. According to Protagoras, the individual subject is the “measure of all things” (Plato, Theaetetus 152a). He uses the examples of heat and cold: it is true that a wind is cold if I feel it as cold, and it is true that the same wind is warm if you feel it as warm. The wind is both cold for me and warm for you. And so it is true for me that the wind is cold and true for you that the wind is warm. A problem with this view is that it takes as its paradigm cases what might be called subjective states of the individual. While we may disagree about how the wind feels to us, there seems to be an objective fact of the matter as to whether or not it is a certain temperature.

A more sophisticated relativism is explored by John MacFarlane. (See his technical paper, “Making Sense of Relativism” for details.) On his view, the truth-value of at least some sentences is relative to the context in which that truth-value is assessed. These sentences tend to have normative content, which means that their truth-value is relative to a standard. Judgments of beauty are of this kind. The reason for relativity is that we bring different standards to different contexts of assessment. If such a semantical theory is adopted for at least some classes of sentences, then the condition that if S knows in c that p, then it is true that p, will have to be understood relativistically in cases where p contains the kind of content which invokes different standards.

Some contemporary epistemologists (and many scholars in the humanities and social sciences) embrace the thesis that knowledge is relative to the knower. They are willing to attribute to people knowledge in c that p as well as knowledge in c that not-p, albeit usually not by the same person at the same time. One reason for allowing the possibility of knowledge of both a proposition and its negation is a relativistic view of knowledge attribution. One attributes knowledge “correctly” when the attributee meets the epistemic standards of the attributor at the time of attribution. Because standards vary from person to person and from time to time, correctness of attribution varies as well. Thus one might wish to say that Aristotle knew that the sun moves around the earth, while Newton knew that the earth moves around the sun. (This is, of course, a curious result only if one holds that the two descriptions of motion are incompatible.) If this kind of assertion is made, there is the apparent consequence that one may know what is false. One might respond by holding simply that knowledge has no truth-condition or else by embracing some form of relativism about truth.

Knowledge attributions are frequently made independently of any relation between what is said to be known and what is true. Most often, the claim that I, you or someone else “knows” is an indication of the intensity of belief. For example I might proclaim, “I just know that I will win the lottery.” Even after I lose, I might still maintain, “I knew I would win, even if it didn’t turn out that way.” Attributions made without any regard to truth are far from the core of attributions which are of interest to epistemologists. This is not to deny that they are made, or even frequently made. But this kind of attribution is generally regarded by epistemologists as lying outside the range of epistemological inquiry, where some connection to truth, even a relativistic one, is presumed.

The denial or relativization of the truth condition on knowledge may be made on independent grounds, such as a view about the nature of truth. However, it is most natural to reject or relax the condition because it removes a constraint on knowledge attribution, with the result that it becomes much easier to attribute knowledge. This is most apparent in the case of the relativization of truth to epistemic standards. In the case of disagreements across cultures, it seems to many to be presumptuous to claim that one culture has knowledege and another lacks it, since any such judgment must itself be made from within a cultural context. One remedy for this discomfort is to allow knowledge all around.

Knowledge and Belief

It is commonplace for analyses of knowledge to contain a condition resembling the following:

Belief: If S knows in c that p, then S in c believes that p (or Ks,cp → Bs,cp).
It is held by Bayesians that belief is a matter of degree: one can believe that p to a greater or lesser extent. This extent might be measurable by the kind of wager one might make that the belief is true. The higher the odds a subject is willing to accept against a proposition p’s being true, the more strongly the subject believes that p. The strength of belief is formally represented by a probability function, the (subjective) probability for S that p in circumstances c. If a Bayesian account of belief is to be adapted to the analysis of knowledge, then it must be determined how high a probability p must have in order to meet the belief condition. On the face of it, it would seem arbitrary to select any given threshold. One might retreat to vagueness and hold that the belief must be very strongly held, but such a condition runs counter to the quantitative nature of the Bayesian account of belief.

A contrasting approach is to take belief, at least for the purposes of the analysis of knowledge, to be binary in nature: S either believes in c that p or S does not believe in c that p. It would then be the task of the epistemologist to determine the conditions under which one believes or does not, and this might call for an analysis of belief. Alternatively, it may be helpful to recall Austin’s account of knowledge as giving one’s word. The giving of one’s word appears to be binary in nature and not to admit of degrees.

One way to look at the belief condition is in terms of what seems to be a feature of knowledge itself. We make attributions of knowledge, at least frequently, to indicate that what is said to be known is, so to speak, a settled matter, not calling for further discussion. (Do you know that it’s safe? Yes? Good, then let’s go!) In this way, we can look at S’s belief that p in terms of its being settled in S’s mind that p. We frequently think of the kind of belief required for knowledge in terms of being certain. A fallibilist would add that certainty does imply truth, but only a high degree of assurance or confidence in the truth of p.

Most epistemologists regard knowledge as a species or kind of belief. Given the truth condition, it would be said that knowledge is the kind of belief that is true (and perhaps has other properties). We will see shortly, however, that not all epistemologists view knowledge as a species of belief.

That belief, however understood, is a necessary condition of knowledge and that knowledge is a kind of belief seems to be supported by the fact that standard analyses of knowledge contain a further condition that makes essential reference to belief. This kind of reference can be found in both “internalist” and “externalist” analyses to be discussed below. A typical internalist analysis might contain a condition of this sort:

Evidence: If S knows in c that p, then S’s belief in c fits S’s evidence for p.
The whole point of having evidence is to provide backing for belief. A variant of externalism might have as a condition:
Reliability: If S knows in c that p, then S’s belief in c that p was produced by a reliable belief-forming process.
Here, belief is the end-product of the kind of process which the externalist claims to be necessary for knowledge.

One might reject a belief condition because it is incompatible with a very liberal policy of knowledge attribution. A robotic vacuum cleaner that has no consciousness might be said to “know” that it should turn when it hits something, but it is hard to attribute any beliefs to it. All it is doing is following a set of instructions embedded in a microchip. Perhaps all that is needed for “commitment” is that the potential knower be able in some way to represent the world and be disposed to act on those representations.

Or take those situations in which someone is willing to attribute knowledge to a subject who is unwilling to express a commitment to the truth of p. Suppose that you have learned that p but have forgotten that you have learned it. You use what you have learned to give the correct answer to a question, but without committing yourself to the answer at that time. It is tempting for someone who is aware of the fact that you have learned that p to attribute knowledge to you, saying “You knew it all along.” (See Keith Lehrer, Theory of Knowledge, Chapter 2.)


We will begin with the requirement of stability, which is not much discussed in contemporary epistemology but which was quite important in traditional epistemology. Ancient philosophers, as well as Descartes, emphasized that knowledge must be stable. According to the Stoics, “Scientific knowledge is a cognition which is secure and unchangeable by reason” (Stobaeus, 2.73,16-74.3). These philosophers held that nothing could epistemically undermine S’s believing p to be true. (It is, of course, quite possible for a belief to be destroyed in any manner of non-epistemic ways, such as through brainwashing, brain damage, etc.) A fallibilist could require only that belief need only be highly stable.

A belief might be stable in the sense that it is hard to dislodge by new information.

Psychological Stability: If S knows in c that p, then S’s belief in c that p cannot be given up on epistemic grounds in the face of any further information that S might receive.

This condition is best suited to an infallibilist view of knowledge as held by Plato and Descartes, as the following two passages show.

For true opinions, as long as they remain, are a fine thing and all they do is good, but they are not willing to remain long, and they escape from a man’s mind, so that they are not worth much until one ties them down by (giving) an account of the reason why. . . . After they are tied down, in the first place they become knowledge, and then they remain in place. (Plato, Meno, 97b-98a)
Accordingly, even if I am no longer attending to the arguments which led me to judge that this is true, so long as I remember that I clearly and distinctly perceived it, there are no counter-arguments which can be adduced to make me doubt it, but on the contrary I have true and certain knowledge of it. (Descartes, Meditation Five, 1641)
To put the Cartesian version of psychological stability in Platonic terms, the clarity and distinctness of perceptions is what “ties down” beliefs when the perception is current or remembered. One contemporary account of knowledge that incorporates psychological stability (though as a consequence, rather than a condition, of knowledge) is Williamson. He claims that someone who knows that p will remain more firmly committed to the truth of p in the light of evidence against the truth of p than will someone who merely has a true belief that p.

We may distinguish from psychological stability a practical or pragmatic kind of stability, in the sense in which knowledge requires that one treat p as a proposition that can be counted on, or whose truth can be taken as given in the course of future inquiry or action.

Pragmatic Stability: If S knows in c that p, then p is so settled as to be available as the source of future inquiry. (John Dewey, Logic: The Theory of Inquiry, 1938, p. 9)
Dewey took this to be a sufficient condition for knowledge as well, implicitly rejecting an independent truth-condition. For Dewey, truth is an epistemic concept of “warranted assertability.”

In contemporary epistemology, the idea that knowledge cannot be undone is generally supplanted by the normative condition that knowledge should not be undone, in the sense that there is no further information in the face of which S should give up the belief that p. Another factor for the lack of attention paid to stability is that it may seem to follow from the belief condition on knowledge. A commitment to the truth of p would seem to have an inherent stability.

Belief and Acceptance

Some epistemologists, among them Keith Lehrer, have proposed that knowledge (in its most central sense) requires that the belief that p be rational in character. Rational belief is called by Lehrer “acceptance.”

Acceptance: If S knows in c that p, then S accepts in c that p is true.
Rational belief is the outcome of rational deliberation on the basis of evidence. Hence, an account of knowledge that requires acceptance is at least in this respect internalist. Lehrer understands the goal of rational deliberation to be the fulfilment of an “epistemic goal” of being committed to as many truths and as few falsehoods as possible (Theory of Knowledge, p. 11). William James originally expressed the matter in terms of epistemic duty in the face of uncertainty: “We must know the truth; and we must avoid error--these are our first and great commandments as would-be knowers” (“The Will to Believe,” reprinted in The Will to Believe and Other Essays, p. 17) James here has in mind the deliberate choice of what to believe, and he suggests that this choice is best made on the basis of rational deliberation.

Lehrer goes on to assign to acceptance a certain functional role in our thought, inference, and action (Theory of Knowledge, second edition, p. 35). The role is that of being assumed to be true, which we might describe as being settled in one’s mind. Lehrer claims that a subject who accepts that p will be willing both to draw certain inferences and perform certain actions assuming the truth that p (Theory of Knowledge, second edition, Chapter 2). Two such actions are to be ready, in appropriate circumstances, to affirm or concede that p, and to be ready to justify the claim that p.

The fact that some philosophers try to eliminate a belief condition and others to strengthen it seems to indicate that knowledge attributions are made in many ways, and that attempts to find some preferred manner of attribution are unlikely to succeed. One can appeal, as does Lehrer, to a conception of knowledge “that is characteristically human” (Theory of Knowledge, second edition, p. 4), but it is not clear that there is much prospect for agreement about whether there is a “core” conception of knowledge or what it would look like.

Knowledge as a Mental State

Recently, there has emerged a new kind of challenge to the thesis that knowing is a state of believing. Timothy Williamson in Knowledge and its Limits (2000) has argued that knowing is a “state of mind” which is not a state of believing. Williamson’s argument is complicated, but it is not hard to understand its basic form. We begin with a rough classification according to which knowledge is a species of belief.

Williamson begins by comparing knowing with the mental state of perceiving, which is a “factive” mental state. One perceives that something is the case (in what can be called a “narrow” sense of perceiving) only if it really is the case or is a fact. We are told by Williamson that the essence of a factive state “includes a matching between mind and the world.” (Knowledge and its Limits, p. 40.) Knowing is like perceiving in this respect. Williamson calls knowledge “the most general factive state” (p. 48).

Belief is not a factive mental state: one can believe what is not the case. Williamson’s argument is that believing does not become a factive state by being true, or by being true and justified, or by being true and reliably formed, etc. (Or in terms used here, being true and sufficiently warranted is not sufficient to make belief factive.) Belief by its very nature is not factive, and so any mental state that is a kind of belief is by its very nature not factive either, even though some beliefs are true. Given that knowing is the generic factive mental state, and that such states as perceiving and remembering are factive states as well, we can construct a new kind of hierarchy.

Note that Williamson’s proposal does not preclude describing the state of mind that is knowledge as being in some way settled. Indeed, Williamson takes the value of knowledge to lie (in part) in our willingness, when we have knowledge, to continue investigation even in light of unfavorable evidence. When one knows that p, “one is less likely to lose belief in p in the course of interacting with the environment by discovering new evidence which lowers the probability of p” (p. 8). (Perhaps perceiving has this character as well, as when someone discounts negative perceptual evidence by exclaiming, “I see it with my own eyes!”)

Nonetheless, Williamson contends that knowledge cannot be analyzed in the traditional manner. He claims that no analysis is capable of ruling out Gettier cases (which will be discussed below) except through circularity. This is a serious claim that will be discussed briefly at the end of the module.

Knowledge and Warrant

The third category of belief-related conditions on knowledge is by far the most extensively treated in the literature. Alvin Plantinga uses the term ‘warrant’ to indicate “that, whatever precisely it is, which together with truth makes the difference between knowledge and mere true belief” (Warrant: The Current Debate, p. 3). Warrant may be thought of as a variable quantity, of which one may have more or less, so the condition on knowledge is that there be “sufficient” warrant.

Sufficient Warrant: If S knows in c that p, then S in c has sufficient warrent with respect to p (or Ks,cp → WKs,cp).
Since the very definition of sufficient warrant is that it is the ingredient that produces knowledge when added to truth and belief, the three conditions Truth, Belief, and Sufficient Warrant are meant to be sufficient for knowledge when all three are present.
Tripartite Analysis of Knowledge:
Ks,cp ↔ (p ∧ Bs,cp ∧ WKs,cp).

But what is warrant? Generally it is expressed in conditions which involve other, more specific, notions. A rather diverse array of warrant conditions on knowledge have been proposed over the ages. In this section, a number of them will be displayed in a systematic way, with some references to classical sources. This classification will be followed by a critical discussion of a number of the proposed accounts of sufficient warrant. The Gettier problem will be dealt with separately.

Perhaps the most generic way of understanding warrant is as involving a connection between the first two components of the analysis, truth and belief. A minimal, and negative, way of understanding that connection is in terms of its not being an accident that the belief is true, when one knows.

No-Accident: If S knows in c that p, then it is not at all accidental that S in c is right about its being the case that p. (Peter Unger, “An Analysis of Factual Knowledge,” 1968)
For this condition to be fleshed out, it must be determined what it is that makes being right accidental as opposed to non-accidental. Some paradigm cases of being accidentally right is are when one’s belief is based on a guess, on bad reasoning, or on wishful thinking. But the subject is also right by accident in many of the Gettier cases, where no suspect bases of belief are involved. In many of the cases, the subject’s reasoning is impeccable, but it is still only by coincidence that the belief is true.

We have expressed something like this condition in our running metaphor that knowledge requires that one’s epistemic position be relatively strong.

Epistemic Position: If S knows in c that p, then S in c is in a strong enough epistemic position with respect to the truth of p.
This condition is quite schematic, since it does not specify what an epistemic position is, what makes an epistemic position stronger or weaker, or when an epistemic position is strong enough. It only suggests that in some cases S is better situated with respect to the state of affairs expressed by the proposition p than in other cases. If S’s belief that p is a consequence of S’s being in a strong epistemic position, then we might say that it is no accident that the belief is true when it is.

One finds something like this requirement in Plato, who describes an event that can be known to have taken place only by someone who has witnessed it. “And when a jury is rightly convinced of facts which can be known only by an eyewitness, then, judging by hearsay and accepting a true belief, they are judging without knowledge” (Theaetetus, 201b). According to this description, an eyewitness is in a strong enough epistemic position to allow knowledge, while the recipient of hearsay is not in a strong enough epistemic position. Or, it is not at all an accident when the jury members form true beliefs based upon what they actually see, while it is to some extent an accident when the jury members form true beliefs on the basis of the testimony of witnesses. Descartes in the Third Meditation describes circumstances in which a “natural light” in his intellect makes his perception clear and distinct enough to give him knowledge, while in other circumstances the prejudices of his youth and education dulls his intellectual perception and makes him ignorant.

There is considerable disagreement in epistemology concerning each of the three questions just raised about epistemic position and warrant. Perhaps the deepest disagreement concerns how to understand the nature of the epistemic position required for warrant. The position taken on this point strongly influences the account of what it is that makes an epistemic position stronger or weaker. There is also disagreement with respect to how strong a subject’s epistemic position must be in order for there to be knowledge. Before turning to these topics, we shall take a brief look at the traditional, but now largely abandoned, claim that whatever the nature or source of one’s epistemic position, it must be so strong as to avoid the possibility of error.

A widely-adopted approach to a “no-accident” condition is to require that S’s epistemic position be strong enough that S is able to rule out alternatives in which p is false. One can guarantee that the subject is right non-accidentally with the following condition:

Elimination of Alternatives: If S knows in c that p, then S in c is in a position to rule out all alternatives in which p is false.
This condition, which is similar to that of Lewis discussed above, is tantamount to being an infallibilist condition. If one really can rule out all alternatives in which p is false, then S cannot go wrong about the truth of p. A fallibilist version of the condition would weaken it so that only a certain range of alternatives can be ruled out by S.
Relevant Alternatives: If S knows in c that p, then S in c is in a position to rule out all relevant alternatives in which p is false. (Fred Dretske, “Epistemic Operators,” 1970)
Given that S really is in a position to rule out all relevant alternatives in which p is false, the truth of S’s belief would not be an accident, but would stem rather from the strength of S’s epistemic position.

A variant of the relevant alternatives condition is one which requires that S have the ability to discriminate between those situations in which p is true and those in which p is false.

Discrimination: If S knows in c that p, then S in c can discriminate the actual state of affairs in which p is true from relevant alternative states of affairs in which p is false. (Alvin Goldman, “Discrimination and Perceptual Knowledge,” 1976)
Presumably, if S could so discriminate between the relevant alternatives, then S is in a strong enough position to rule out all relevant alternatives in which p is false, since S can discriminate between them. Conversely, if S can rule out all the relevant alternatives in which p is false, then S must be able to discriminate between states of affairs in which p is true from those in which p is false.

The variants of the condition that alternatives in which p is false must be ruled out do not specify how these alternatives are to be eliminated. Differing accounts of the nature of epistemic position, specifically the internalist and externalist accounts discussed below, will fill out these conditions in different ways.


We have now taken a look at the two most basic positions regarding the nature of the kind of epistemic position required for knowledge. A further question is how strong that epistemic position must be. The most powerful epistemic position with respect to any proposition would be one in which the subject is not capable of being wrong about the truth of p.

Infallibility: If S knows in c that p, then S in c could not be mistaken about whether p is true.
The tradition in epistemology until the twentieth century was to assume an infallibility condition for knowledge. The demand for infallibility was generally confined to knowledge in the fields of logic, mathematics, and metaphysics, where (it was thought) we are in an infallible epistemic position because our knowledge is of “eternal truths” about these matters which can be discovered a priori, or independently of experience. Infallibility with respect to existence was generally limited to cases involving one’s present mental state or one’s own existence when thinking about whether one exists.

The infallibility condition is so strong that it seems to rule out most of what in everyday life we are inclined to call “knowledge,” which concerns contingent matters of fact. So it has largely been abandoned by contemporary epistemologists, who are primarily concerned with everyday knowledge, which they think exists. Two philosophers who were influential in this regard were the American C. S. Peirce and the Englishman G. E. Moore.

A couple of apparently infallibilist conditions have been placed on knowledge by contemporary philosophers, David Lewis and Fred Dretske.

Uneliminated Possibilities: If S knows in c that p in c, then p holds in every possibility left uneliminated by S’s evidence. (David Lewis, “Elusive Knowledge,” 1996)
As noted in the last module, Lewis’s condition is misleading, in that on his account, the word ‘every’ is implicity qualified. One is “infallible” only with respect to a certain limited range of possibilities, i.e., ones that are not “properly ignored.” (This condition is discussed further below, in its application to the Gettier problem.)

Dretske’s condition is an internalist condition limited to knowledge on the basis of reasons or evidence.

Conclusive Reasons: If S knows in c that p simply on the basis of reasons R, then R would not be the case unless p were the case. (Fred Dretske, “Conclusive Reasons,” 1971)
The idea is that one would not have the reasons one has unless one were correct with respect to the truth of p. So given that S has such reasons, p must be true. This condition is important less because of its content than because of the way in which it is formulated, using a subjunctive conditional, which can be re-phrased as, “If p were not the case, then R would not be the case.” We will see the subjunctive conditional used in other proposed conditions of knowledge.

Dretske’s condition may be infallibilist only on the surface. It is stated in the subjunctive mood, which appears to be weaker than the following indicative modal condition: it is impossible for R to be the case unless p is the case. However, Dretske himself claims in “Conclusive Reasons” that the subjunctive rendering can be understood in this way: having reasons R “eliminates not-P as a possible state of affairs.” As Pappas and Swain point out, the condition understood in this way is very strong, and would rule out most knowledge attributions (“Some Conclusive Reasons against ‘Conclusive Reasons’,” 1973). They suggest that Dretske might have had in mind a weaker condition, which requires only that the actual situation be changed only enough to make p false. This is how the subjunctive conditional is treated in standard semantics, as will be seen below.

As has been noted, infallibilism has fallen out of favor in terms of fallibilism, according to which sufficient warrant may fall short of infallibility. The requisite strength of epistemic position needed for knowledge then becomes an issue, and will be treated as it arises in the context of our treatment of various accounts of what warrant is and how it arises.

Internalism vs. Externalism

Internalists and externalists differ fundamentally over the nature of the required epistemic positions. The differences between the two can be described in different ways. One way, suggested by Wilfrid Sellars (before the introduction of the terms ‘internalism’ and ‘externalism’), is to say that internalists require that the requisite epistemic position with respect to a belief be understood in terms of the location of that belief with the logical space of reasons. [I]n characterizing an episode or a state as that of knowing, we are not giving an empirical description of that episode or state; we are placing it in the logical space of reasons, of justifying and being able to justify what one says (Empiricism and the Philosophy of Mind, Section 36).

Logical Space of Reasons: If S knows in c that p, then S’s belief that p must have a place in the logical space of reasons.
Thus, if a belief (or what one says) is merely caused by the the world in a certain way, or is merely formed in a reliable way, it does not stand in the logical space of reasons and hence does not qualify as knowledge, given the Sellars condition.

We shall use the term ‘evidence’ to indicate the reasons within the logical space of reasons. We can then re-formulate the previous condition in terms of evidence,

Evidence: If S knows in c that p, then S’s belief that p in c fits the evidence for the truth of p. (Adapted from Feldman and Conee, Evidentialism, 1985).
Of course, the term ‘fits’ must be explained in order to flesh out how fitness of evidence constitutes warrant, an issue that will be discussed in what follows. One thing that can be said is that for evidence to play a role in the activity of justifying a subject’s beliefs, the subject must have some kind of access to one’s evidence for the truth of what one says or believes. This suggests a derivative conditions for internalists:
Awareness: If S knows in c that p, then S must be able to be aware of the evidence for S’s belief that p.

We may understand externalism to be the denial of internalism. Thus, for an externalist, a belief may be an item of knowledge despite its standing outside the logical space of reasons. Thus, an externalist might identify warrant with the cause of a belief or the fact that it was reliably formed, though beliefs about them may lie within that space. We will turn to a discussion of externalism after exploring a number of issues that arise from attempts to give internalist accounts of warrant.

One question that arises immediately is why one might adopt the internalist condition. One can certainly cite tradition: most of the accounts of knowledge developed before around the middle of the twentieth century are most easily interpreted as being internalist. Some are driven to internalism because of their responses to purported cases of knowledge, one of which will be discussed shortly. But there are some principled reasons why internalism might be adopted.

The most influential internalist account of knowledge can be traced back to Descartes in the seventeenth century. The question Descartes raised in his Meditations was the extent to which he himself had knowledge. In other words, he was concerned with epistemic self-attribution. In order to assess whether his epistemic position was strong enough regarding a number of different kinds of propositions, Descartes had to take a perspective on himself, what Sosa calls a meta-perspective. To answer his own question of whether he knows, he had to appeal to evidence placed within the logical space of reasons. The requirement of evidence in self-assessment then trickles down to become a condition of knowledge itself: if you know, then you must be able to tell how you know, and this requires being able to justify your beliefs.

But that is not the end of the story. Descartes required that for knowledge that p, one’s epistemic position with respect to p must be as strong as it can be, leaving no room for doubt. And the only way to be in this maximally strong position is to possess accessible evidence that is capable of dispelling possible doubt. Thus, constraints on the strength of one’s epistemic position may drive one to internalism.

Epistemic Deontology

There is a third reason for internalism that has been attributed to Descartes, as well as to internalists that have followed him. Put in its most general form, the idea is that beliefs that amount to knowledge must be “permissible” to hold in order to be warranted.

Permissibility: If S knows in c that p, then it is permissible for S in c to believe that p.
This feature of the condition is built into a number of knowledge conditions.

One way to construe permissible believing is as “responsible” believing, in a sense of “responsible” that is either straightforwardly moral in character or else modeled on the notion of moral responsibility.

Responsibility: If S knows in c that p in c, then S in c has come to believe p in a responsible way. (Attributed to Descartes and Locke by Plantinga, Warrant: The Current Debate, 1993; see also Alvin Goldman, “Strong and Weak Justification”, 1988)
Plantinga contends that accounts of knowledge in Descartes and Locke are “deontological” in the sense that moral requirements are relevent to knowing. If one knows, one has done one’s moral duty to use one’s God-given faculties responsibily and therefore cannot be morally blamed for believing what he does. The contemporary epistemologist Roderick Chisholm claims that the requirements accompanying knowledge have to do with intellectual integrity than morality (Theory of Knowledge, second edition, 1977, pp. 14-15). One’s intellectual duty is to believe as many truths, while avoiding the belief in as many falsehoods, as possible.

A related notion was articulated by A. J. Ayer.

Right to Be Sure: If S knows in c that p, then S in c has the right to be sure that p. (A. J. Ayer, The Problem of Knowledge, 1956)
Ayer contends that there is variation in the concession that S has the right to be sure, and that trying to fix the exact point at which one has that right does not belong in the analysis of knowledge, any more than the standards of goodness belong in the analysis of ‘good.’ (Here he anticipates contextualism, which is discussed in the previous module.)

The condition of permissibility suggests that what the epistemic subject believes is (at least to some extent) up to the epistemic subject. Plantinga argues that this connection between warrant and choice gives rise to a general condition that ties knowledge to something which would allow the subject to believe responsibly or not. Whether or not this is correct, the notion of permissibility can be worked into the analysis by requiring that S have internally available resources that would allow S to make a permissible choice. And what makes the choice permissible is that it lies in the right part of the logical space of reasons.

The Current Debate

The current debate between internalists and externalists is played out in different ways. One way is simply a battle of intuitions. Internalists construct cases in which beliefs which lie outside the logical space of reasons, but which otherwise have strong epistemic credentials, claiming that the epistemic subject lacks knowledge. Thus in BonJour’s example, Norman is a reliable clairvoyant who forms a true belief based on his reliable clairvoyance, but with no access to the fact that he has this ability. BonJour concludes that he does not know, but this intuition is not shared by externalists.

The basis for the internalist intuitions may lie in two (or more) places. One possibility, suggested by Sosa, is that knowledge that requires a meta-perspective is of a special kind, what he calls reflective knowledge. This is distinguished by Sosa from what he calls animal knowledge, which lies outside the logical space of reasons. Thus, Descartes adopted an internalist standpoint because he was seeking reflective knowledge.

A second basis for rejecting the claim that Norman knows is that Norman does not meet the permissibility condition or its variants. Norman has no right to be sure, even if he is sure, because he lacks accessible evidence for the truth of his belief. Of course, this leaves the door open for a further distinction between deontological or non-deontological senses of knowledge. Perhaps one way to make such a distinction is to think of knowledge as deontological insofar as knowledge attributions are made the basis of our own actions. If I am to act responsibly, I ought to have a permissible belief, i.e., one for which I have justification within the logical space of reasons. For Descartes, believing is a kind of action for which we are responsible.

On the other hand, there are few current defenders of Cartesian certainty as the degree of strength of epistemic position required for knowledge. If certainty is not needed, then one of the motivations for requiring internal accessibility is removed. The virtual abandonment of certainty as a condition for knowledge thus opens the door for externalist accounts of warrant.

As was noted above, one motivation for internalism is the taking of a first-person point-of view, in which one is deciding whether to attribute knowledge to himself. This requires that the person take into account the evidence that is available to him. But third-person attributions make no such demands. When deciding whether to attribute knowledge to another subject, one may take into account that subject’s external, so to speak, circumstances. If I am interested in whether Norman knows, it is open to me to take into account powers that Norman has to which he is unaware himself. Perhaps the clash of intuitions over this case is largely a matter of whose point of view is taken. I, from the outside, might be willing to concede that Norman knows, but Norman, from the inside, might resist the claim that he knows.

There are other motivations besides point-of-view that favor externalism. One such motivation is that internalism is simply too restrictive. Internalism denies knowledge to those who operate outside the logical space of reasons and who at the same time seem to be in very strong epistemic positions regarding what they believe. This includes not only those like Norman who possess powers of which they are unaware, but also small children and non-human animals.

Other motivations for externalism are more technical in nature. Externalists have claimed that confining warrant to the logical space of reasons leads to problems. One claim by externalists is that even if access to evidence is a necessary condition for knowledge, it cannot be a sufficient condition, at least if one’s epistemic position need not be maximally strong. As the Gettier cases show, one might have a true belief that is well-situated with respect to the logical space of reasons but violates the No-Accident condition on knowledge, that one does not know if the truth of one’s beliefs is accidental. As a result, some internalists with respect to warrant have held that sufficient warrant requires an additional non-internal condition. One such condition will be discussed below.

A second technical problem for internalist accounts of warrant comes to light when we take a deeper look at the logical space of reasons. Internalists would agree that not only must a warranted belief be supported by evidence, but the evidence that supports it must be good evidence. The question for the internalist is what makes the evidence for a belief good. In a small range of cases, it can be claimed that beliefs are self-evident, but most beliefs are backed by something else which provides support for them.

Now the question arises as to how the supporting evidence can be good evidence. If appeal is made to other supporting evidence, a regress threatens, in that the same question can be asked about the goodness of that evidence. Here, the externalist offers a life-line. What makes a piece of evidence good, he claims, is its origin or source. For example, evidence which has its source in reason alone might be considered good evidence. Or evidence whose source is reliable, as with sense-perception in people with well-functioning senses.

The Logical Space of Reasons

Let us suppose, with the internalists, that knowledge that p requires that p have an appropriate place in the logical space of reasons. And let us suppose further that sufficient warrant is fallible. This opens the door to myriad other conditions, many of which will be discussed below. For now, the question arises as to the organization of that logical space.

Most, if not all, internalists would agree that an appropriate metaphor for the proper organization of the logical space of reasons is one of evidence supporting a belief. Corresponding to this metaphor is that of evidence that undermines a belief. The former is conducive to knowledge; the latter is conducive to ignorance. We shall here state a condition on knowledge that is generally accepted by externalists as well as internalists:

Negative Coherence: If S knows in c that p, then the sufficiency of S’s warrant for believing that p is not blocked by evidence or beliefs in S’s possession.
Thus, even if evidence, in the internalist’s sense, is not a necessary condition for knowledge for an externalist, it is capable of blocking knowledge. For example, if Norman were in possession of evidence that he was not clairvoyant, then even an externalist would allow that he lacks knowledge. As BonJour (himself an internalist) set up the example, Norman has no idea whether he is or is not clairvoyant, because if he thought he was not, that fact in itself would be sufficient to discourage any claim that he knows.


The internalist camp is deeply divided, and one of the main reasons is disagreement about the nature of evidence. Notice that the examples given earlier are all in sentential form. One view of evidence is that it must have conceptual content, which is expressed in the predicates used in the sentence. If I am to be able to make use of evidence, on this view, it must be in the form of a conceptualized judgment.

Presumably the judgments which constitute evidence must be my beliefs or acceptances—at any rate, something to which I would be willing to take as settled premises in inferences. In that case, evidence is identical to belief, or perhaps to rational belief.

Doxastic Assumption: If E is evidence for S in believing that p, then S believes that E.
John Pollock calls accounts of evidence of this kind “doxastic theories.”

A non-doxastic account of evidence would allow that at least some evidence is not in the form of belief. Typically, it is held that pre-judgmental “experience” can constitute evidence. Pollock, for example, claims that we make reference to such evidence by a kind of mental “pointing” at it. (See his Contemporary Theories of Knowledge).

Doxastic theorists criticize this kind of view by claiming that uninterpreted experience cannot play the role of evidence. One way of seeing the argument is modeled on Plato’s paradox of the learner. If we have not already conceptualized our experience, then how could we tell which experience it is appropriate to “point” to as evidence for what we believe? As Immanuel Kant put it in the eighteenth century, “intuitions without concepts are blind” (see Critique of Pure Reason, A51/B75). This line of defense of the doxastic assumption is developed by Wilfrid Sellars in the mid-twentieth-century; see especially Empiricism and the Philosophy of Mind,, Section 6).

As was indicated above, this is not a problem for externalists. They can hold that experience can be evidence without our being able to identify it as evidence. It seems that the non-doxastic position is favored by some primarily because of difficulties with doxastic theories. We will discuss the non-doxastic position more fully after looking at the main outlines of doxastic accounts of the nature of evidence within the logical space of reasons.


For a doxastic theorist, support is a relation among beliefs. The primary dispute among doxastic theorists concerns the way in which the beliefs constituting evidence support further beliefs. Is the relation linear or non-linear?


It is a widely-held view that the support relation has a tree structure which is transitive and asymmetrical. If support is transitive, then given that the belief that p supports the belief that q and the belief that q supports the belief that r, it follows that the belief that p supports the belief that r. Transitivity is needed if chains of evidential support are to be built (and it seems that they must be if we are to be warranted in believing very much).

The support relation is asymmetrical if it meets the following condition. If the belief that p supports the belief that q, then the belief that q does not support the belief that p. Asymmetricality is favored because it (along with transitivity) precludes the possibility of circularity in the support relation, and circularity is taken to be incompatible with the notion of support.

This kind of account of the support relation invites a regress problem, noted above, of the kind encountered by Aristotle in his account of “demonstrative knowledge.” The regress is stopped by claiming that some beliefs are “self-evident.” This could mean that they are evidence for themselves, but that account violates the asymmetry condition that blocks circularity in justification. So the preferred meaning is that some beliefs are evidence “on their own” without needing anything else to make them evidence. These beliefs are foundational.

We have already seen the basis for rejecting the possibility of self-evident beliefs. All beliefs have propositional content, and all propositions make some connections between the elements (notably concepts) which make them up. There must, it is argued, be a basis for making that connection. The only resource available for connecting the elements of the propositions are beliefs about those elements. And since these other beliefs are in part what make a belief evident, no belief is self-evident.


If the anti-foundationalist argument is taken by a doxastic theorist to be sound, skepticism can be avoided only by dropping one of the features of foundational justification. The coherence theorist rejects the asymmetry requirement. The coherentist picture is that the support relation is “mutual,” constituting a “web” of beliefs.

The best worked-out coherence theory is that of Keith Lehrer, in Theory of Knowledge. (Laurence BonJour presents an elaborate account of coherence in The Structure of Empirical Knowledge, 1985, but he has not developed it further and indeed has abandoned it altogether.) The chief difference between the two approaches is the role they assign to coherence. For Lehrer, it is a condition for S to be justified in accepting that p that p coheres with the body of what S accepts. He then goes on to lay down an account of the conditions for coherence. For BonJour, p is justified for S just in case p is a member of a coherent body of beliefs. Much of BonJour’s work is then to spell out the conditions for the coherence of a body of beliefs. We will not describe such conditions here.

Lehrer’s account takes as evidence the rational acceptances of a person. He assumes that all propositions can be ranked in a partial ordering in terms of how reasonable it is to accept them, given that one accepts other propositions. (That is, some are more reasonable to accept than others, but some may be equally reasonable to accept.)

Accepting a proposition may raise or lower the reasonableness of accepting others. Lehrer calls the reasonableness-lowering propositions “objections” to those whose reasonableness is lowered. So, accepting that I am now dead drunk lowers the reasonableness of accepting that pink elephants are dancing before me. Note that the notion of ruling out objections is related to the generic condition that the truth of the belief not be accidental. One way of fleshing this out is to say that the subject can rule out all alternatives in which the belief is false. Lehrer proposes that the range of alternatives in question includes not only situations in which the belief is false, but also those which make it less reasonable to believe it. On the other hand, such situations need not be ruled out, but need only be not as reasonable to accept as the belief in question.

The coherence idea is that accepting p may make it more reasonable to accept q, while accepting q at the same time makes it more reasonable to accept p. So, accepting that I see a cow makes it more reasonable to accept that I hear mooing, and accepting that I hear mooing makes it more reasonable for me to accept that I see a cow. This relationship of mutual support is not straightforwardly circular. While accepting p may raise the reasonableness of accepting q, it is only one of many other things that sets the level of reasonableness of accepting q. Certainly I do not accept that I see a cow solely on the basis of the belief I have that I hear mooing!

The main foundationalist objections to coherence account of warrant (other than the circularity objection) are directed to the question of the adequacy of the support relation, and so we will turn to that now.


Doxastic foundations theories place the support relation on an allegedly firm basis of self-evident beliefs. These beliefs support other beliefs in the manner in which premises support the conclusions of inferences. If the truth of a belief that p follows deductively from the truth of self-evident beliefs, then the support is all that it could be: it is impossible for p to be false given the truth of the self-evident beliefs.

But given the limitation of self-evident beliefs to experience (and perhaps rational intuition as well), what is deductively supported by self-evident beliefs will be very limited as well. Thus, most non-self-evident beliefs, including beliefs about states of affairs beyond experience, will have to be made evident non-deductively. The problem then is to show how any non-deductive mode of support can ever be adequate. In particular, we can never by deductive means eliminate alternatives in which the non-self-evident belief is false, since it is at least logically possible that they are true.

Coherence theorists can point to the coherence relation as overcoming the limitations of non-deductive support. Consider Lehrer’s coherence account of warrant. Support offered to p by what one accepts is adequate when one is able to overcome any objection to p by appeal to the pool of acceptances. The coherence relation does allow “elimination” of skeptical hypotheses through the mechanism of overcoming objections.

The foundationalist argues that this solution comes at too high a price. In the first place, it does not discriminate among the acceptances that constitute the pool of evidence. Anything that one accepts (in the earnest search for the truth) will count equally well as evidence. The fact that the whole system of acceptances fits together very well, it is argued, is not enough to confer adequacy when no acceptance is supported independently. The system of acceptances might well be “isolated” from reality. Lehrer responds to this objection by noting that conformity to reality is a condition of knowledge, not of warrant.

Numerous other objections against coherentism have been raised. They are nicely summarized by BonJour in BonJour and Sosa, Epistemic Justification: Internalism vs. Externalism, Foundations vs. Virtues.


One strategy that has been attempted by some philosophers is to add to a foundationalist approach to support some coherentist elements. Roderick Chisholm was perhaps the first to do so in a systematic way. Another attempt has been made by Sosa. Susan Haack (Evidence and Inquiry, 1993) calls her own hybrid account of support “foundherentism.” We will not investigate such efforts here.

Non-Doxastic Accounts of Warrant

The above discussion assumes that what counts as evidence is beliefs or acceptances. If this view is rejected, a non-doxastic version of foundationalism can be developed. According to Pollock and Chisholm, what counts as basic evidence is experience. The support relation is not modeled on inference. Instead, it is defined by special rules that operate, so to speak, “in the background.” Pollock claims that we come to know in the manner of riding a bicycle, rather than looking something up in a manual.

This approach has the advantage of side-stepping the main anti-foundationalist argument. It is not necessary to interpret our experience in order for it to count as evidence. This would be needed only if evidential support requires “consulting” experience in the way that we consult a manual. Instead, support arises from the use of internalized epistemic norms.

The epistemic norms Pollock cites operate purely internally, sometimes upon beliefs, and sometimes on non-doxastic psychological states, and in particular, perceptual states. These states function as reasons, in a manner of speaking, in that they provide input for reasoning. But they are not reasons in the traditional sense as being able to serve as premises in inferences, which only the propositional contents of belief can do. The psychological states must be accessible in order for them to serve as reasons, and Pollock claims that we are in fact directly aware of them. Thus, Pollock is an internalist, so long as ‘logical’ in ‘the logical space of reasons’ is interpreted in Pollock’s extended sense in which reasons can function non-inferentially.

The doxastic theorist will object that the appeal to the alleged epistemic norms is not justified. We ought to be picking our norms because they are truth-conducive, or at least because they appear to be truth-conducive. In effect, the non-doxastic foundationalist justifies his choice of norms simply because they are the ones we use. Pollock and Chisholm do not deny this.

Externalist Accounts of Warrant

Our investigation of the internalist/externalist dispute over warrant now turns to the side of externalism. We can divide externalist conceptions of warrant into two types. The first type introduces a condition which connects belief and truth in a very abstract and indeterminate way, by “tracking” the relation between belief and truth. The second approach to warrant is to give a relatively concrete account of how belief and truth are connected.

Subjunctive Accounts of Warrant

One way to think of something happening by accident is to say that there is some unusual, unexpected, or unintended circumstance that brought about its happening. Suppose, for example, that I run into you the store by accident as I am entering and you are leaving. If I had left for the store five minutes later, I would not have seen you at all. On the other hand, if you were waiting (patiently) for me at the store, my having left five minutes later than I had planned would not affect the outcome.

We can apply this model to knowledge. One non-accidentally “gets it right” with respect to the truth that p when “getting it right” is not affected by small changes in one’s situation. We will suppose that in every case in which p is true, there is a range of alternative possible cases in which p is false, which are as close as can be to the actual case while still making p false. It is these cases which are used to evaluate subjunctive conditionals. Consider the case of my accidentally meeting you at the store. I can truly say that if I had left five minutes later, I would not have encountered you at the store. To evaluate this sentence, I would fix everything as being the same except for my time of leaving and whatever is required to accommodate my leaving at the later time. This limitation is essential, for I could imagine that you had left five minutes later as well, in which case I still would have met you at the store.

Reasoning about such situations is done using subjunctive conditionals. Grammatically, a subjunctive conditional is of the form, if it were the case that p, then it would be the case that q. Some such conditionals are contrary to fact or counterfactual, whose truth depends not on what is the case, but on what would be the case if things were other than what they are. The standard way of evaluating the truth of counterfactuals is to consider whether in all of the situations nearest the actual one in which p (which is false in the actual world) is true, q is true as well.

Now let us consider some conditions that use subjunctive conditionals to co-ordinate truth with belief in such a way that accidentally true belief is ruled out. The first such condition is a variant of Dretske’s Conclusive Reasons condition introduced above. Recall that it could be construed as stating that if p were not the case, then S’s reasons would not be the case either. We can transpose this from involving reasons to involving belief.

Sensitivity: If S knows in c that p, then if p were not the case, then S would not believe in c that p. (Robert Nozick, Philosophical Explanations, 1986)
Nozick’s account is sometimes known as the “tracking” account. It is externalist because whether one’s belief tracks the truth is not a fact that is internally accessible to an epistemic subject. (Indeed, it is even questionable whether tracking could be determined by an outside observer.)

The tracking conditions on knowledge discussed above can be thought of as requiring a kind of atemporal stability. It does not focus on whether a belief would continue to be held under various future circumstances, but rather on whether it would continue to be held under circumstances which are just like the present circumstances but which do not presently obtain. The emphasis is not on future information or action, but on variation in external circumstances.

To see how insensitivity can rule out knowledge, consider the case of hearsay evidence. I have heard third-hand that my friend is getting a divorce. If I am sensitive to the truth of the proposition that my friend is getting a divorce, then if my friend were not getting a divorce, I would not believe that he is getting a divorce. It is easy to see why this condition is not met. Take a case where my friend is not getting a divorce, but someone in the chain of gossip has misheard that he is getting divorced. This possibility would be close enough to the actual situation that it would be relevant. And since I would still believe that my friend is getting divorced even if my belief was ultimately based on a mistake in the chain of communication, I do not know (on the basis of the hearsay evidence) that my friend is getting a divorce.

A further (or alternative) condition one might require has been advocated by Ernest Sosa. His idea is that what is known should be “safe” in the sense that if S were to believe that p, p would be true.

Safety: If S knows in c that p, then if S were to believe in c that p, then p would be true. (Ernest Sosa, “Postscript” to “Proper Functionalism and Virtue Epistemology,” 1996)
S is safe in believing that p when, in the nearest conditions in which S does believe that p, p is true. S would not “get it wrong” in the situations most like the present one in which S believes that p. For example, a belief based on witnessing an event first-hand might be taken to be “safe” in this way.

It may occur to the observant reader that the sensitivity and safety conditions might be equivalent. The reason is that the two are contrapositives of each other. If the conditional in the sensitivity condition is reversed and the negation signs dropped, the safety condition results, and if the conditional in the safety condition is reversed and the negation signs added, the sensitivity condition is generated. But while the equivalence holds for indicative conditionals, it does not hold for subjunctive conditionals of the sort we are considering in the Sensitivity and Safety conditions.

The following is a case where safety apparently is satisfied but sensitivity is not. Suppose that S is a typical human being in ordinary circumstances, which include his being awake. Let p be the proposition: that S is now awake. We shall also suppose that this is not a belief that S has in c, perhaps because S is preoccupied with other matters. If S is to satisfy the safety condition with respect to p, if S were to believe that p, then p would be true. That is, if S were to believe that he is awake, then he would be awake. This seems right. In terms of the semantics, in the closest worlds where S does believe that p, S is awake. Such a world might be one in which S is reading Descartes’s First Meditation, which sets S to wonder whether he is asleep or awake. But it seems that S is not sensitive to p in S’s circumstances. That is, it is not the case that if S were not awake, S would not believe that S is awake. One very close world is a world in which S is asleep but believing he is awake because he is dreaming that he is, a kind of situation suggested by Descartes.

Sensitivity, then, seems to be in at least one way a stronger condition than is safety. Indeed, the present example seems to show that if sensitivity is required for knowledge, S does not know that he is awake. This is in fact the conclusion drawn by Keith DeRose. On his view (discussed in the previous module), if the sensitivity standard is operative, then we do not know that “skeptical hypotheses” such as that S is not awake are false.

A final condition in this family of “tracking” conditions was given by Nozick. As it is the converse of the safety condition, Sosa refers to it as “counter-safety.”

Counter-Safety: If S knows in c that p, then if p were true, S would believe that p. (Robert Nozick, Philosophical Explanations, 1986)
In all close worlds in which p is true, S believes that p. This condition does not get much discussion, given that the truth of and belief that p are already conditions of knowledge.

In fact, it can be argued that given the truth and belief condition, the Counter-Safety condition follows logically. If those conditions are met, then p is true and S does believe that p. Given the standard account of the semantics of subjunctive conditionals, this implies that if p were true, then S would believe that p. But in that case, the Counter-Safety condition is redundant. (Note that this argument applies to Safety as well, and its consequences will be discussed below.)

Still, Nozick had a reason for laying down the Counter-Safety condition and rejecting the standard semantics. This has to do with the fact that there are various propositions which (it is claimed by most) are necessarily true or cannot be false. Propositions of logic and mathematics are generally taken to be of this kind. Since in such a case p must be true, the Sensitivity condition could not be triggered. If it is impossible for p to be false, then, again by the standard semantics for subjunctive conditionals, if p were false, S would not believe that p. (And for good measure, if p were false, S would believe that p!)

However, S may lack knowledge in these cases. Perhaps S believes that p (a difficult mathematical theorem) on the basis of being told that p by an unreliable classmate. Now consider an alternative situation, not far from the truth, where his classmate mistakenly tells S that p is false. In that case, p would be true, but S would not believe that p. So S does not know that p, given Counter-Safety.

Objections to Subjunctive Accounts

While the requirement of truth-tracking has a certain intuitive appeal, the tracking conditions have generally been rejected by epistemologists. This seems to be due to some devastating counterexamples due to Saul Kripke. These appear in a paper which is was only recently published, but which was widely presented as a talk. What Kripke convincingly argued is that tracking is influenced by external conditions in a way that does not reflect what is essential to warrant. Other objections to the sensitivity condition have been given by Jonathan Vogel (“Tracking, Closure, and Inductive Knowledge,” 1987).

One of Kripke’s arguments may be summarized as follows. Consider Goldman’s case in which Henry is driving in the countryside and correctly identifies a barn, unaware that the area is littered with imitation barns. There are close possible worlds in which what Henry sees is not a barn, but in such worlds, he would still believe that he sees a barn, as he is unable to discriminate between the imitations and the real things. So, Henry does not meet the Sensitivity condition, and on Nozick's analysis, does not know that he sees a barn.

So far, so good, Kripke says. But now modify the case a bit, such that it is causally impossible, for some reason or other, that there be imitation red barns in the area. Then there are no close worlds in which Henry mis-identifies an imitation red barn as a real red barn. He sees a red barn and believes that he sees one. Given some obvious assumptions about Henry’s powers of discrimination, he meets the Sensitivity condition: if it is false that he sees a red barn, he would not believe that he sees a red barn. Since he does see a red barn and sensitively believes that he sees one, then the analysis forces the conclusion that he knows that he sees a red barn.

Kripke claims that this result is unacceptable. On the analysis, Henry knows that he sees a red barn, but he does not know that he sees a barn. This is counter-intuitive, and it presents us with a case of failure of Closure Under Detachment, since the proposition that one sees a red barn implies the proposition that one sees a barn. One might contest Closure principle on some grounds, but the present case does not seem to pose any problems for the principle. Kripke’s diagnosis of the problem is that the logical behavior of counterfactuals does not mirror that of epistemic principles such as Closure Under Detachment.

Perhaps the related principle of “safety” is more suitable as a condition for knowledge. As noted above, this is that belief be “safe,” in the sense that in all situations similar to the present one, if S believes that p, then p is true. This condition is called “Cartesian” by Sosa. For Descartes, if a person is absolutely careful in believing only what is evident to him, then if he believes that p under these conditions, p is true. Sosa’s condition is weaker than that of Descartes. According to the standard semantics of counterfactuals, the only situations in which p must be true given that S believes that p are those which fairly closely resemble the actual situation. In very remote situations, such as one in which I am the victim of radical deception, I might believe that p without p’s being true.

One logical feature of about the safety condition has been noted by James Tomberlin (“Skepticism, Tracking, and Warrant,” 2000). If S knows in c that p, then S in c believes that p, in which case by safety, p is true. Thus the safety condition in conjunction with the belief condition would render the truth condition redundant. If safety is taken to be equivalent to warrant, then it would follow that every warranted belief is true, which is infallibilism. So it seems that safety is at most a necessary condition for warrant.

Causal Accounts of Warrant

We now turn to attempts to specify more concretely which external factors are sufficient for knowledge. Perhaps the most straightforward externalist account of warrant was given in an early paper by Alvin Goldman, “A Causal Theory of Knowing” (1967). Goldman’s condition for warrant presupposes truth and belief, so he gives a single condition for knowing.

Cause: S knows in c that p if and only if the fact that p is causally connected in an “appropriate” way with S’s believing that p.
The appropriate causal processes that produce knowledge include perception, memory, other mental processes that make connections between the fact that p and the belief that p. We might think of appropriateness as capturing the way in which the truth of the belief of p is not accidental.

While this account of knowledge has great intuitive appeal, Goldman was quick to abandon it. The problem he found with it was that it lacks sufficient generality, in that the conditions for knowing cannot be satisfied with respect to those facts (such as those of mathematics) which apparently do not cause us to have the beliefs we do. Goldman was looking for a single standard to apply to all “knowledge.” Perhaps the causal condition would fare better if it were restricted to perceptual knowledge.

Reliabilist Accounts of Warrant

Goldman’s own response to what he thought was the inadequacy of his condition was to propose that warrant requires that belief be caused in a reliable way, a view known as “reliabilism.” Here is a generic form of reliabilism:

Reliability: If S knows in c that p, then S’s belief that p was produced by a reliable belief-forming process.
This is more general than the causal account because it does not require any direct causal relation between what is known and the production of the belief. The connection instead is indirect. The belief that p is of a kind that is (objectively) likely to be true when it is formed by the process that produced it. If we can reliably detect mathematical truths, for example, then they can be objects of knowledge.

There are a number of technical objections to generic reliabilism that have led to a number of refinements. One obvious problem is that the very notion of reliability is vague, and it is hard to see how to calibrate it to produce exactly the results we want.

Another problem is that the reliability of a process is relative to the conditions under which it operates. This is known as the “generality” problem, because the degree of reliability varies with the generality of the description of the process. For example, “perception” (in the “wide” sense which allows for perceptual error) is not very reliable, given that it takes place under a variety of conditions that can produce false belief: dreams, madness, the conditions which the Pyrrhonian skeptics described in such lush detail. The problem is how to find just the right specific kind of perception that is productive of knowledge.

One reason internalists reject reliabilism is because it produces knowledge attributions which they reject. The typical anti-reliabilist counterexample involves situations where a person is reliable in a way that cannot be internally detected. Someone might have a brain abnormality which makes that person highly reliable in the production of certain kinds of beliefs. The internalists want to deny that such a person has knowledge, though the reliabilists are willing to attribute knowledge in such cases.

Relevant Alternatives

As noted above, one way of understanding the claim that knowledge requires that the truth of a belief not be accidental is that one can rule out alternatives to the truth of the belief. The externalist version this approach is closely related to that of the reliability approach. In order to know, a subject must be in a position to rule out certain alternatives which are “relevant” in that person’s situation.

Relevant Alternatives: If S knows in c that p, then S in c is in a position to rule out all relevant alternatives in which p is false.
This is intended to be a fallibilist condition on knowledge, since the ability to rule out all alternatives in which p is false would be infallibility. The intuition that supports this condition is elicited by two famous cases.

Dretske’s case from “Epistemic Operators,” already discussed, involves a man at a zoo looking at a cage labeled “zebras.” He sees a zebra and comes to believe that he sees a zebra. He cannot, however, distinguish the situation in which he sees a zebra from one in which he sees a mule cleverly painted to resemble a zebra. It seems that for him to know, he must be able to rule out the alternative that he sees an elephant, but must he be able to rule out the implausible alternative in which something very zebra-like, but not a zebra, is before him?

A case from Alvin Goldman, described above in its application to the tracking condition, is quite similar. Henry is driving in the countryside and sees a barn. He is driving slowly enough to get a good look at it, etc. He seems to know that he sees a barn. But suppose the countryside is full of cleverly constructed barn-façades, which he could not distinguish from a real barn. Then he seems not to know that he sees a barn, because he cannot rule out the alternative that what is before him is a barn façade and not a barn. This alternative becomes relevant because of the presence of barn façades. So Goldman suggests a “discrimination” condition.

Discrimination: If S knows in c that p, then S can discriminate the actual state of affairs in which p is true from relevant alternative states of affairs in which p is false.
Being able to discriminate perceptually seems to be a specific case of being able to rule out alternatives. What Goldman’s case adds to Dretske’s is an illustration of a condition under which alternatives that are not ordinarily relevant become relevant.

Relevant alternatives conditions seem plausible enough as they stand, but there is some difficulty in fleshing them out. The problem is to specify exactly what counts as relevance. Relevance might be understood in terms of probability, but how improbable must an alternative be in order to be irrelevant? And as Goldman’s case shows, what is relevant may depend on the circumstances of S, and the factors in S’s circumstances that influence relevance might be hard to pin down. How many barn façades are needed to make their presence relevant? How nearby must they be? These problems are similar to those facing the reliability account. And indeed, one way to understand reliability itself is in terms of the ability to rule out relevant alternatives.

Virtue Accounts of Warrant

Some externalists are persuaded by the same kind of example, though for different reasons. They see the problem as being that a person with a brain abnormality produces true beliefs in a way that is, so to speak, accidental, which makes the truth of the belief accidental in a way that precludes knowledge. As Alvin Plantinga puts it, knowledge can result only when one’s cognitive faculties are functioning properly (see his 1993 book Warrant and Proper Function). Here is a paraphrase of his account:

Proper Function: If S knows in c that p, then S’s belief that p is produced by the proper functioning of S’s cognitive systems in the environment for which it is designed.

For Plantinga, the relevant design plan is the product of a supernatural being. However, it seems open to interpret the “design plan” naturalistically in evolutionary terms. This would be awkward, though, because the notion of a “design plan” is teleological (end-directed) in nature, whereas standard accounts of evolution are not framed teleologically.

One can avoid the teleology of the “design plan” in the manner of Ernest Sosa, who puts the matter in terms of a requirement that the belief-producing faculties be “virtuous” epistemically (see his 1991 anthology Knowledge in Perspective). The approaches of both Plantinga and Sosa have come to be called “virtue epistemology.” Its warrant-condition can be put this way:

Aptness: If S knows in c that p, then S’s belief that p is produced by virtuous cognitive faculties of S.

An important limitation of Sosa’s proposal is that it applies as a sufficent condition for warrant only to non-reflective or “animal” knowledge. Belief produced by virtuous faculties is described merely as “apt.” The production of “apt” belief does not confer what Sosa calls “justification,” which he describes in internalist terms. Justification is required for reflective knowledge, that is, knowledge that is a product of reflection upon one’s epistemic situation.

The chief internalist complaint about virtue epistemology is that the fact of virtue or proper function is not itself internally accessible, and hence that one is attributed knowledge without having access to what it is that makes it knowledge. This is in essence the reason why Sosa thinks that for “reflective” knowledge, one needs a “meta-perspective” or a view about the virtuousness of one’s faculties. The standard for the meta-perspective for Sosa turns out to be coherence. So we can say that Sosa is an externalist concerning animal knowledge and an internalist concerning reflective knowledge. The reason an internalist might demand a meta-perspective is that he thinks that without some kind of access to the fact of one’s virtue, one will be left wondering whether one actually knows or not. Another criticism of externalist accounts in general is that they omit an important dimension of knowledge, namely our accountability for what we believe—something which is easily explained by internalism.

At any rate, if some of the external factors (such as reliability) required for warrant are not internally accessible, it seems that we cannot always be held accountable for believing what is not warranted. So someone who links epistemic responsibility necessarily with knowledge will almost certainly reject externalism. And someone who does not will find this objection to be of no merit.

Interestingly, Alvin Goldman, one of the original externalists, came to acknowledge a deontological element in knowledge attribution. For this reason, he adds to reliability in his account of warrant a condition which he calls “weak justification” (See “Strong and Weak Justification,” 1988). The requirement is that for S to know that p, S must be “blameless” in believing that p.

The Gettier Problem

As we have seen, there is a great deal of disagreement about what constitutes warrant. The arguments for and against each position are disputed, with no resolution in sight. But even if there is agreement about what warrant is, there is a further problem that has bred further disagreement—what has come to be known as the “Gettier problem.”

We begin our discussion of the problem by noting a feature that the nearly all the accounts of warrant have in common: fallibilism. What one is warranted in believing does not determine the truth-value of what is believed. One consequence of this is that one can have a warranted belief that is false. Another is that one can have a warranted belief that is true, but true in a way that is disconnected from the way in which the belief is warranted.

The earliest examples of this phenomenon are found in writings of Meinong and Russell. Meinong (Über der Erfarungsgrundlagen unseres Wissen, 1906) imagined a situation in which S has an auditory hallucination of the sound of a harp. This (we will suppose) warrants S’s belief that a harp is playing. And in fact a harp is playing, only it is not heard by S—only the hallucination is. So S has true belief that is warranted to some extent. But does S have sufficient warrant that amounts to knowledge?

Russell’s case (Human Knowledge: Its Scope and Limits, 1948) was that of looking at a clock that has stopped, say at 12:00. S looks at the clock and forms a belief that it is now 12:00 which is warranted (we will suppose). And indeed, at that very moment it is 12:00. Does S know that it is now 12:00?

Edmund Gettier, in 1963, gave two examples in the context of an analysis of knowledge as “justified” true belief (see, “Is Knowledge Justified True Belief?”) The publication of these examples set off a blizzard of articles. Virtually all of them agreed that the cases Gettier presented were not cases of knowledge, and as a result they set out to patch the traditional analysis so as to make them safe from “Gettier cases.” To date, the creativity of the crafters of Gettier cases has outrun that of those who would amend the analysis of knowledge.

One response to the cases is to try to tighten up the account of warrant, so that the subjects in the cases are not in fact warranted in believing what they do. So perhaps looking at a malfunctioning clock or seeming to hear a sound should not count as warrant. But it seems that no matter how much the analysis of warrant is restricted, the underdetermination of truth by warrant, which is due to the adoption of fallibilism, will allow the construction of a Gettier case.

Timothy Williamson raises a problem of circularity for this kind of approach.

If someone insists that knowledge is justified true belief on an understanding of ‘justified’ strong enough to exclude Gettier cases but weak enough to include everyday empirical knowledge, the problem is likely to be that no standard of justification is supplied independent of knowledge itself. (Knowledge and its Limits, p. 4)
The proposed solution to the problem of tightening up the standards for warrant in a fallibilist framework, according to Williamson, yields a circular analysis of knowledge. The account of warrant (justification) would be derived from an understanding of what knowledge is, rather than providing an understanding of what knowledge is.

Externalist Solutions to Gettier Problem

Goldman’s causal account of warrant was introduced precisely to preclude Gettier cases. The causal account actually looks promising, because it appears that in cases of knowledge sanctioned by the causal account, truth is not underdetermined by warrant. For S to know that p, it must be the very fact that p which causes in an “appropriate” way my belief that p. So in the clock case, it must be the fact that it is now 12:00 that causes me to believe in an appropriate way that it is 12:00.

One might object to the causal account by noting that the fact that it is 12:00 could set off some Rube Goldberg-like chain of events that causes me to believe that it is 12:00, without my knowing it. But the causal theorist could reply that that chain is not “appropriate,” and that in general any Getterizing monkey-business will not be appropriate. We will not discuss this initially promising solution further, because virtually nobody subscribes to the causal theory any more.

Note that the other externalist accounts of warrant are not so successful as the causal account. A reliable process of belief-formation need not be an infallible process. What we are warranted in believing, then, does not determine the truth of what is believed. It seems that the only way the reliabilist can escape the possibility of Gettier cases would be to define reliability in such a way that one’s belief is not reliably formed if it can be true for a reason not related to its warrant.

For example, suppose I cannot distinguish two identical twins from each other. The one with whom I am acquainted lives in the United States; the other, of whose existence I am unaware, lives in Australia. The Australian twin comes to the United States and conspires with her sister to fool me. While the American twin hides in the room, the Australian twin presents herself to me. I form the belief that the American twin is in the room, and that belief is true.

Are we to say that my belief that the American twin is in the room is not warranted because I cannot distinguish her from her sister? That would solve the problem, but it seems to have steep consequences. For it seems that my beliefs about my friend are warranted when her sister is in Australia. The generality problem arises here, for it does not seem that there is a principled way to establish the circumstances in which one is reliable from those in which one is not.

Internalist Solutions to the Gettier Problem

The internalist seems to face even greater problems than does the externalist in solving the Gettier problem. For what counts as warrant is an internal matter. There is no way to say that the internal support one has for a belief in a Gettier case is not appropriate, due to the external circumstances of the case.

The standard internalist response is to add a fourth condition of knowledge, designed to rule out knowledge in Gettier cases, while leaving internal warrant intact. The most common approach is based on the notion of a defeater.

In the case of the harp, the fact that S is having an auditory hallucination defeats one’s knowledge. In the clock case, the fact that the clock has stopped serves as the defeater. In the twins case, the fact that S cannot distinguish one twin from another is the defeater. So in general, a defeater is a piece of information that explains (at least to some extent) why it is the case that p, where it is clear that this explanation does not conform to what makes the person warranted in believing that p.

So we can set down a fourth condition of knowledge:

Undefeated Warrant: If S knows in c that p, then there is no true q which would defeat S’s warrant in believing that p.
This condition needs to be fleshed out in order to be of any value. Specifically, we must be given conditions for some proposition being a defeater. One suggestion is that a defeater is a proposition which, if believed or accepted by S, would lead to S’s not being warranted in believing that p.

One problem with this simple “no defeater” approach is that it seems to be too strong. It may be that there is a defeater q which does not block knowledge of p. It may be a misleading piece of evidence. The problem then becomes one of trying to screen out the non-damaging defeaters while not excluding the defeaters that actually block one from having knowledge. At this point, the analyses become so elaborate that it is difficult to tell even how they work, let alone whether they succeed in blocking all the possible Gettier cases.

A variant of the “no defeater” internalist approach is an outgrowth of the full version David Lewis’s Uneliminated Possibilities condition for knowledge, described above.

Qualified Uneliminated Possibilities: If S knows in c that p in c, then p holds in every possibility left uneliminated by S’s evidence (except those possibilities which are properly ignored).
Lewis lays down a set of circumstances that do not allow a possibility to be properly ignored, including what is actually the case and what S believes to be the case. He adds to this a rule of resemblance, that what saliently resembles what may not be properly ignored may not itself be properly ignored. Lewis claims that this approach “solves the Gettier problem” (“Elusive Knowledge”).

Take, for example, the barn example described above. Henry’s circumstances include the fact that he is in an area full of barn façades, which he cannot distinguish from real barns, yet he correctly identifies a barn. On Lewis’s view, Henry cannot properly ignore actuality, and thus he cannot properly ignore the presence of the fake barns. Further, a possibility which saliently resembles actuality is one in which Henry mistakenly takes a fake barn for a barn. Henry cannot eliminate this possibility, which in effect defeats his justification, in which case he lacks knowledge. Other cases are treated similarly. (In effect, what we have here is the relevant alternatives approach with a clear account of what makes an alternative relevant, though one using a vague notion of salience.) This interesting approach will not be pursued further here.

Available Evidence

Some epistemologists have thought it necessary to add another condition for knowledge beyond what is needed to block Gettier cases. The alleged need for such a condition stems from a case due to Gilbert Harman (Thought, 1973).

Suppose that a person S has read in an early edition of a newspaper the report of the assassination of a prominent civil-rights leader. S has every reason to trust the source of the report and indeed does trust it, forming the belief that the event had occurred as described. Suppose further that the event had in fact occurred and had been described adequately by the paper. One might be inclined to attribute knowledge to S.

Suppose further that soon after the publication of the early edition of the paper, a new edition was published, this one retracting the original account and declaring that the civil-rights leader was alive. The reason for the deceit was to prevent race-riots from breaking out when word of the (real) death became generally circulated.

After the retractions have been issued, S goes about his business, oblivious to them and as well to what people around him are saying—that the leader is alive. Harman held that knowledge should not be attributed to S in this case, because of S’s ignorance of relevant “social” information, misleading though it was. His reaction was to require as a condition of knowing that p that S not overlook relevant beliefs of people around S.

Some epistemologists, such as Keith Lehrer, think that the presence of misleading evidence of any kind should not block one’s knowing. Harman argues that it can block one’s knowing when it is readily available. We can formulate a principle embodying this claim:

Available Evidence: If S knows in c that p, then S in c has taken into account all relevant evidence regarding p that is readily available to S.
The addition of a condition of this kind might be defended by appeal to intuition. But those people who think misleading evidence should not block knowledge do not have this intuition. The condition might be justified more systematically by claiming that S would not be justified without taking all relevant evidence into account because S would be epistemically irresponsible in believing that p.

A potential problem with this kind of condition is that it will rule out many attributions of knowledge just because some small bit of readily available relevant evidence has been overlooked. It is hard to see how one would modify the proposed condition in order to prevent this from happening. One would have to specify somehow that only significant evidence of this sort has to be taken into account, but the concept of significance is inherently vague.

Prospects for Analyses of Knowledge

We have surveyed most of the main attempts to provide necessary and sufficient conditions for S’s knowing that p. There is pretty widespread agreement that the truth of p and S’s commitment to the truth of p are necessary conditions for S’s knowing that p. Disagreement begins when a further condition of warranted belief begins to be fleshed out.

Many ingenious attempts to explain warrant have been made by epistemologists over the last fifty years or so. There have been proposals and retractions, crude sketches and complex refinements. But there seems to be little agreement about what warrant is. Now it may be that someone, some day, will come up with an account which will satisfy almost everyone. (It is impossible for me to imagine an account that would satisfy absolutely everyone, barring a drastic reduction in the size of the educated populace.)

It seems more reasonable to think that the source of much of the disagreement lies in variety of ways in which we understand knowledge, especially given the general adoption of fallibilism. What seems in store for us in the future is the production of more sophisticated versions of established views, plus a few new approaches that will appeal to some as-yet untapped ways in which we have always been thinking about knowledge.

[ Previous Module | Next Module | Table of Contents | Assignment Page | Course Web Page ]

[ G. J. Mattey’s Home Page | UC Davis Philosophy Department Home Page | UC Davis Home Page ]

This page and all subordinate pages copyright G. J. Mattey, 2004, 2005, 2008, 2009, 2010, 2011, 2012, 2014. All rights reserved.