Contemporary Epistemology II: Language

UC Davis Philosophy 102

Theory of Knowledge

Instructor: G. J. Mattey, Senior Lecturer

Version 1.4.3, April 27, 2015

The linguistic project concerns the way in which the words ‘know,’ ‘knows,’ ‘knowledge,’ etc. are used in ordinary language. In carrying out the project, it is customary to focus on knowledge attributions of the form ‘S knows that p’ and ignorance attributions of the form ‘S does not know that p.’ (We shall use the generic expression ‘epistemic attribution’ to include attributions of knowledge and attributions of ignorance.) One might ask several questions about epistemic attributions. Why are they made? When are they made correctly, and when incorrectly? What is the meaning of the terms ‘knows’ and ‘does not know’ when they are embedded in attributions of knowledge and ignorance? We will look at proposed answers to these questions in this module.

Syntax, Semantics and Pragmatics

Generally, linguistic analysis is divided into three main categories: syntax, semantics, and pragmatics. Syntax deals with the rules by which sentences are generated. In the standard treatment, the words ‘know’ or ‘knows’ are verbs which form verb-phrases when completed with clauses that begin with ‘who,’ ‘why,’ ‘when,’ ‘what,’ ‘how,’ ‘whether,’ ‘that,’ and others. Joining the verb-phrase form ‘knows that p’ (the predicate of the sentence) to the noun-phrase ‘S’ (the subject of the sentence) produces a sentence of the form ‘S knows that p.’ The verb-phrase itself consists of the verb ‘knows’ followed by a complement ‘that p.’ The complement contains the complementizer ‘that’ and the sentence ‘p,’ which itself can be broken down in terms of noun-phrase, verb-phrase, etc. We will not have anything more to say about the syntax of sentences using the verb ‘knows’ in what follows.

Semantical analysis is concerned with such matters as meaning and truth. One semantical issue is what the predicate ‘knows that p’ means. Given that the predicate is meaningful, it would appear that sentences of the form ‘S knows that p’ which attribute it to an epistemic subject have truth-values. Then the question becomes what determines the truth-value that a given such sentence has (or what are the sentence’s “truth-conditions”).

It is generally held that meaningful sentences express propositions, and that it is the truth or falsehood of the proposition expressed by a sentence that determines whether the sentence itself is true or false. Thus we distinguish between a sentence of the form ‘S knows that p’ and a proposition of the form that S knows that p. The question of the truth-value of the sentence is then shifted onto the question of the truth-value of the proposition it expresses.

Sentences of the form ‘S knows that p’ are semantically underdetermined. First, the proposition expressed by that p can depend on the context of the use of the sentence. If I say, ‘Barack Obama knows that the Democrats lost control of the House of Representatives,’ I have in mind the election held in 2010, but the sentence could be about the election held in 2008. So to apply semantics to the sentences requires precise specification of the meaning of the sentence p.

In non-epistemic use of language, there often occur in sentences indexical pronouns such as ‘I’ and ‘now.’ Sentences using ‘I’ that you utter will express propositions different from those that I utter, even if the syntactical structure of the sentence is the same. (For example, take ‘I am sitting.’) Another kind of ambiguity involves vague terms such as ‘tall.’ The sentence ‘Steve Nash is tall’ might be taken to express one proposition when uttered in the company of ordinary people and another when uttered in the company of a professional basketball team. (He is officially listed as 6'1''.) In both cases, the different propositions expressed may have different truth-values, in which case the sentences expressing them might have different truth-values as well.

Aside from ambiguities in sentences occurring after the complementizer of epistemic attributions, the circumstances of the epistemic subject (who is also the subject of the sentence) lead to ambiguities. If I say that Barack Obama knew that the Democrats would lose control of the House of Representatives, the time at which knowledge is attributed to him makes a difference as to whether my attribution is true or false. So we will consider only knowledge attributions which are precise enough to include (at least implicitly) a reference to the circumstances of the epistemic subject.

Finally, sometimes it is possible for a single sentence to express different propositions, depending on the context in which it is used. It may be that even when fully specified, the sentence ‘S knows that p’ is ambiguous due to the context of its use. In what follows, we will be examining some apparent cases of this kind. Some epistemologists have claimed that sentences expressing epistemic attributions contain ambiguities similar to those found with indexicals or vague terms.

Let us now turn to the question of the conditions under which sentences of the form ‘S knows that p’ are true. Consider the standard view, on which the sentence expresses a proposition of the form that S knows in circumstances c that p. The proposition, in turn, can be understood as stating that there is a relation of knowing holding between the epistemic subject S, a circumstance c, and a proposition p. Each relation is said to have an extension, which is an array of objects that are related to one another by the relation. In the case of knowledge, we distinguish three relata: the epistemic subject S, the circumstances c of S, and the proposition p. So one semantical question on this account is which ordered triples of objects <S,c,p> are in the extension of ‘knows.’ If the triple <S,c,p> is in the extension of ‘knows,’ then the proposition that S knows in c that p is true. Otherwise, it is false.

For example, suppose the sentence ‘Barack Obama knows that the Democrats have lost control of the House of Representatives’ is asserted on the morning of November 3, 2010. Then the proposition expressed by the sentence might be put this way: that Barack Obama knows in the circumstances in which he finds himself on November 3, 2010, that the Democrats have lost control of the House of Representatives. If this proposition is true, then the triple <Barack Obama, being in the circumstances in which Barack Obama finds himself on November 3, 2010, that the Democrats have lost the House of Representatives in 2010> is in the extension of ‘knows.’ If it is false, then that triple is not in the extension of ‘knows.’

As we will see, some epistemologists hold that there is more than one relation (or set of ordered triples) expressed by the word ‘knows.’ In that case, each of the relations will have its own extension, and sentences containing ‘knows’ will express different propositions, depending on which relation is indicated by the proposition that S knows in c that p. If that is the case, then it is possible for a triple to be in the extension of one of the relations and not in the extension of the other, in which case the sentence will have two different truth-values, one for each of the knowledge relations it expresses.

Pragmatics is the part of linguistics dealing with the conditions of the use of language. Some epistemologists are interested in trying to determine rules for the correct usage of sentences of the form ‘S knows that p.’ Of these, some hold that there are conditions dictating the correct use of the expression in a given context that are independent of the conditions which make the sentence true. Others claim that the truth-conditions of knowledge and ignorance attributions are themselves dependent on the contexts in which they are made or on the circumstances of the attributee. Some would take the matter further and say that variations in context of attribution affect the propositions and knowledge-relations expressed by sentences of the form ‘S knows that p.’

Variability in Epistemic Attribution

We shall begin with pragmatics and examine the conditions for the use of sentences of the form ‘S knows that p.’ In the process, we will examine the consequences of the pragmatic treatment for the semantics of such sentences. The most striking phenomenon in the use of sentences of this kind is the variability of conditions in which it or its negation are asserted. There are cases in which a knowledge-attributor K (in K’s context of attribution) and an ignorance-attributor N (in N’s context of attribution) differ in their attributions to S with respect to p without differing in their assessment of either the identity of S, the S’s circumstances, what the proposition p is, whether p is true, whether S believes that p, or S’s epistemic position. Most importantly, they may agree that S’s epistemic position with respect to p has a certain degree of strength, yet they may not agree about whether S knows that p. We will take this to be a fundamental datum in linguistic practice and will seek to explain this datum.

Variability in Epistemic Attribution: Without difference in their beliefs about the truth of p, whether S believes that p, and the strength of S’s epistemic position with respect to p, epistemic attributors disagree over whether to attribute knowledge or ignorance that p to S.

As evidence for the existence in variability in epistemic attribution, we will modify an example due to Stewart Cohen. Consider the sentence ‘Parker knows that his car is where he parked it an hour ago.’ K is willing to assert it, while N stands ready to deny it. K might reason as follows: Parker remembers having parked his car there; it is where he has always parked the car; the parking location is relatively secure; the time between his having parked it there and now is relatively short, etc. N might acknowledge all these facts about the strength of Parker’s epistemic position, yet refuse to attribute knowledge to Parker, on the grounds that Parker is not in a position to rule out the possibility that the car has been stolen within the last hour.

One reason for the disagreement seems to lie in the fact that the mere strength of S’s epistemic position with respect to p is not enough to determine whether knowledge or ignorance of p is attributed to S. Perhaps some other features of S’s circumstances play a role. A further factor is whether S’s epistemic position is “strong enough” (or whether S is “sufficiently warranted” in believing that p). Whether S’s position is strong enough depends on what standard must be met by S’s epistemic position in order for S to be said to know that p. Thus, if K and N are judging relative to different standards, and S meets the standards used by K and not by N, then the attributions of both K and N will be different. It is usually said that S is said to know on K’s standards because they are low relative to N’s standards, while S is said not to know on N’s standards because they are high relative to K’s standards.

One might be inclined to treat examples such as these as curiosities of linguistic usage and let them go at that. One way to try to understand them systematically would be to claim that they tell us something about the way in which ‘S knows that p’ is used, but that the manner of usage has no semantic significance. Such a purely pragmatic approach might consider general patterns of usage, and even pronounce that some usages are correct and some incorrect linguistic practice. (For example, they may be appropriate or inappropriate in various conversational contexts.) But it would draw no semantical conclusions from the results it uncovers.

Most epistemologists think that there are semantic consequences of the way we use sentences of the form ‘S knows that p.’ Specifically, they think that there is correlation between the correct use of such sentences and the sentences being true. Such a thesis would be that K correctly asserts ‘S knows that p’ only if ‘S knows that p’ is true, and N correctly asserts ‘S does not know that p’ only if ‘S knows that p’ is false. We might call this a semantical criterion of the correctness of usage. Such a criterion seems desirable, but at the same time it seems to be challenged by the phenomenon of variability of attribution.

The semantical approach to the correctness of epistemic attributions allows for four possible semantical values of the sentence asserted by K and denied by N. Two are of the same kind: either K asserts a true sentence and N a false one, or N asserts a true sentence and K a false one, but it is not the case that both K and N assert a true sentence. On this view, one of K or N is in error about what is attributed. Such a claim then requires an “error theory“ to explain why either K or S is in error in making his attribution. A third possibility is that (at least in some contexts), both K and N may make true attributions, and neither is in error. One explanation for this outcome would be that the same attribution may be both true and false, in which case neither K nor N is in error. A fourth possibility is that both attributions are false. It could alternatively be held that the attributions have no truth-value at all, perhaps because they do not express propositions but rather perform some other linguistic function.

The second and third approaches share the view that conditions of the use of ‘S knows that p’ dictate the standards to be used in attribution, and they share the further view that variability in context of attribution allows for variation in the semantics of sentences of the form ‘S knows that p.’ Beyond these rather general claims, there is a great deal of disagreement as to how the standards used in attribution are affected by context and as to the specific semantic consequences of any such variability in standards.

Why We Make Knowledge Attributions

With this framework in place, let us now take a look at the pragmatic question of why epistemic attributions are made in the first place. There seem to be two somewhat different reasons: one theoretical and one practical. On the theoretical side, our attributions are meant to indicate the quality of the subject’s information about the world. On the practical side, attributions of knowledge and ignorance are tied directly to action.

On reflection, it seems that not many of our attributions of knowledge and ignorance are purely theoretical. To be sure, we might wish in some contexts to make purely disinterested evaluations of the epistemic positions of epistemic subjects, perhaps out of curiousity or for purposes of documentation. For example, I might simply wonder whether Franklin Roosevelt knew that the Japanese were planning to attack Pearl Harbor. A somewhat artificial class of such attributions are those made in the course of the practice of epistemology itself. We often consider hypothetical subjects (like Parker in the example above) and ask clinically whether that subject meets some standard of knowledge we might have in mind.

The variability in our theoretical attributions of knowledge and ignorance may be accounted for in different ways. One might begin with a set of standards but subsequently run across different standards, higher or lower, which seem better than the standards that were originally in use. This frequently happens as a person becomes more familiar with norms of good reasoning and raises his standards accordingly. (Perhaps students who have successfully completed a course in critical reasoning emerge with higher standards. Descartes in his Meditations was consciously seeking a higher standard than that to which he had subscribed in his youth.) Another way one’s standards might vary is when the presentation of unconsidered possibilities bring to mind the relative weakness of the standards one currently uses. In the Parker case, if one takes into account the possibility of the car’s having just been stolen, one might think that this possibility should be ruled out, even if ruling it out was not taken into consideration in the original attribution.

The second way of thinking about knowledge and ignorance attributions concerns actions that are relevant to the attribution. We tend to think of knowledge as a proper basis for action. Frequently we explain unsuccessful actions by pointing to their having occurred as a consequence of the agent’s ignorance. The expression, “If I had only known that so-and-so, I would never have done that” is unfortunately heard all too often. Thus when we make an attribution of ignorance that p to ourselves and others, it may be that we are signaling that (other things being equal) the ignorant person ought not to act, or to have acted, on the assumption that p is true. Conversely, when we attribute knowledge that p to a subject S, we often indicate that it is in some way permissible for S to assume the truth of p as a basis for going ahead with a given course of action. (A major international business-news agency has adopted the slogan “Knowledge to Act.”) This normative language of permissibility is a reminder of the close parallel between epistemic assessments and moral assessments, discussed in the introductory module.

Risk in Knowledge Attribution

Given that we do in fact attribute knowledge and ignorance based on whether or not it is permissible for S to assume the truth of p as the basis for action, the question of the normative standard of correctness arises. Whether it is correct to attribute knowledge when knowledge is relevant to action seems to depend on several factors. One factor is the attitude toward risk held by the attributor himself. It seems that there is no single threshold of acceptable risk that applies to all people. It is a fact that different people have different tolerance for risk. In the case of Parker, K might think that it is not too risky for Parker to act on the assumption that his car is where he parked it, while N might think that he ought not to act on this assumption because N takes the risks of Parker’s being wrong to be too high by his own standards.

The differences in attribution may involve something other than (or in addition to) different thresholds of risk. Two attributors might have the same degree of risk-tolerance but have different views of the consequences of the action. If one attributor thinks that the consequences of being wrong are more harmful than does the other attributor, the latter might invoke a higher standard of knowledge than the former.

A third factor concerns the individual(s) who would suffer the consequences of S’s being wrong about the truth of p. When the attributor takes into account the consequences of the subject’s being wrong, or of being right, this may take place in several ways. The consequences for the attributor, for the subject, for both, for others, etc. may all be salient for the attribution.

Consider an adaptation of another example due to Stewart Cohen. Anxious and Carefree are at the airport. They meet a person, Casual, who has looked up the departure time of one flight on his smart phone and says that it is 12:00. If it does not matter to Carefree who might be on that flight, Carefree might attribute knowledge of the departure time to Casual. If it is important to Anxious that the time of departure be right, because it is Anxious’s connecting flight, Anxious might attribute ignorance to Casual. Anxious might do the same if the flight is Casual’s own connecting flight. It seems that the relevant factor is the person for whom being wrong would be harmful. If we combine the variable of who is at risk if S is wrong with those of the attributor’s conception of the consequences of being wrong or right, and the attributor’s tolerance for risk, we might begin to wonder whether a single rule can cover all the combinations.

Culpability in Knowledge Attribution

If someone’s action turns out badly because the person possesses bad information, we might merely feel sorry for him. On the other hand, there are times in which a person might be blamed for assuming the truth of p. In particular, if the person has sincerely asserted that he knows that p, when p is not the case, then the person has assessed own his epistemic position and found (at least) that the assumption of the truth of p is safe for present purposes. Now suppose that I act on your assertion that you know that p, where p turns out to be false. I might then blame you as being to some extent responsible for the failure of my action. (Conversely, I might praise you as contributing to the success of my action if it turns out the p is the case.)

The moral responsibility of the person attributing knowledge in a practical context is explained in an interesting way by J. L. Austin (in “Other Minds”). On his view, the act of self-attribution of knowledge is a special kind of action that we now call a “speech act,” in which one gives one’s word that p is indeed the case. In this way, one is assuring the hearer of the truth that p, and this act is analogous to a promise, in which one gives one’s word that one will act in a specified way. Any time someone (sincerely, not under coercion, etc.) gives one’s word, one incurs a moral obligation. In the epistemic case, it is the obligation to be right. If one fails in this obligation, then one may correctly be censured for having done so. For example, suppose that Carefree alone has heard what Casual said and tells Anxious that he knows that the plane will leave at such-and-such time. If Anxious misses the plane, then Carefree is in trouble.

Austin’s approach can be generalized to some extent. Austin claims only that I give my word that p is true when I assert that I know that p. But it seems plausible to extend this description to cases in which I attribute knowledge that p to someone else. When I do this, I am saying that S meets standards of knowledge that I endorse. If I make that attribution in your presence, and you act on the basis of my attribution, I am rightly blamed if I turn out to be wrong. So Carefree’s telling Anxious that Casual knows the time the plane will leave is enough to make Carefree blameworthy if Casual is wrong.

This “deontological” approach to knowledge attribution does not have any direct semantical consequences. That is, it tells us nothing about the truth-conditions for sentences of the form ’s knows that p.’ One extreme view suggested by Austin’s approach is that knowledge attributions have no truth-value at all, but rather function solely to give assurance. We will shortly turn to views according to which pragmatic considerations are said to have consequences for the truth-values of sentences. Before that, we will leave pragmatics for a moment and look at a purely semantical approach to the truth-conditions of sentences of the form ‘S knows that p.’


According to what has come to be known as “invariantism,” the truth-values of sentences involving the use of ‘knows’ are determined by a single, invariant, standard of knowledge. Thus, if there is disagreement between two users of ‘knows,’ one of the two would assert a false sentence. Classically, the invariant standard has been the strength of a subject’s epistemic position. The strongest such position is infallibility. On the standard of infallibility, ‘S knows that p’ is true only if S’s epistemic position is so strong that S could not be wrong in believing that p.

If the standard of infallibility is rejected in favor of something less demanding, it becomes questionable whether there is a single epistemic standard that determines the truth of sentences of the form ‘S knows that p.’ As we have seen, it seems that real-life attributions of knowledge are made according to shifting standards. We might dismiss this as erroneous use of language, but then it becomes mysterious as to how competent speakers of language could be so wrong so much of the time in their use of ‘s knows that p.’ Timothy Williamson has defended invariantism on the grounds that every rival theory must account for error in the use of ‘S knows that p,’ and that he can give a better account by appealing to invariant standards (“Contextualism, Subject-Sensitive Invariantism and Knowledge of Knowledge,” 2005). We will not develop Williamson’s approach here.

One way to try to save invariantism and yet account for the variability in knowledge attributions is to adopt what John Hawthorne (Knowledge and Lotteries) calls “moderate sensitive invariantism.” (In most of the subsequent literature, this position has come to be called “subject-sensitive invariantism.”) According to this approach, there is a single, invariant, standard for knowledge attribution. However, the standard is more complex than one that is sensitive (so to speak) only to the subject’s epistemic position. The standard must be sensitive to other factors regarding S and S’s circumstances. (Hence the name “subject-sensitive” invariantism.)

More formally, the subject and circumstance parameters contain information that, in addition to the subject’s epistemic position, plays a role in determining whether a sentence of the form Ks,cp is true. To put it another way, they help determine whether the ordered triple <S,c,p> is in the extension of ‘knows.’ The question for the attributor is not whether S’s epistemic position is strong enough in some absolute way, but whether S’s epistemic position is strong enough relative to the other factors involving S and S’s circumstances.

For example, the practical interests of the subject in the truth or falsehood of p may be something to which the standard is sensitive. In the airport case, Carefree is not so interested in being right about the time of the flight as is Anxious. Thus, the effect on the two of Causal’s testimony will be different. The standard might give the result that Carefree is attributed knowledge but Anxious is not. Although they are both in the same epistemic position, that position is strong enough for Carefree but not strong enough for Anxious, given their circumstances.

There is an important consequence of subject-sensitive invariantism that must be borne in mind from the start. There will be cases where a subject’s epistemic position remains fixed while his practical interests change. Suppose that according to some fixed standard and Casual’s circumstances at time t1, Carefree attributes knowledge that the plane will leave at such-and-such time. Then at t2, Casual receives a phone call which makes it urgent for him to make the flight on time. At that point, Carefree switches his view and attributes ignorance of the plane’s departure time to Casual. As Hawthorne puts it, the mechanisms which govern knowledge attribution on this view “are mechanisms for making knowledge come and go” (Knowledge and Lotteries, p. 158, note 3).

Keith DeRose has proposed some elaborate counter-examples to this approach (“The Ordinary Language Basis for Contextualism and the New Invariantism,” 2005). The examples have to do with an incongruity between attributions of knowledge to one’s self and to others. Suppose Greg and I have the same epistemic position with respect to p, but that Greg’s position is strong enough in his circumstances but my position is not strong enough in my circumstances. Then it would seem that I am forced to attribute knowledge that p to Greg but ignorance to myself. To some, this is an unacceptable consequence of subject-sensitive invariantism.


We can use the generic term ‘variantism’ to signify any denial of invariantism. Variantism with respect to the use of ‘knows’ is commonly called “contextualism,” which is the claim that the standards governing assertions of sentences that contain the word ‘knows’ vary with the context of the individual making the assertion. (There may be other possible forms of variantism, but none will be discussed here.) We must note at this point that there are two distinctive linguistic practices involving such sentences. The most standard practice is that of epistemic attribution, that is, attributing knowledge or ignorance that p to an epistemic subject. But there is a semantical practice as well, which is that of assessing whether the attributions are true or false. Focus on one or the other generates distinct varieties of variantism, as will be seen below.

There are cases outside epistemology in which different standards for the use of the same word are employed at different times, generating sentences with different conditions for use. For example, there are different standards governing the use of indexical terms such as ‘I,’ ‘here,’ and ‘now.’ The sentence “It is now noon“ is appropriately used at noon but not at any other time. Variations in standards governing usage can be found in sentences containing vague terms such as ‘tall.’ The tree in my neighbor’s yard may correctly be called “tall” relative to the other trees in the neighborhood but wrongly be called “tall” relative to the giant Sequoias. Finally, sentences with tensed verbs such as ‘is’ vary in the appropriateness of their usage depending on when they are used. “Socrates is sitting” would not be correctly asserted at the present time but could have been used correctly during the life of Socrates. All of these kinds of variations in standards have been used as models for contextualist accounts of epistemic attribution.

Contextualism takes it that what prompts shifts from higher to lower standards (or from lower to higher) with respect to strength of epistemic position is determined by something about the context in which the attribution of knowledge or ignorance, or the semantic evaluation of such attributions, is made. A single attributor K might attribute knowledge of p to S in one context of attribution and attribute ignorance to S in another context of attribution. Or attributor K might (in K’s context of attribution) attribute knowledge to S, while attributor N attributes (in N’s context of attribution) ignorance to the same subject S at the same time. Yet again, an attributor K might be willing to assert or endorse the truth of the sentence ’S knows that p’ in one context while N rejects it in another.

Thus, we can distinguish between accounts according to which it is the context of the use of ‘knows’ in epistemic attribution which accounts for the shifts in standards and accounts according to which the relevant factor accounting for the shifts in standards is the context of the assessment of the truth of knowledge-attributions. Accordingly, we will distinguish between “use-contextualism” and “assessment-contextualism.” Further, within the camp of use-contextualists there is a division between those who would model the pragmatics of ‘knows’ after the pragmatics of indexical terms, and those who would model the semantics of ‘knows’ after pragmatics for sentences whose truth-values are not determined in the manner of indexicals. They may be treated as containing vague terms or as sentences whose meaning depends in part on a time-factor. Thus, we will distinguish between “indexical-contextualism” and two varieties of “non-indexical-contextualism.” Finally, it might be thought that there is no approach that is able to account for the datum of variability in attribution, what MacFarlane calls “eliminitivism.” The options are summarized in the following chart.

Historically, the path to contextualism began with work in the 1970s by Peter Unger, who recognized that invariantism with very high standards leads to skepticism. (For a brief history of the development of contextualism, see DeRose, The Case for Contextualism (2009), Introduction, Section 12.) Contextualism was a response to this “great rival” (as DeRose put it), and it was in part developed by Unger, who eventually endorsed contextualism himself.

One kind of contextualist response to invariantism is an account of the pragmatics of conversation by David Lewis in 1979 (“Scorekeeping in a Language Game”). A second is a development of the “relevant alternatives” response to skepticism laid out by Fred Dretske in 1970 (“Epistemic Operators”). The two come together in a 1988 paper by Stewart Cohen, “How to be a Fallibilist.” Since the publication of that paper, Cohen has defended his contextualist thesis in a number of papers. Lewis in 1996 published “Elusive Knowledge,” which melds his earlier approach with a version of the relevant alternatives condition on knowledge. Another prominent defender of contextualism has been DeRose. Recently, a number of philosophers, and especially philosophers of language, have entered the debate, leading to the formulation of the other varieties of contextualism.

Relevant Alternatives

We will begin our discussion of contextualism with Cohen’s version, which is a natural response to a basic question posed by fallibilism. According to a fallibilist, there is at least one standard of knowledge which falls short of absolute infallibility or certainty. Fallibilism accommodates our ordinary use of epistemic terms: we are willing to attribute knowledge to subjects whose epistemic positions fall far short of the ideal position of being incapable of error with respect to a potential object of knowledge.

But once we give up an absolute standard, we are faced with the fact that, at least in ordinary usage, the standards used by different people vary enormously in their demands. Moreover, it is a fact that even the standards used by a single person vary a great deal. When standards are low, knowledge is attributed quite liberally, but when they are raised, we become skeptical. One way to calibrate the level of the standard in play is in terms of which skeptical alternatives can be “ruled out” or “eliminated” by the subject S, given S’s epistemic position.

Consider a case described by Fred Dretske in “Epistemic Operators.” You are in a zoo, you see an animal that looks, in every respect that you can discern, like a zebra. It is in an enclosure with a sign that says that this is the home of the zebras. This information might be enough for me to attribute to you knowledge that p: the animal in front of you is a zebra.

But now consider the fact that p is incompatible with q: that the animal in front of you is a mule cleverly painted by the zoo staff to resemble a zebra. To put it another way, if p is true, then ~q (the animal in front of you is not a mule cleverly painted by the zoo staff to resemble a zebra) is true. I might then be inclined to attribute ignorance of ~q to you and say that you do not know that ~q, because you are not in a position to detect this kind of fraud. But if you do not know that ~q, and q is incompatible with p, how could you know that p? That is, if you do not know that the animal is not a painted mule, how is it that you know that it is a zebra?

Dretske has an answer to this question. He says that knowing that p is relative to a set of contrasts to the truth of p. These contrasts are alternatives to the state of affairs in which p is true. In our example, a contrast might be a state of affairs in which the animal is a horse, or in which what you saw is not an animal at all, and so forth. If I attribute knowledge to you, it is (in part) because I think that you are in a position to rule out these alternatives. The alternatives that need to be ruled out for an attribution of knowledge are those that are relevant to whether or not you know. If you can rule out all of these relevant alternatives to the truth of ‘The animal in front of you is a zebra,’ then you are in a position to know. So the epistemic standard I use in this case is that of your being able to rule out all relevant alternatives in which p is false.

However, in the case of ~q, that the animal in front of you is not a cleverly painted mule, a further set of alternatives becomes relevant: those in which the animal is a cleverly painted mule. The alternatives in which p is false, where the animal is a painted mule, become relevant to ~q because q is its negation. It seems reasonable to demand that in order to attribute knowledge that ~q to you, I must at least think that you are able to rule out the alternatives in which q is true, i.e., in which ~q is false. The alternative in which q is true seems to be as relevant to the knowledge that ~q as anything could be.

Yet you are not in a position to rule out these alternatives. The reason you cannot rule them out is that you are not close enough to the animal to detect signs of paint, you have no information about the motives of the zoo staff, and so forth. Since you are not in a position to rule out alternatives that are relevant, I must conclude that you do not know that the animal is not a cleverly painted mule.

The only difference between the case where I attribute knowledge that p to you and the case in which I attribute to you ignorance that ~q is the proposition involved in the attribution. If Dretske is right, the knowledge attributions we make will vary depending on which proposition is a candidate for knowledge. The proposition that the animal is a zebra is associated with one range of relevant alternatives, and the proposition that the animal is not a cleverly painted mule is associated with another.

But as just noted, we are faced with a puzzle. Suppose I agree with Dretske that you do not know that the animal before you is not a cleverly painted mule. How, then, can I also claim that you know it is a zebra, since zebras are not, and cannot be, mules? Dretske’s account of relevant alternatives allows what DeRose calls the “abominable conjunction”: you know that the animal is a zebra but you do not know that it not an animal that cannot be a zebra. This pair is not logically contradictory. It does not consist of the propositions that the animal is a zebra and that the animal is a mule. Nor does the pair assert that you know it is a zebra and you know that it is not a zebra. Rather, it is a pair of epistemic attributions according to which you know something but do not know what is implied by what you know. But despite the two claims’ being logically consistent, it does sound deeply suspicious to think that both attributions can be correct. The basis for the suspicion seems to concern S’s epistemic position. If S’s epistemic position is strong enough with respect to p, and p is not true unless ~q is, why is S’s epistemic position not strong enough with respect to ~q? Should not the same standard of strength of epistemic position that applies to p apply to ~q as well?

Dretske answers in the negative, and he takes this consequence of his view in stride. As far as he is concerned, this is just how the verb ‘knows’ works. As he puts it, “What I am suggesting is that we simply admit that we do not know that some of these contrasting ’skeptical alternatives’ are not the case, but refuse to admit that we do not know what we originally said we knew” (“Epistemic Operators”). The fact that skeptical alternatives (like the painted mule alternatives discussed above) that cannot be ruled out is relevant to knowing one proposition does not mean that one lacks knowledge of another, albeit intimately related, proposition. ‘Knows’ functions in language as an epistemic operator, and epistemic operators have the property of being “nonpenetrating.”

These days, it would be said that Dretske accepts that ‘knows’ violates a “closure” condition, specifically with respect to conditionals that are known to hold. As discussed in the introductory module, one such closure condition is that if S knows that p, and S knows that if p then q, then S knows that q. Here we will represent a simple instance of this kind of closure in argument form.

Ks,c(p ⊃ q)
Therefore, Ks,cq.
An equivalent formulation is that if S does not know that ~q, and S knows that p ⊃ ~q, then S does not know that p.
Ks,c(p ⊃ ~q)
Therefore, ~Ks,cp.
Thus if you do not know that the animal is not a painted mule, and you know that if an animal is a zebra then it is not a painted mule, then you do not know that the animal is a zebra. For Dretske, closure threatens skepticism, and he meets this threat by rejecting closure.

Cohen’s Indexical Contextualism

One appeal of contextualism is that it provides a way of saving closure. Cohen, in his original formulations of contextualism, rejects Dretske’s solution to the skeptical problem. He does so because he thinks that closure must be preserved, or in Dretske’s terms, that ‘knows’ is a “penetrating” operator. Cohen’s approach, in effect, is to say that the closure principle can be upheld because the word ‘know’ cases such as the “zebra” case is used ambiguously. That is, you do not know (in one sense) that ~q, but you do know (in another sense) that p. We can represent this by using the superscript ‘H’ to indicate high standards (by which it is not known that the animal is not a mule) and ‘L’ to indicate low standards (by which it is known that the animal is a zebra). No index need be placed on the second premise, though it can be presumed that the biological knowledge that zebras are not mules may be presumed to meet the higher standard.

Ks,c(p ⊃ ~q)
But, KLs,cp

What determines the sense of ‘know’ is the context of attribution. When I say that you know that the animal is a zebra, I only expect your epistemic position to be as strong as it need be where deception is not an issue. But when I say you do not know that the animal is not a cleverly painted mule, I expect your epistemic position to be strong enough as it need be if deception is an issue.

As Cohen describes the matter, what has happened is that consideration of q creates a “skeptical context,” in which the standard for knowledge is raised. In such skeptical contexts, what counts as a relevant alternative is different from what counts as a relevant alternative in an “everyday context.” Thus, in the skeptical context, more exotic alternatives must be ruled out, and you cannot rule out the alternative that the animal is a cleverly painted mule.

The contrast between Dretske’s and Cohen’s approach comes down to this. For Dretske, there is a single meaning of ‘knows,’ but the range of relevant alternatives that determines whether one knows varies with with respect to the proposition in play. Because there is a single meaning of ‘knows,’ closure must be denied. Cohen holds that closure can be saved by allowing that the meaning of ‘knows’ changes from context to context of attribution. Within the same context, closure holds, but when the context of the premise and the conclusion of closure operate with a different sense of ‘know’ due to shifting contexts of attribution, there is no failure of closure.

Cohen holds that knowledge attributions function in the same way as the use of indexical terms such as ‘I,’ ‘here,’ and ‘now.’ From the standpoint of pragmatics, the appropriateness of the use of sentences containing these terms varies with the context of use. Cohen draws a semantic consequence from this: that the truth-values of these sentences differ as well. Whether sentences such as ‘I am a philosopher,’ or ‘My computer is here,’ or ‘G. J. Mattey is typing now’ are true or false varies depending respectively on who is asserting it, where the person asserting it is, and/or when the assertion is made.

The reason given by Cohen for the variation in the truth-values is that these sentences express different propositions in different contexts of utterances. If Cohen utters, ‘I am a philosopher,’ the proposition expressed is that Stewart Cohen is a philosopher at the time he utters the sentence. If I (the author of these notes) utter the sentence ‘I am a philosopher,’ the proposition expressed is that G. J. Mattey is a philosopher at the time he utters the sentence.

We might express the indexical view by thinking of ‘knows’ as an elliptical verb that contains a hidden parameter. When I say, ‘S knows that p,’ what I am really expressing is the proposition that S knows-by-the-standards-dictated-by-my-context that p. Here the phrase ‘my context’ is explicitly indexical. Suppose a context of attribution results in two propositions being generated by two standards of knowledge, a high standard H and a low standard L, applied to a single sentence. Presumably the difference lies in there being two different knowledge-relations invoked by the context: knows-by-H and knows-by-L. The extensions of these relations may vary, so that the triple <S,c,p> might be in the extension of the relation knows-by-L but not in the extension of knows-by-H.

An Objection to Indexical Contextualism

If indexical contextualism is correct, then the word ‘knows’ functions like an indexical. A fact about indexicals is that variation of the context of use does not affect the semantic assessments we make of them. For example, if I asserted late last night, “I am very sleepy,” while I assert this afternoon, “I am not at all sleepy,” I now take it that “I am very sleepy” was true last night. I will stand by the assertion and will not retract it.

Now consider the case of Parker, who asserted an hour ago, “I know that my car is where I parked it.” Then the context of use is changed by the mention of the possibility that Parker’s car may have been stolen in the last hour. Now his standards are higher, and Parker says, “I do not know that my car is where I parked it.” If asked about whether he knew it a few moments ago, when he asserted that he did know it, Parker might be inclined to say that he did not not know it earlier. This phenomenon must be accounted for, but the claim that ‘knows’ is indexical is of no help here, indexical sentences are not subject to retraction when the context of their use changes. I do not reject the truth of the sentence ‘I am at home’ as it was asserted this morning even though at this time I am in my office. It seems that the standards in play in the current context (where higher standards are in play) govern all semantical evaluations made in other contexts (where lower standards are in play).

The indexical-contextualist might respond in the following way. It is appropriate for Parker to say, “In fact, I did know, according to the standards at play at the time, that my car is where I parked it, but I do not know this by present standards.” This response can be criticized as being one which we do not ordinarily make. Instead, under the influence of higher standards, we tend to think that we really did not know at the time when lower standards were in play. (This idea is developed a little further in the discussion of Lewis’s contextualism, below.) This fact might be explained in one way or another by an indexical-contextualist, but we will not pursue such explanations here.

How the Standard for Knowledge Changes

So far, we have been working with the claim that certain features of the context of attribution either raise or lower the standard for knowledge. It would be good to make this claim more concrete by specifying more precisely how the standard is raised or lowered, or what it is about the context of attribution that dictates one standard or another. If we think of the standard of knowledge in terms of being able to rule out relevant alternatives, then the standard is raised when the range of relevant alternatives is increased and lowered when the range is decreased. That is, a higher standard makes relevant more alternatives in which p is false and a lower standard makes relevant fewer alternatives.

We have stated that the motivation for Cohen’s contextualism lies in the fact that in some contexts of attribution, the standards governing how strong an epistemic position a subject must be in to be attributed knowledge are different. If we think of the epistemic position required for knowledge attribution in terms of the range of relevant alternatives that must be capable of being ruled out, the next question is exactly which alternatives become relevant or irrelevant in which contexts of attribution. For Cohen, the factor that would do the ruling out is S’s evidence, but any factor that might be used to account for warrant could be invoked instead.

One way in which context of attribution increases the range of relevant alternatives that must be ruled out is through the awareness by the attributor of skeptical possibilities. For example, if I as the attributor consider the possibility of fraud at the zoo in which zebras are being disguised as mules, then your epistemic position must be stronger if you are to know that you see a zebra. It may be a general principle governing the use of ‘knows’ that if a skeptical possibility is taken into account in the context of potential attribution of knowledge that p, then any alternative in which p is false and the skeptical possibility holds becomes relevant.

More generally, it seems that whenever a possible state of affairs is brought to the attention of an epistemic attributor, it becomes relevant and can have an effect on the standards in play. David Lewis describes a “rule of attention” to this effect. For a given alternative, “if in this context we are not in fact ignoring it but attending to it, then for us now it is a relevant alternative” (“Elusive Knowledge,” 1996).

Thus, if the alternative is that what you see is a mule cleverly disguised as a zebra, then any situation in which I, the attributor, consider the possibility of a mule cleverly disguised as a zebra makes that possibility relevant to whether you know. Perhaps even my consideration of the generic possibility of your being deceived could make relevant the alternative in which you are deceived by the zoo's staff’s disguising a mule as a zebra.

DeRose’s Sensitivity Condition

Keith DeRose has presented a version of contextualism that makes no semantic claims but which may be discussed profitably at this point. He has argued that the relevant alternatives approach leaves one fact unexplained. If it is granted that in a context in which an alternative in which p is false is mentioned, it must be ruled out, the question still remains whether that alternative can be ruled out. To put it another way, the question as to whether S’s epistemic position is strong enough is not answered simply by the specification of what is relevant and what is not. DeRose is concerned about cases in which a powerful skeptical hypothesis is advanced by a skeptic. For example, a skeptic might tell the the attributor that you do not know that you do not see a mule cleverly painted to resemble a zebra. At this point, the alternative that what you see is a painted mule becomes relevant, and as a result the attrbutor must think that you are in a position to rule out this alternative in order to be able to attribute knowledge to you. This relative weakness of your epistemic position must be explained. In what way is the mule alternative not excluded by the strength of your epistemic position?

The answer to this question depends on how the relative strength of one’s epistemic position is understood. One proposal, made originally by Dretske but popularized by Robert Nozick, is that S’s epistemic position with respect to p is strong enough if S’s belief that p “tracks the truth.” The basic idea is that if S’s belief that p tracks the truth, then if p were not true, then S would not, in the altered circumstances in which p is not true, believe that p. So in the zebra case, if you believe that you do not see a painted mule, that belief does not track the truth, since if what you saw were a painted mule, then you would still (in that circumstance) believe that you do not see a painted mule. In the literature, it is said that if S’s belief that p tracks the truth, then S is “sensitive” to the truth of p.

DeRose proposes that the reason your epistemic position with respect to not seeing a painted mule is not strong enough is that your belief that you do not see a painted mule is not sensitive. DeRose goes on to propose a conversational “rule of sensitivity” which is more specific than Lewis’s “rule of attention.” In contexts of epistemic attribution, where it is asserted that S knows that p or that S does not know that p, the standard of knowledge requires that S’s belief (or would-be belief) that p be sensitive to the truth of p. So if p is ‘you see a zebra,’ then when someone claims that you know that you see a zebra, that person’s standard demands that if you did not see a zebra, then you would not believe that you see a zebra.

What about your belief that you see a zebra? Does it track the truth? It is generally held that whether the belief at some time tracks the truth depends on the actual state of the world at that time. If you are in a normal zoo and there is no deception at play, then if you were not seeing a zebra, then you might be seeing some other animal in its ordinary condition or no animal at all. For example, the zoo-keepers might be quite negligent and sometimes put the animals in the wrong enclosures. They are prone to put antelopes in the zebra pen. Given that you can distinguish zebras from other ordinary animals and from non-animals, your belief tracks the truth.

When we say, if it were the case that q, then it would be the case that r, we have in mind those cases where q is true to be as close as it can be to the actual situation. On the other hand, if there is some deception going on in the environment, the non-zebra possibilities are close enough that you would have to be sensitive to them. If the deception would not fool you, you would be sensitive to the truth of the proposition that you see a zebra. On the other hand, if you were not seeing a zebra, you might still believe you are seeing a zebra, in which case you are not sensitive to the truth of the proposition that you see a zebra.

So if someone is trying to evaluate whether you know that you see a zebra, that person will have to ask, “If he were not seeing a zebra, would he still believe that he was?” Keeping things as close to the actual situation (as understood by the attributor) as possible, the attributor might be inclined to say “No, he would not. He can tell the difference between a zebra and an antelope or some other thing that is not a zebra.” So imposing the sensitivity condition need not affect knowledge attributions in ordinary contexts.

On the other hand, consider the question of whether to attribute to you knowledge that you are not looking at a mule cleverly painted to resemble a zebra. For me to attribute knowledge to you, I would have to regard you as sensitive to the proposition that you are not looking at a mule cleverly painted to resemble a zebra. For you to be sensitive to that proposition, if you actually are looking at a painted mule, you would not believe that you are not looking at a painted mule. The attributor would have to answer this question: “If he were looking at a painted mule, would he still believe that he not looking at a painted mule?”

Now the attributor has to take into account cases as close to reality as possible where you are looking at a painted mule and consider what would happen if that is what you saw. As the circumstances have been described, your epistemic position is weak, because from your vantage point you cannot tell the difference between zebras and painted mules. Because of your relatively weak epistemic position, the answer should be, ‘Yes, he would still believe that he is not looking at a painted mule, because he believes he is looking at a zebra.’ So the attributor would not take your belief that you are not looking at a painted mule to be sensitive, in which case he would say that you do not know that you are looking at a painted mule.

DeRose notes that his rule of sensitivity comes into play only in contexts of knowledge attribution or denial. The original application of sensitivity to knowledge by Nozick took the form of a necessary condition in the analysis of knowledge. So for Nozick, sensitivity is part of the concept of knowledge itself. This analysis has been effectively criticized by Kripke and Vogel, but only as a condition of, or part of the concept of, knowledge itself. It remains to be shown whether the limited application of the sensitivity condition does not fall prey to these criticisms.

Bringing sensitivity into play only in contexts of knowledge attribution or denial may be too restrictive. It may be that sensitivity is called for in other contexts. According to the “rule of attention” described above, a skeptical possibility becomes salient or relevant for knowledge attributions when it is being actively considered. We may want to say in parallel fashion that a conversationally live skeptical possibility should trigger a demand for sensitivity. Suppose that someone has casually mentioned to me that the animal before you might be a mule painted to resemble a zebra. Then it seems that I should be forced to consider the conditional: if the animal were not a zebra, you would not believe believe that it is a zebra.

It should be noted that neither the relevant alternatives nor the sensitivity approaches as described here have anything to say about cases where practical considerations raise or lower the standard of knowledge. While we can say generally that the more important it is to the attributor that he be right, the more alternatives become relevant, it is hard to see how to quantify this in a precise way. It does not seem that the specific reason why one needs the attribution to be right would dictate which specific relevant alternatives would have to be ruled out. There probably is no systematic connection. The weaker “rule of attention” seems to be as much as actually comes into play in conversational contexts. The same reasoning holds for differences in the practical situation of the attributee. The consequences of being wrong that come to mind are the ones that have to be taken into account.

Lewis’s Contextualism

There is a powerful linguistic reason to think that the truth-conditions of sentences of the form ‘S knows that p’ are determined at least in part by pragmatic considerations. The reason is that ‘knows’ seems to function in ordinary language as a vague term, which is used in a way analogous to the way we use words like ‘tall,’ and ‘bald.’ It is context which helps to determine the range of application of vague terms. Usually the “precisification” (or making precise) of vague terms depends on some comparison class. In a context in which the attributor has in mind one class of people, say students in a philosophy class, ‘tall’ would apply to someone who is six-feet four inches in height. But when the comparison class is professional basketball players, the range of application of ‘tall’ might not include a person who is “only” six-feet four. Similarly, in ordinary usage, we make knowledge attributions liberally in some contexts but not in others. (But as Jason Stanley has pointed out in “On the Linguistic Basis for Contextualism,” the fact that we do not say that S knows that p better than does S′ is reason to think that ‘knows’ does not function as a vague term, even if we can speak of being in a better epistemic position. On the other hand, vague terms allow for comparisons, such as that one thing is taller than another.)

An early version of attributor-contextualism may be found in David Lewis’s 1979 article “Scorekeeping in a Language Game.” It appears rather briefly in a discussion of vague predicates and is concerned with certainty rather than knowledge. The context in question is that of conversation, rather than knowledge-attribution per se.

The main point in the article is that successful conversation often requires the implicit use of “rules of accommodation” which are sensitive to what is being said. The predicate ‘certain’ is vague in that it is subject to different standards of precision. It is the conversational context which determines the standards of precision. What is “true enough” in a context subject to one standard becomes no longer “true enough” when a higher standard is introduced.

Lewis criticizes an argument for skepticism by Peter Unger (Ignorance, 1975). Unger had argued that almost nothing is flat. Take something, say a desktop, we call “flat,” there is something we would call “flatter.” But nothing can be flatter than what is flat, so the desktop is not flat after all. Unger contends that ‘certain’ is an “absolute term” like ‘flat.’ If nothing can be no more certain than what is certain, he concludes, there is little if anything that is certain, and hence we are either entirely or nearly entirely ignorant. “Our language habits might serve us well in practical ways, even while they involve us saying what is not true,” i.e., that we are certain (Ignorance, p. 48). Lewis agrees that it is inconsistent to say that something is more certain than something that is certain. (As noted above, Stanley thinks that ‘knows’ is not a vague term, from which we might conclude that it functions as as what Unger calls an “absolute term.”)

But we do often compare degrees of certainty. Here is such a case. The clouds are dark, the air is heavy, the wind is blowing, and a radar map shows heavy rain very nearby and upwind. I might say, “It is certain that it will rain soon.” My interlocutor might add, “It is even more certain that the sun will set today.” This brings to mind a way in which I might be in a better epistemic position (albeit with respect to some other proposition). Because ‘certain’ is an absolute term, this seems to create a problem: it is inconsistent of me to say that something is more certain than something that is certain, so I must conclude that it is after all not certain that it will rain in a few minutes.

Lewis argues that there is a sense in which this is correct, but at the same time, he thinks that to deny that I am certain is quite misleading. When my interlocutor has spoken, he has raised the standard in play. According to that raised standard (given that ‘certain’ is an absolute term), it is not now “true enough” to say that it is certain that it will not rain soon, since something is more certain than that. But this does not imply that it was not “true enough” to say that it is certain that it will rain soon in the context in which it was initially said.

Technically, we treat a sentence as being “true enough” in a conversational context when we treat it “more or less as if it is simply true.” We treat a sentence more or less as if it is simply true “if it is true over a large enough part of the range of delineations of its vagueness.” How large is large enough is determined by conversational context. So if a sentence of the form ‘S is certain that p’ is true given the present and lower standards for certainty, it is treated as if it were simply true because it is true within a large enough range of cases we are willing to take as being certain.

So Lewis’s response to Unger is that if we succeed in raising the standards extremely high, then in that context, very little is certain. But this does not affect claims to certainty in other contexts where the standards are lower. Applying this to knowledge, we get the result that Lewis’s “rules of accommodation” make attributions of knowledge appropriate in some conversational contexts and inappropriate in others. If he is right, we can say that the pragmatics of ‘knows’ treats it as a vague term whose use is good enough for some contexts and not good enough in others. The semantical counterpart of this pragmatic claim is that epistemic attributions are true enough in some contexts but not true enough in others.

An observation Lewis makes is that when standards change in conversation, they tend to become more rather than less strict. When someone invokes the highest standards possible, it may change the conversational context so much that it interferes with the purpose of the conversation. If we apply this to epistemic attributions, someone could introduce infallibility as a standard when the discussion concerned some ordinary matter. Even though that might make it impossible to attribute knowledge even when it ordinarily would be appropriate, the person introducing the higher standard may be able to “get away with it.” The reason is that introducing higher standards in general “seems commendable.” Thus a skeptic in a certain way has an automatic advantage when introducing a very high standard.

Lewis’s rules of accommodation are quite loose. All that is required is that “something is said” which raises or lowers the standards at play in the conversation. DeRose (“Solving the Skeptical Problem,” 1995), argues that a rule of accommodation of this sort is too loose. DeRose’s primary objection is that the rule is too powerful. Let p be ‘I have two hands.’ A “simple skeptic” might simply deny that I know that I have two hands. This simple denial, according to DeRose, “should raise the standard of knowledge to such a point as to make her claim true.”

But in fact, DeRose maintains, the normal reaction to the simple skeptic would not be to raise the standard, but merely to dismiss her claim. On the other hand, if a more sophisticated skeptic were to raise the possibility that I am a victim of a Cartesian demon, then the standard of knowledge would be raised by the introduction of this possibility.

It must be said that it is not clear why DeRose would think that the simple skeptic would raise the standard in the first place. The idea seems to be that when the simple skeptic says, ‘You do not know that you have two hands,’ on the basis of the skeptic’s standard (whatever it might be), then the skeptic’s standard is the new standard in play. Suppose I said ’she is tall,‘ and someone said to me ‘She is not tall.’ I would perhaps have to inquire as to the person’s standards of tallness in order to keep the conversation going. Now if I inquired of the simple skeptic what her standards were, and she reported some standards higher than my own, then there is a situation like that with the sophisticated skeptic.

A secondary objection by DeRose is that the standards should be raised in a non-skeptical context, as when someone says, non-skeptically, that he can rule out certain possibilities. But by DeRose’s understanding of Lewis’s rules of accommodation, the standards are not raised.

In his 1996 paper, “Elusive Knowledge,” Lewis revisits the effect of rules for conversational success on epistemic attributions, focusing now on knowledge rather than certainty. The contextualist strategy of Cohen and DeRose makes a concession to skepticism. In some contexts of attribution, we are forced to attribute ignorance because (in the language of relevant alternatives) we cannot rule out extreme alternatives that the context of attribution has made relevant. But as Cohen would have it, in attributing ignorance, we express a different proposition from that expressed when we attribute knowledge to the same subject in the same epistemic position.

Lewis claimed that one such context that forces ignorance is that of doing epistemology.

Do some epistemology. Let your fantasies rip. Find uneliminated possibilities of error everywhere. Now that you are attending to them, just as I told you to, you are no longer ignoring them, properly or otherwise. So you have landed in a context with an enormously rich domain of potential counter-examples to ascriptions of knowledge. In such an extraordinary context, with such a rich domain, it can never happen (well, hardly ever) that an ascription of knowledge is true. Not an ascription of knowledge to yourself (either to your present self or to your earlier self, untainted by epistemology); and not an ascription of knowledge to others. This is how epistemology destroys knowledge. (“Elusive Knowledge”)
According to Lewis, knowledge is destroyed only temporarily, when the rich domain of possibilities of error has our attention. In other, more ordinary, contexts, we have the knowledge we think we have.

One feature of Lewis’s view that distinguishes it from Cohen’s is that Lewis eschews fallibilism in favor of infallibilism. “I propose to take the infallibility of knowledge as my starting point.” As Lewis recognizes, infallibility has been abandoned as a condition of knowledge because it is thought to lead to skepticism. Consider Lewis’s initial definition of knowledge:

Subject S knows proposition P iff P holds in every possibility left uneliminated by S’s evidence.
When we let loose our skeptical fantasies, we find that our evidence does not eliminate every possibility in which p does not hold. So it seems that “infallibilist epistemology [must] end in skepticism.” Yet Lewis thinks that skepticism can be avoided when we are not doing epistemology. How could this be?

The way around the problem is to exploit our everyday use of the quantifier term ‘every.’ In conversational contexts, we restrict the range of the quantifier to a certain group of objects. If I say in the classroom, ‘Everyone will be present at the final examination,’ I am not talking about all people, but only about all members of the class. Other people may be properly ignored. Similarly, the context in which the definition of knowledge is stated (or thought) will determine the scope of ‘every’ within it. In ordinary contexts of use of the definition, ‘every’ has as its “domain” a range of possibilities which does not include skeptical possibilities, in which p does not hold, that cannot be eliminated. So the only possibilities that must be eliminated by S’s evidence are those which are properly ignored. Semantically, a sentence containing a universal quantifier expresses different propositions depending on how the context determines the range of the quantifier.

When we are doing epistemology in a way that invokes skeptical possibilities, we cannot properly ignore them. The reason lies in the afore-mentioned “rule of attention,” according to which what has our attention cannot be ignored, and therefore cannot be properly ignored. (Lewis describes six other rules delimiting what can and cannot be properly ignored.) When we are not doing epistemology, these possibilities are properly ignored, which allows us to have knowledge.

One is left wondering whether Lewis has remained faithful to his stated infallibilist starting-point. His exploitation of the conversational restriction on the range of the quantifier ‘every’ requires a distinction between absolute and relative infallibility. If the range of the quantifier is universal, that is, if every possibility in which p is false must be eliminated by S’s evidence, then the infallibility of knowledge is absolute. If, on the other hand, the range is restricted, then to know, one must only be relatively infallible. So in ordinary contexts, knowledge is infallible, but only relatively so. It appears that any distinction between relative infallibility and fallibility is only a verbal one. For non-epistemological contexts, one is “infallible” with respect to p in name only. Although in a formal sense what is meant by ‘knows’ is invariant across contexts, when the definition of ‘knows’ is interpreted by determining the range of the quantifier, it is seen that the word has different meanings in different contexts.

Non-Indexical Use-Contextualism

A quite different version of use-contextualism has been proposed by Nikola Kompa (“The Context Sensitivity of Knowledge Ascriptions,” 2002). On her view, the verb ‘knows’ is “unspecific,” in the sense that sentences containing it are evaluated by different standards. A single property is expressed by ‘knows,’ but what objects fall under that property varies across contexts of use of the sentences. Thus the truth-value of ‘S knows that p’ is variable, despite the fact that ‘knows’ expresses a single property and ‘S knows that p’ expresses a single proposition regardless of its context of use.

This approach is developed in more detail in MacFarlane’s 2007 paper “Nonindexical Contextualism”. MacFarlane cites the example of ‘temporalism’ in the semantics of time-sentences. According to temporalism, the sentence ‘Socrates is sitting’ expresses a single proposition no matter when it is asserted, but it will vary in truth-value with the time at which it is asserted. Just as the temporal sentence’s truth depends on a time parameter, a knowledge attribution will depend on an epistemic standard parameter. According to a low standard, the sentence ‘Parker knows that his car is where he left it’ is true, but according to a higher standard, the sentence ‘Parker knows that his car is where he left it’ is false. A key feature of MacFarlane’s approach is that commits him only to variable truth-values of sentences and not to variable propositions, as Cohen would have it.


MacFarlane had earlier proposed an alternative account of the semantical features of ‘knows.’ His idea was that the variability in knowledge attributions is not a matter of our general way of using ‘knows’ (pragmatics). Rather it has to do with the specific use of the term when making semantical claims about it. The source of variability lies in “contexts of assessment.” “A context of assessment is a situation in which a (past, present, or future, actual or merely possible) use of a sentence might be assessed for truth or falsity” (“The Assessment Sensitivity of Knowledge Attributions”). MacFarlane proposes a new kind of contextualism (which he calls “relativism” but shall here be called “assessment-contextualism”), according to which different epistemic standards are in play in different contexts of assessment. Moreover, the context of assessment is the only context which invokes different epistemic standards. To put it another way, context comes into play only in evaluating knowledge sentences semantically, but not when otherwise using the terms ‘knowledge’ and ‘knows,’ as in ordinary epistemic attributions.

This allows him to solve problems faced by the use-contextualist and subject-sensitive invariantism. Recall that the use-contextualist must account for why it seems right to retract a low-standards knowledge attribution when the standards are raised. This does not parallel our use of indexicals such as ‘I.’ The problem for subject-sensitive invariantism is that the same subject can correctly be attributed both knowledge and ignorance, without any change in his epistemic position, yet this is not how we speak. (Hawthorne seems to have regarded this consequence of the view as a “feature” rather than a “bug.”)

On MacFarlane’s view, retraction occurs when the earlier belief is being assessed. Then the high standard in play applies to all potential knowledge attributions in that context, including those made in the past when lower standards were in play. The variability in correct attribution that is a consequence of subject-sensitive invariantism simply does not occur with assessment-contextualism. So this version of contextualism conforms more closely to ordinary ways of speaking.

According to assessment-contextualism, the proposition expressed by knowledge attributions is the same in different contexts of assessment. This differentiates it from indexical-contextualism. But it also raises the question as to how the same proposition can be true in one context of assessment and false in others. MacFarlane claims that the context of assessment determines the standard by which the proposition is either true or false. Given that some standards are higher than others, it follows that the truth of propositions is relative to the context of evaluation. The upshot is that “knowledge attributions are not as robustly objective as ordinary claims about the world” (“The Assessment Sensitivity of Knowledge Attributions”).

MacFarlane compares the relative subjectivity of knowledge attributions to the relative sensitivity of aesthetic attributions, such as deliciousness. In different contexts of assessing the deliciousness of a meal, we bring different standards to play. And we are not tempted to say that it is simply true or simply false that the meal is delicious.

An objection to MacFarlane’s truth-relative semantics is that it seems to be incompatible with notion that truth is the “internal aim of assertion.” If the truth of a proposition may vary according to the context in which it is assessed, what is the point of making knowledge attributions in the first place? Is the semantics of ‘knows’ coherent with the pragmatics of ‘knows?’ MacFarlane responds by rejecting the claim that our aim in making assertions is to state the truth. Our aim is, he suggests, instead to committing ourselves to the truth of a proposition in a context of assessment. The commitment is to vindicate the claim when it is challenged.

The final step is to adapt this account of assertion to the assessment of knowledge claims. The idea is that in a given context of assessment (say, with high standards), one must decide whether to vindicate the attribution of knowledge according to the standards at play in that context. If one is unwilling to do so, then one will withdraw the assertion. Suppose I, in a low-standards context of assessment, attribute to Parker knowledge that his car is where he parked it. I have thereby committed myself to defend that claim, by those standards. But in a high-standards context of assessment, where the possibility of theft is relevant, I withdraw my original claim.

MacFarlane acknowledges that making the truth of a proposition relative to the conditions of its assessment is a hard sell. Not having to accept this kind of truth-relativity makes his later non-indexical contextualism more attractive. MacFarlane had proposed propositional truth-relativity because he had assumed that tying standards to the context of use would require treating ‘knows’ as an indexical term. He later came to realize that the context of use does not have to be treated indexically. This provided an opening for him to think of ‘knows’ without making the truth of propositions relative to contexts of assessment.

We may summarize the various claims about the semantics of ‘knows’ in a table. Let p be a sentence and let C1 and C2 be two different contexts, of use or of assessment depending on the case. And let the epistemic position of an epistemic subject S be fixed. But let c1 and c2 be two distinct circumstances for S. Now we can form a chart that delineates the various positions.

Truth-value Proposition Expressed
Same Different
Same Subject-Insensitive Invariantism (circumstances ci irrelevant)
Subject-Sensitive Invariantism (circumstances ci relevant)
Different Kompa (Ci: use. Sentences differ in truth-value)
MacFarlane (Ci: assessement. Propositions differ in truth-value)
Cohen (Indexical)
Lewis (Restricted Quantifier)

On no view does the sentence ‘KS,cp’ express different propositions with the same truth-value depending on either the circumstances c of S or the context C of use or assessment.


MacFarlane notes that there is an alternative to invariantism and various forms of contextualism: “eliminativism about knowledge.” According to this view, “there may be no fully coherent way to assign truth-conditions to our knowledge-attributing sentences. The rational course of action would be to reform our thought and talk by introducing new, unfocused terms of epistemic assessment” (“The Assessment Sensitivity of Knowledge Attributions”). MacFarlane calls such a proposal “radical” and claims that “it should not be accepted unless there is no good alternative.” As to whether there is a good alternative, the jury is still out. We have here explored a rather bewildering array of accounting for our linguistic behavior with respect to a single sentence-type, ‘S knows that p.’ In fact, there are many more arguments and positions than could be described here. As Patrick Rysiew points out in his Stanford Encyclopedia of Philosophy article on contextualism, “Unsurprisingly, no clear consensus has emerged.” What would be surprising is if a clear consensus ever were to emerge and eliminativism were not to win the day.

One reason for pessimism concerns the practical dimension of knowledge-attribution. Thus far, the focus has been on the theoretical component of knowledge attribution and on the context of attribution. That is, we have been focused merely on the theoretical possibility of error. But as has been noted above, the way we make attributions of knowledge and ignorance rests to a great extent on the practical consequences, for the attributor and/or the subject, of being wrong in the attribution. It is difficult to see how these practical considerations can be summarized in any simple rule that would specify which alternatives are relevant, or which can be properly ignored.

One problem with devising such a rule is that it seems that there is a trade-off between the theoretical and the practical. The situation of the subject may be such that practical urgency renders irrelevant skeptical alternatives that are relevant from a theoretical point of view. Suppose the attributor is in a life-or-death situation, and whether S knows that p is directly connected to his chances of survival. The attributor may somehow still have skeptical alternatives fresh in his mind. Yet he might not judge an evil demon or brain-in-a-vat alternative to be relevant when pondering whether S knows that, say, that the door leads to an escape route.

[ 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. All rights reserved.