Contextualism in Epistemology

A knowledge attribution is a claim that someone knows that something is the case. What conditions must be met in order for a knowledge attribution to be true? Over and above true belief, knowledge involves various “epistemic” factors, such as justification, warrant, reliability, and so on. We can refer generically to such factors by saying that knowing involves being in a good epistemic position. But how good is good enough? How good an epistemic position must you occupy in order for it to be true to say that you know something? Contextualists say that the answer varies with the context in which the attribution is made. A knowledge attribution might require one epistemic position when made in a “low standards” context, but a different, stronger epistemic position when made in a “high standards” context. The truth-conditions of knowledge attributions are thus, in a distinctly epistemic sense, context sensitive. What it means to claim that someone knows depends in part upon the context in which the claim is made.

The traditional opponent of the contextualist is the invariantist. Invariantists hold that the truth-conditions of knowledge claims are fixed across contexts. Invariantists differ among themselves about what those truth-conditions are, and in particular about the epistemic position required to know. Skeptical invariantists hold that knowledge is extremely epistemically demanding, while moderate invariantists think it is not so difficult to know things. Subject-sensitive invariantists think that the epistemic position required for knowledge depends on one’s situation. But all invariantists agree that what it means to claim that someone knows something doesn’t vary with the context in which the claim is made. Relativism is a newer competitor to contextualism, and a more radical departure from invariantism than is contextualism itself. According to the relativist, a particular knowledge claim can have different truth-values relative to different contexts of assessment. Even holding fixed the meaning of a knowledge claim and the underlying facts of the matter, the truth-value of the claim is not settled: it may be true for some, but false for others.

The debate about contextualism does not directly concern knowledge itself, though as we’ll see in section 4, different versions of contextualism reflect different ideas about the nature of knowledge. Contextualism is a theory about knowledge claims, and hence is not a theory of knowledge but a “complement to any theory of knowledge” (Heller 1999a, 115). Still, as Keith DeRose, a leading contextualist, points out, “it’s important in studying knowledge to discern what it means to say someone knows something” (DeRose 2009, 19). Contextualists hold that their view can be used to make progress on a range of epistemological problems, most notably the problem of radical skepticism.

The terms contextualism and invariantism were first used as names for these views about knowledge claims by Unger 1984. Classic developments include Stine 1976, Cohen 1988, DeRose 1995, Lewis 1996, and Heller 1999a; Rysiew 2011a (sec. 1) provides a thorough historical introduction. Early contextualism grew out of the “relevant alternatives” tradition in epistemology associated with Goldman 1976 and Dretske 1981; Pryor 2001 and Black 2003 situate contextualism in that framework. The present article provides an opinionated entry point to a number of current discussions concerning contextualism. After discussing the contextualist approach to skepticism (section 2), I will survey two key arguments for contextualism (section 3), and give an overview of debates among contextualists about how best to understand and develop their view (section 4). In the last decade a number of objections to contextualism’s adequacy as a linguistic thesis have emerged. These objections have made many epistemologists wary of endorsing contextualism. However, detailed and persuasive responses are available to each objection. In section 5, I discuss several of the most prominent issues and show that there is little reason to balk at endorsing contextualism on empirical grounds. I close by considering some alternative accounts of our ordinary knowledge talk (section 6).

“Contextualism” is used as a label for many philosophical views. Three distinct uses are relevant here. First, one can hold that the truth-conditions of any particular class of claims vary with the context of utterance, and thus endorse “contextualism” about such claims of a sort parallel to contextualism about knowledge claims. There are contextualist approaches to the semantics of causal-explanatory claims; modal language (“might,” “must,” “possibly,” “necessarily,” and so on); conditionals; evaluative terms (“good,” “bad”); normative claims involving “ought,” “should,” and so on; predicates of personal taste; and more. The language of epistemic possibility is obviously of central concern to epistemologists, and there are important connections between contextualism about epistemic modals and contextualism about knowledge claims (DeRose 1991; Egan, Hawthorne, and Weatherson 2005; Egan and Weatherson 2011). Nonetheless while there is overlap between the literature and traditions relevant to contextualism about other terms and the work discussed in this article, once the details of the positions and arguments come into focus, there are few obvious or uncontroversial connections. So the focus here is exclusively on contextualism about knowledge claims.

Second, “contextualism” is used by philosophers of language as a label for the idea that linguistic meaning is deeply context-dependent. Recanati, for example, says that contextualists hold that “only in the context of a speech act does a sentence express a determinate content” (Recanati 2005, 171; see also Cappelen and Lepore 2005 (ch. 1) for an influential definition). Those who embrace contextualism in this sense are likely to be friendly to some form of contextualism about knowledge claims, but the converse needn’t hold. One may be a contextualist about knowledge claims while endorsing an austere view about the nature and extent of context-sensitivity in natural language.

Third, “contextualism” has been used by epistemologists for theories that have nothing directly to do with semantics. Most notably, Annis 1976 introduced it as a name for a distinctive view about the nature of epistemic justification. Williams 1991 (191) follows Annis’s use of the term, and in later work characterizes it as “the view that all justification takes place in an informational and dialectical context” (Williams 2001, 179). Though the Annis-Williams version of contextualism is not immediately connected with contextualism about knowledge claims, the views are not unrelated; elsewhere Williams says that “the fundamental idea of contextualism is that standards for correctly attributing or claiming knowledge are not fixed but subject to circumstantial variation” (Williams 2001, 159).

Every major contextualist has developed his or her version of the view with an eye towards relieving perennial skeptical anxieties. Most of the discussion has focused on a particular form of skeptical argument that DeRose 1995 calls the argument from ignorance (AI).

Consider one of the many obvious facts O that it ordinarily seems that you know, for example, that you have hands, and a radical skeptical hypothesis H on which things would seem to you just as they actually do, but O would be false, for example, that you are the victim of a Cartesian evil demon. It is not difficult to get into a frame of mind in which it seems that you don’t really know that H is false. In that state of mind, the following argument will appear troubling:

Given that AI’s premises are compelling, there is pressure to accept its conclusion. But (AI3) is profoundly at odds with our ordinary perspective. If we wish to retain our ordinary perspective, how can we respond?

The defender of our ordinary perspective may deny (AI1), insisting that we do know Not-H. For this response to be satisfying, she needs an explanation for how we know Not-H. When H is well chosen, providing such an explanation is not an easy task. Alternatively, she may deny (AI2), holding that knowing O does not require knowing Not-H. But this response is at odds with independently plausible “closure” principles for knowledge: it seems that when you know P, you can extend your knowledge to include whatever you recognize that P entails. Moreover, the claim that one knows that O but not that Not-H is, in DeRose 1995’s phrase, an “abominable conjunction.” If this claim can be true, why does it seem so bizarre?

Contextualism offers a different path of resistance. Contextualists think that serious consideration of radical skeptical possibilities tends to raise the epistemic standards in the context to an extremely high level. Even our mundane beliefs rarely meet such extremely high epistemic standards. So serious consideration of Not-H tends, by the contextualist’s lights, to transform the context into one in which (AI1) expresses a truth. And given (AI2), which contextualists are keen to accept as true with respect to any context (though Heller 1999b provides a contrary view), the troubling conclusion follows. However, given that knowledge claims are context sensitive, the fact that (AI3) is true with respect to a skeptic-friendly context does not imply that it is true with respect to an ordinary context, where the epistemic standards are less demanding. Hence, the fact that AI is sound with respect to skeptic-friendly contexts is consistent with the truth of our ordinary knowledge claims. Cohen, 1986  1988, 1999, DeRose 1995, Heller 1999a, Lewis 1996, Neta 2004, and Stine 1976 all offer accounts in this vein.

Unlike straightforward anti-skeptical responses to AI, the contextualist treatment denies neither premise. On the contrary, the contextualist accepts that, as uttered in skeptic-friendly contexts, AI’s premises are often true. The contextualist’s aim is not to defend our ordinary perspective by engaging with skeptics on their own turf. It is rather to limit the force of the skeptic’s challenge, obviating the need to provide such a defense in the first place (Cohen 1988, 113-114, 1999, 79–80; DeRose 1995, 49–50). Once we understand how serious consideration of a skeptical hypothesis affects the truth-conditions of knowledge claims, we can recognize that our ordinary knowledge attributions are not undermined by the skeptic’s conclusion. In addition, the contextualist has a simple explanation for why AI is nonetheless so intuitively compelling: with respect to contexts where it’s considered seriously, its premises and conclusions are all apt to be true.

The contextualist treatment is successful, of course, only if contextualism itself is true; we’ll consider that question in sections 3, 5, and 6. But even if contextualism is true, one may doubt that its attempt to disarm the skeptic is successful.

A number of theorists have argued that because of its epistemic neutrality, contextualism is of little relevance to the concerns that have animated discussion of the skeptical problem.

Suppose that knowledge claims are context sensitive and that our ordinary knowledge attributions are true. As Sosa points out, from the fact that people sometimes speak truly when they utter “I know that I have hands,” it does not follow that people ever know that they have hands (Sosa 2000, 4). Even if we accept contextualism we may still wonder whether we know that we have hands. Relatedly, Feldman complains that contextualists “make it correct to use the word ‘know,’ but they don’t address the interesting doubts that have been raised about whether we know much” (Feldman 2001, 62). Sosa and Feldman both think that the contextualist’s description of the conversational dynamics exploited by AI is plausible. But since contextualism is neutral concerning the question of what we can know, it does not engage with the issue at the heart of the skeptical dialectic.

The objection misses the point of the contextualist response. Given contextualism, the answer to the question “What do we know?” will vary with the epistemic standards in place in the context in which it is asked. Asking whether people know that they have hands without regard to a particular epistemic standard is, for the contextualist, like asking whether it is raining without regard to any particular location. If the question is whether we know anything by extremely high standards, then contextualists generally concede that we don’t. However, recognizing the standards-relative nature of the question removes the sting of that concession. DeRose says, “[O]nce I start to get a clear look at what it would take to ‘know’ according to the skeptic’s absolute standards, I find the distress caused by my failure to meet those standards to be minimal at best—perhaps to be compared with the ‘distress’ produced by the realization that I’m not omnipotent” (DeRose 2004). On the other hand if the question is whether we know anything by ordinary standards, contextualists will insist that AI has done nothing to undermine our conviction that we do.

Yet it is not obvious that the AI-wielding skeptic leaves our ordinary knowledge claims unscathed. Kornblith considers the “full-blooded skeptic” who uses AI-style considerations to argue that “we have no more reason to believe that the world is as we take it to be than that it is altogether different or, indeed, that there is no such world at all” (Kornblith 2000, 25). The full-blooded skeptic’s intended target is any knowledge attribution, not just those made in skeptic-friendly contexts. Klein 2000 points out that the closure principle implies that our ordinary knowledge claims are true only if we know by ordinary standards that radical skeptical hypotheses are false. But, Klein’s skeptic argues, we “cannot have any evidence whatsoever for believing” that such hypotheses are false (Klein 2000, 110, emphasis in original). Without an account of how we can know by ordinary standards that radical skeptical hypotheses are false, the contextualist’s reassurances are apt to ring hollow.

Contextualists have offered such accounts. Cohen says that while we lack evidence that we are not bodiless brains in vats, it is “intrinsically rational” to believe that we are not (Cohen 1988, 112; Cohen 1999, 76–77). DeRose thinks, that provided the closest possible worlds where we falsely believe P are “quite distant” from the actual world, we can meet the relatively undemanding standards in place for knowing P in ordinary contexts (DeRose 1995, 35). Lewis says that in contexts where the possibility that P is false is “properly ignored,” one can “know P just by presupposing it”; the assumption is that skeptical hypotheses are properly ignored in ordinary contexts (Lewis 1996, 561; Stine 1976; Heller 1999a). However plausible these proposals are, they go beyond what Feldman 2001 calls “bare contextualism,” but are substantive epistemological claims. The challenge of full-blooded skepticism shows that while contextualism provides a key element of a solution to the skeptical problem, direct engagement with epistemology is still required for a philosophically satisfying resolution.

If sound instances of AI do not threaten our ordinary knowledge claims, why do we find them troubling in the first place? Cohen and DeRose suggest that the appearance of trouble arises from a kind of error. On Cohen’s view, by failing “to distinguish between the standards that apply in skeptical contexts, and the standards that apply in everyday contexts,” we are misled “into thinking that certain knowledge ascriptions conflict, when in fact they are compatible” (Cohen 1999, 77). DeRose says that “we fail to realize [ … ] that the skeptic’s present denials that we know various things are perfectly compatible with our ordinary claims to know those very propositions” (DeRose 1999, sec. 6; 1995, sec. 14; 2004, sec. 8).

Schiffer 1996 argues that this “error theory” is implausible: we shouldn’t expect to be “bamboozled by our own words” in the way that it requires. If, because “knows” is context sensitive, knowledge denials made in skeptic-friendly contexts are consistent with knowledge attributions made in ordinary contexts, then we ought to recognize, more or less immediately upon careful reflection, that sound instances of AI leave our ordinary knowledge claims unscathed. Hofweber and Rysiew refine the objection. Hofweber allows that while there is plenty of what he calls “hidden relativity” in language, the contextualist error theory implies that “sameness, difference, and incompatibility of contents” is inaccessible to speakers. But, Hofweber thinks, it is implausible to posit such inaccessibility (Hofweber 1999). Rysiew says that while speakers may be blind to what proposition a sentence literally expresses, they are rarely wrong about what they mean by uttering a sentence. But, he argues, the contextualist error theory requires this second claim (Rysiew 2001, 482–485).

Cohen responds that our ignorance of the context-sensitivity of other terms leads to analogous puzzles (Cohen 1999, 78–79; 2001, 90–91; 2005, 61). For example, a “flatness skeptic” argues that, owing to the ubiquity of tiny bumps, no actual surfaces are flat (Unger 1975, 54–62). The intuitive force of that argument is dissolved once we realize that flatness ascriptions are context sensitive. However, while it is easy to see that flatness denials made in contexts where even tiny bumps are important are compatible with flatness attributions made in ordinary contexts, the puzzle posed by AI is not readily dissolved by accepting that knowledge claims are context-sensitive. Cohen suggests that this difference in the puzzle-dissolving power of the contextualist approach in each case may be due to the fact that knowledge attributions have a “normative force” that flatness ascriptions lack (2005, 62). McKenna 2014 develops a position that makes Cohen’s suggestion more substantive. McKenna says that knowledge attributions have an important normative function; as a result, an ordinary knowledge attribution and the corresponding skeptical denial reflect a conflict that is not ameliorated by their truth-conditional compatibility. Such a residue of non-truth-conditional disagreement may obscure the fact that the truth-conditional content of the skeptic’s claim is consistent with that of our ordinary knowledge attributions.

DeRose says that when ordinary speakers are asked whether the conclusion of a skeptic’s argument contradicts an ordinary knowledge attribution, some will say “yes” and some will say “no” (DeRose 2006, 333–334; DeRose 2009, 177–179). “Apparently,” he observes, “that’s just a tough question.” Given that there is disagreement among ordinary speakers about this issue, some must be mistaken. If the contextualist treatment of AI is independently plausible, then, it is reasonable to regard those who think that skeptical knowledge denials conflict with ordinary knowledge attributions as the ones in error.

Whether or not the contextualist error theory successfully explains why we find AI troubling in the first place, it is distinct from the contextualist treatment of AI itself. The latter is valuable independently of the former. Contextualism implies that compelling knowledge denials made in contexts in which skeptical hypotheses are in play are consistent with ordinary knowledge attributions. Contextualism may neither explain nor alleviate the concerns of an epistemologist who is keen to show that we know by even very demanding standards that there is an external world, but it does tell us that her failure will not undermine our ordinary knowledge claims. Thus contextualism provides a bulwark against the spread of skeptical anxieties, regardless of whether the error-theoretic hypothesis about the root of those anxieties is correct.

Insofar as contextualism enables us to make progress on solving the skeptical problem, it is worth taking seriously. But contextualism is also independently plausible. As contextualists have argued, our ordinary knowledge talk appears to show that the truth-conditions of knowledge claims vary with the context in which they are made.

At the heart of the case for contextualism are pairs of imagined conversations concerning mundane concerns such as we encounter in everyday life. In one (the Low case), a speaker utters an instance of “S knows that P”; in the other (the High case), a speaker utters the corresponding instance of “S does not know that P.” In Low, it isn’t a matter of much practical significance whether P is true or false, and no unusual error possibilities are salient. In High, the practical stakes are high, and unusual error possibilities are salient. Considered from the perspectives of the conversations in which they are made, the attribution in Low and the denial in High are both natural and intuitively true. Contextualism accommodates and explains the truth of both “surface-contradictory” claims. Invariantism does neither, at least not straightforwardly. Thus Low-High pairs indicate that contextualism provides a better account of our ordinary knowledge talk than does invariantism.

The best-known Low-High pairs are DeRose’s Bank Cases (1992, 913) and Cohen’s Airport Cases (1999, 59). DeRose’s more recent Thelma and Louise Cases (2009, 4–5) are designed to repair a dialectical shortcoming of the earlier Bank Cases (section 7.3):

Thelma, Louise, and Lena are co-workers; their colleague John was in his office all day with his the door closed. All three saw John’s hat in the hallway and overheard a conversation whose participants presupposed that John was in his office. On these grounds they believe that John was in, though they did not see him. At the end of the day, Louise and Lena each head home, while Thelma stops at the local tavern to settle a small bet concerning whether John would be in that day. After her tavern-mates pay up, they ask her whether Lena knows that John was in, since she also had a small bet going on the question. Yes, Thelma answers:

(L) Lena knows that John was in.

Meanwhile, Louise is stopped by the police on her way home. They are investigating a serious crime, and need to verify whether John was at work today. They have no reason to doubt that he was, but need Louise’s testimony. She demurs, describing her evidence and belief that John was in, but pointing out that he may have left his hat on the hook the previous day, and that her co-workers who thought he was in may have been mistaken. After all, she points out, she didn’t actually see him. They follow up by asking whether Lena could testify to John’s whereabouts. No, Louise answers; she didn’t see him either:

(H) Lena doesn’t know that John was in.

Considered with respect to the contexts in which they are made, both Thelma’s utterance of (L) and Louise’s utterance of (H) are natural and intuitively true. But their utterances are simultaneous, and concern the same subject (Lena) and the same fact (that John was in). Contextualism accommodates the intuitive truth of both claims, and explains why Thelma’s knowledge attribution and Louise’s knowledge denial are both true. Invariantists cannot regard both as true. Thus contextualism reflects our ordinary knowledge talk better than invariantism does. Prominent non-contextualist strategies for accommodating the data from Low-High pairs will be addressed in sections 6.1, 6.2, 6.3, and 6.4.

One response to the argument from Low-High pairs is to hold that what varies between the cases are not the truth-conditions of the knowledge claims, but their assertability conditions. The epistemic standards for warranted assertion appear to vary with context; this may explain why both claims are natural and intuitively true (this response is related to the invariantist strategy discussed in section 6.2).

The warranted assertability response leads to a distinct argument for contextualism (Hambourger 1987, DeRose 2002, DeRose 2009, ch. 3). What epistemic position warrants assertion? According to a popular account, a speaker’s epistemic position is sufficient for making a warranted assertion of P just in case she knows that P (Unger 1975, ch. 6; Williamson 2000, ch. 11). A contextualist will rephrase the account along the following lines: a speaker’s epistemic position is sufficient for making a warranted assertion of P in context C just in case it would be true in C to attribute knowledge to her that P (DeRose 2009, 98–102). Invariantists should not have qualms with the rephrasing, since on their view the relativization to context is superfluous. But given that the epistemic position needed to make a warranted assertion varies with context, the link between assertability and knowledge implies that the epistemic position required for the truth of a knowledge attribution varies with context as well.

Blackson 2004 objects that the argument, even if sound, does not establish contextualism. It establishes rather that the epistemic position required for a subject to truly claim that she knows something can vary with features of the context in which she finds herself. But this variability can also be accommodated by subject-sensitive invariantism (section 6.3). DeRose agrees with Blackson’s assessment, but thinks contextualism is superior to subject-sensitive invariantism on independent grounds (DeRose 2009, 108).

A number of theorists deny that knowledge is necessary for assertability (e.g., Douven 2006, Kvanvig 2009, Lackey 2007, and Weiner 2005); others deny that it is sufficient (e.g., Brown 2010 and Lackey 2011). If knowledge and assertability are delinked, the argument is undermined. The claim that the epistemic requirement for warranted assertion varies with context has also been challenged. Turri 2010 proposes that the type of speech act one performs by making an assertoric utterance varies with context; he calls this view speech-act contextualism. Speakers in epistemically demanding contexts do not (merely) assert, but guarantee; the epistemic requirements for warranted guaranteeing are more stringent.

Contextualism is a broad tent. In this section, I’ll survey two intra-contextualist debates about how best to develop the view. The first concerns the contents of knowledge claims. What, more specifically, do knowledge claims mean, and how does this meaning fluctuate with the epistemically relevant features of the context of utterance? Contextualists have proposed a wide variety of answers, and the differences among them reflect differences among theories about the nature of knowledge itself. The second concerns the linguistic mechanism that accounts for the context-sensitivity of knowledge claims. Contextualism has traditionally been treated as the view that the proposition expressed by a knowledge sentence varies with the context of utterance: that “know” is an indexical term. But recent authors have proposed different theoretical frameworks for understanding the relationship between the context sensitivity of knowledge claims and the semantics of knowledge sentences.

The dominant framework for contextualist thinking about the contents of knowledge claims has been the relevant alternatives approach to knowledge, on which knowing that P involves being in a position to rule out relevant alternatives to P. If we understand “relevant” to mean “relevant to those attributing knowledge,” then the relevant alternatives approach naturally leads to a version of contextualism:

(RA-C) “S knows that P” is true with respect to C only if S can rule out the alternatives to P relevant in C.

On the RA-C picture, the epistemic standard in a context is understood in terms of the set of alternatives relevant in that context. Such a proposal is at the heart of many classic and more recent contextualist theories, including Stine 1976, Cohen 1988, Lewis 1996, Heller 1999a, Blome-Tillman 2009, Ichikawa 2011, and McKenna 2013. Opinions among these authors vary widely about how to characterize relevance, the relationship between relevance, psychological salience, and practical circumstances, and what it is to “rule out” an alternative.

Another influential approach echoes the theory of knowledge defended by Nozick 1981, who held that knowledge requires sensitive belief; that is, a belief that would not be held if it were false. DeRose recasts the sensitivity requirement as a “conversational rule” that fixes the epistemic standards governing knowledge attributions in a context. His “Rule of Sensitivity” states that when a speaker attributes knowledge that P to S, the epistemic standards should be set as high as they need to be so that the knowledge attribution is true only if S’s belief that P is sensitive (DeRose 1995, sec. 12). DeRose argues that the Rule of Sensitivity underwrites a contextualist treatment of AI superior to various RA-based treatments.

Contextualists of a more “internalist” bent understand the variation in standards in terms of the type or degree of justification, evidence, or reasons a subject must possess. Though originally couched in the RA framework, Cohen’s view is an example of this approach. He says that what varies is “how justified a belief must be” or “how good one’s reasons have to be” (Cohen 1999, 59–60; Cohen 1988, sections IV and V). Neta’s version of contextualism also starts from an internalist picture of knowledge as involving evidential justification, but on Neta’s view, what varies with context are which states count as evidence; the context-sensitivity of knowledge attributions is parasitic on the context-sensitivity of evidence attributions (Neta 2004).

Not all versions of contextualism are best understood in terms of higher and lower standards. On Schaffer’s “contrastivist” theory, knowledge is a three-place relation between a subject, a proposition, and a contrast proposition (Schaffer 2004a). For the contrastivist one never knows that P simpliciter; one always knows that P rather than Q. As a result, the truth-conditions of “binary” knowledge attributions vary with the contrast proposition supplied by context: when Q is the salient contrast to P, an utterance of “S knows that P” is true iff the subject knows P rather than Q; but when R is the salient contrast, an utterance of “S knows that P” is true iff the subject knows P rather than R. For contrastivist contextualists, what shifts are contrasts, not epistemic standards. (Schaffer himself says that contrastivism is a competitor to, rather than a version of, contextualism (2004a, sec. 1), but the view he characterizes as “contextualism” involves a number of commitments inessential to contextualism as understood in the present essay.)

Another family of contextualist theories starts with the idea that knowledge involves causal-explanatory relations. Greco 2010 argues that knowledge is a kind of “success through ability.” This means that a subject’s true belief that P is “creditable” to her intellectual abilities, which implies that her abilities causally explain her success in believing truly. Since causal explanations are context-sensitive, “knowledge attributions inherit the context-sensitivity of causal explanations” (Greco 2010, 106). Rieber 1998 says that when a subject knows that P, the fact that P explains why she believes that P. But whether A counts as “the explanation” of B varies with what contrast to B is implicit in the context, and so the claim that one thing explains another is context-sensitive. Thus “the context-dependency of explanation can account for the context-dependency of knowledge” (Rieber 1998, 194).

Traditional contextualists hold that the proposition expressed by a knowledge sentence varies with the context of utterance. Cohen says that he “construes ‘knowledge’ as an indexical” (Cohen 1988, 97), and DeRose encourages thinking of contextualism in terms of Kaplan 1989’s distinction between character and content, a central element of his hugely influential theory of indexicals (DeRose 1999, 920; 2009, 3). Indexical terms encode different contents in different contexts of utterance; as a result, sentences containing indexicals express different propositions relative to different contexts of utterance. Following MacFarlane 2009, I’ll call the proposal that “know” functions semantically like an indexical, indexical contextualism. If the content “know” contributes to the contents of sentences that contain it varies with the context of utterance, then the propositions expressed by those sentences vary as well. This explains why the truth-conditions of knowledge claims vary with context.

There are various ways to develop indexical contextualism. One may treat “know” as what Schiffer 1996 calls an “indexical verb”: on this picture, the semantic content of “know” is simply the knowledge relation salient in the context, just as (for example) the semantic content of “I” is simply the speaker of the context. Instead, indexical contextualists may assign “know” a more complex semantic content with a typically unpronounced element that accounts for its indexicality. For example, “S knows that P” may encode the content that S knows that P according to N, where N denotes the epistemic standard of the context. Kompa 2002, Brogaard 2008, and MacFarlane 2009 develop an alternative that Brogaard calls perspectivalism and MacFarlane calls nonindexical contextualism. Suppose that the sentence “Obama was born in Hawaii” encodes the proposition that Obama was born in Hawaii. That proposition is true at the actual world but false at a possible world where Obama was born somewhere else, or never at all: though it contains no indexicals, its truth-value varies with the possible world at which it is being evaluated. Some theorists think that a proposition’s truth-value can also vary with the time of evaluation; for example, that Obama was born in Hawaii may be true at a 2014 circumstance but not true at a 1914 circumstance. Nonindexical contextualists claim that the truth-value of a proposition varies with the epistemic standard at which it’s evaluated. If so, the proposition that Descartes knew that Galileo was condemned may be true at a low epistemic standard but false at a high epistemic standard. Assuming that the epistemic standard relevant to the truth-value of a knowledge claim is that of the context of utterance, nonindexical contextualism implies that the truth-conditions of a knowledge claim may vary with the context of utterance even if “know” is not indexical.

Brogaard and MacFarlane argue that nonindexical contextualism can resist certain influential objections to contextualism discussed later in this article (sections 6.2 and 6.3). Non-indexical contextualism bears a strong resemblance to MacFarlane’s influential version of relativism (section 7.4); the primary difference with MacFarlane’s relativism concerns which context sets the epistemic standard relevant to a knowledge claim’s truth-value.

A third approach distinguishes sharply between the contents encoded by knowledge sentences and the propositions asserted by speakers who make knowledge claims. What a speaker asserts by uttering a sentence is not always the content encoded by the uttered sentence with respect to the context of utterance. For example, consider Bach 1994’s example of a mother who utters “You’re not going to die” to comfort her son who’s gotten a minor scrape. The proposition encoded by that sentence is true only if her son is immortal, but she has asserted merely that he is not going to die from this scrape. When by uttering s in C a speaker asserts a proposition not encoded by s with respect to C, her assertion is pragmatically enriched (Recanati 2012). Pragmatic contextualism says that the epistemic standards in a context affect how knowledge claims made in that context are pragmatically enriched. Thus, even if the contents of knowledge sentences are invariant, the truth-conditions of knowledge claims vary with the epistemic standards of the context.

Rysiew thinks that the speakers in Low and High “say something true,” not because of what is “semantically encoded or expressed by the sentence itself,” but rather because of the “conceptually and pragmatically enriched proposition[s]” conveyed by their utterances (Rysiew 2001, 486, emphasis in original; Rysiew 2007 develops the proposal further). Cappelen 2005 argues that, while instances of “S knows that P” encode an invariant content that is typically false, speakers who utter “S knows that P” typically assert distinct truths. Both Rysiew and Cappelen present themselves as opponents of contextualism, but indexical contextualism is their real target; their views are versions of pragmatic contextualism. Harman 2007, in a similar vein, treats contextualism as a “theory of primary speaker meaning.” And Stainton 2010 defends what he calls “speech-act contextualism” (not to be confused with Turri’s view of the same name discussed in section 3.2) by appeal to the idea that pragmatic factors affect the literal contents of an assertion; Stainton argues that the speech-act contextualist is invulnerable to linguistic objections to indexical contextualism (in particular those discussed in sections 6.1 and 6.2). And Pynn 2015 defends a version of pragmatic contextualism on which High context knowledge denials are instances of lexical narrowing.

Since it first appeared on the scene, contextualism has faced an array of objections. The first wave of criticism focused primarily on the adequacy of the contextualist treatment of skeptical arguments (sections 2.2 and 2.3). The second wave, which I’ll focus on in this section, targeted contextualism’s plausibility as an empirical linguistic thesis. These objections have led many to become wary of endorsing contextualism. In this section I will familiarize the reader with some of the most prominent such objections, most of which interact with complex issues in the philosophy of language, and put forward a piecemeal argument that the objections are not compelling. Contextualists have responded to each in careful detail, and in each case the objection turns out to depend upon observations or theoretical commitments that contextualists may freely deny. I’ll first discuss in some detail three influential “linguistic” objections to contextualism (sections 5.1–5.3), briefly describe three more (section 5.4), and close with an overview of some recent work on knowledge claims in experimental philosophy (section 5.5).

Cohen 1999 defends contextualism in part by appeal to an analogy between “know” and gradable adjectives such as “flat,” “bald,” “rich,” “happy,” and “sad.” Such terms are semantically linked to scales; the degree on the scale specified by a particular use of the term varies with context. Similarly, Cohen suggests, the strength of epistemic position specified by a particular use of “know” varies with context. Stanley presents a number of considerations to undermine the analogy between “know” and gradable terms. For example, unlike gradable terms, “know” does not accept modifiers such as “very” or “really”; nor is it associated with comparative constructions like “happier than.” So, Stanley concludes, we should be “at the very least suspicious” of the claim that “know” is semantically linked to a scale of epistemic position (Stanley 2004, 130; Stanley 2005, ch. 2).

However, Stanley’s view that terms that do not pass his tests for gradability are not linked to scales has been challenged. Halliday 2007 points out that while “sufficiently tall” is both context sensitive and linked to a scale, it fails Stanley’s tests for gradability. Blome-Tillman 2008 says that whether a noise is loud enough to count as a snore varies with context, suggesting a semantic link between “snore” and a scale of loudness. Yet “snore” fails Stanley’s tests as badly as “know” does. Moreover, just as “snore” accepts adverbial modifiers that denote degrees of loudness (e.g., “very / quite / extremelyloudly”), “know” accepts modifiers that denote degrees of epistemic position (e.g., “with very / quite / extremely good evidence / justification”; Ludlow 2005, 19–21 provides a long list of other such modifiers).

Moreover, as Stanley acknowledges, the thesis that “know” is not semantically linked to a scale of epistemic position is entirely consistent with contextualism; there are other forms of context-sensitivity than those associated with gradability (Partee 2004). The gradability objection is part of Stanley’s “larger inductive argument” against contextualism (Stanley 2005, 52). The challenge for the contextualist who accepts the gradability objection is to find another family of context-sensitive terms that can serve as a plausible model for the semantics of “know.” Stanley thinks that other problems indicate that none of the candidates is plausible. As we’ll see in the rest of this section, however, the other linguistic objections to contextualism are less than overwhelming.

Finally, the gradability objection appears to have no force against nonindexical and pragmatic contextualism. Nonindexical contextualists treat the semantics of “know” in a way quite unlike the standard treatment of gradable adjectives. And Stainton argues that, due to pragmatic enrichment, the claims we make using terms that are not gradable can themselves be “subject to degrees” (Stainton 2010, 128).

When someone utters a sentence s, you can often report her utterance by saying, “He said that s”; that is, by using a disquotational indirect speech report. However, when s contains an indexical term like “I,” “here,” and “now,” disquotational reports of utterances of that sentence break down. For example, if Julie uttered “I’m hungry” and you’re not Julie, you can’t report Julie’s utterance by saying, “Julie said that I’m hungry.” Some opponents of contextualism claim that disquotational reports of knowledge sentences are accurate across contexts with different epistemic standards (Cappelen and Lepore 2005, ch. 7; Stanley 2005, ch. 3). Hawthorne makes an analogous claim about disquotational indirect belief reports (Hawthorne 2004, 98–111). These observations undermine the analogy between “know” on the one hand and “I,” “here,” and “now” on the other, and so present a challenge for the indexical contextualist.

Proponents of the objection do not generally argue for the claim that the relevant disquotational reports are unproblematic. Cappelen and Lepore, for example, list it among their “reports of [their] own intuitions” (Cappelen and Lepore 2005, 94). Not all share the intuition. DeRose, for example, disagrees with Hawthorne that the relevant indirect belief reports are unproblematic (DeRose 2009, 162–166; DeRose 2006). When we focus on cases like those involving Thelma and Louise, DeRose says, we’ll find that they are “far from being clearly correct” and even, in DeRose’s judgment, “at least somewhat clearly wrong.” Analogous points apply to disquotational speech reports involving “know” as well.

Even if some cross-contextual disquotational reports involving “know” are wholly unproblematic, this is troubling only if all context-sensitive terms lead to breakdowns of disquotational reports in a like manner. But as DeRose 2006 (337–331; DeRose 2009, 166–174) points out, there are no relevant differences between the behavior of “tall” and “know” in indirect belief reports; Blome-Tillman 2008 (36–38) makes a similar observation about the behavior of “empty” and “flat” in indirect speech reports. Yet “empty,” “flat,” and “tall” are widely held to be context sensitive. Cappelen and Hawthorne 2009 argue that the fact that a term does not block disquotational reports provides little evidence that it is not context sensitive. They suggest that felicitous cross-contextual disquotational reports involving context-sensitive terms often involve “parasitic context-sensitivity”: the values of the context-sensitive terms in the report are “parasitically determined” by the context of the reported utterance (Cappelen and Hawthorne 2009, 40–42). If the suggestion is right, then felicitous cross-contextual reports of knowledge claims do not undermine contextualism.

Finally, the objection has no force against forms of contextualism that reject the thesis that “know” functions semantically like an indexical. MacFarlane argues that nonindexical contextualism is immune: since the content of a knowledge sentence is stable across contexts, we should expect disquotational reports of knowledge claims to be accurate across contexts (MacFarlane 2009, 239–241). And Stainton suggests that the pragmatic contextualist can resist the objection as well (Stainton 2010, 127–128). Pynn 2015 (sec. 5) develops Stainton’s suggestion in detail, showing that the pragmatic contextualist has nothing to fear from the behavior of “know” in disquotational reports.

The argument from Low-High pairs assumes that the correct semantic theory of knowledge claims should comport with our intuitions about the truth and falsity of those claims. Unlike invariantism, contextualism respects the judgments of speakers in Low and High cases about the truth-values of the knowledge claims they themselves make. But according to a number of theorists, contextualism does not respect the judgments of ordinary speakers about the truth and falsity of knowledge claims made in contexts other than their own. Consider a knowledge attribution made in a Low context and the corresponding knowledge denial made in a High context. Participants in High will (the objection goes) not only regard their own knowledge denial as true, but also judge the attribution made in Low to be inconsistent with their knowledge denial. Contextualism implies that the attribution and denial are consistent; indeed, that both are true. The alleged failure of contextualism to comport with our cross-contextual judgments about knowledge claims undermines its claim to respect ordinary speaker intuitions.

The primary evidence cited in favor of the objection rests on observations concerning when expressions of disagreement and retraction are natural. For example, consider the following exchange imagined by Williamson:

John:

I know that this is a zebra.

Mary:

How do you know that it isn’t a mule cleverly painted to look like a zebra?

John:

Hmm, for all I know it is a painted mule. So I was wrong. I don’t know that it is a zebra after all (Williamson 2005, 220; see also Richard 2004; Stanley 2005, 52; MacFarlane 2005, 202–203; Chrisman 2007, 228–229; Brogaard 2008, 411).

John’s use of “I was wrong” and “after all” in response to Mary’s question indicate that he disagrees with and intends to retract his earlier attribution.

But cases like Williamson’s only have force against a view that implies that John’s attribution and subsequent denial are consistent. And while contextualists could regard John’s claims as consistent, contextualism does not imply that they are. On the contrary, since contextualists have no wish to erase genuine disputes about knowledge, they have reason to treat John’s claims as inconsistent. The contextualist may deny that Mary’s intervention has shifted the epistemic standard. Montminy points out that unless we know more about the “presuppositions, purposes, intentions, etc. of the conversational participants, it is unclear whether the epistemic standards associated with John’s knowledge attribution are different from those associated with his knowledge denial” (Montminy 2009a, 642). Rather than shifting the standards, Mary may have made explicit a possibility that John should have been attending to in the first place, or she may have pressed John to consider a possibility he is free to ignore, and so led him astray. Alternatively, the contextualist may accept that Mary’s intervention has shifted the standard but still regard the claims as inconsistent. This is DeRose’s treatment. He gives an account of cases like Williamson’s on which the latter claim conflicts with the former even given a shift in standards; see his discussion of the “assymetrical gap view” of what he calls “one-way disagreements” (DeRose 2009, 150–151).

Still, while Williamson’s case needn’t make trouble for the contextualist, she must hold that some “surface-contradictory” pairs of knowledge claims are consistent; this is, after all, the basis on which contextualism is claimed to capture a key feature of our ordinary knowledge talk. But DeRose points out that when we focus on “properly constructed cases”—that is, the sorts of cases best suited to underwrite the argument from Low-High pairs—it is far from obvious that our cross-contextual judgments are at odds with contextualism (DeRose 2009, 160–161). Suppose that the police, in their conversation with Louise, played her a tape of Thelma’s knowledge attribution. Provided that Louise is aware of the casual nature of Thelma’s tavern conversation and carefully attending to the contrast with her own very serious situation, it is odd to think that she would say, “I disagree with Thelma; she is wrong.” It would be at least as natural for her to say, “Thelma was only speaking loosely; she wouldn’t say that if she were here right now,” or, “She didn’t mean that Lena knows.” Unless a firm expression of disagreement would be natural, the case provides little grist for the anti-contextualist mill.

Finally, even if there are clear examples of a tension between contextualism and our intuitive cross-contextual judgments, this is hardly a decisive objection. McKenna 2014 argues that the normative function of knowledge claims gives rise to a form of cross-contextual disagreement that can persist even when the relevant claims are both true. Cross-contextual judgments of inconsistency may reflect the “conflicting recommendations” of the speakers in each context. Brogaard 2008 (453–455) argues that given nonindexical contextualism, a High speaker can truly judge that the Low knowledge attribution was false (though see MacFarlane 2014, 190–192, who argues that while nonindexical contextualism handles the disagreement data better than indexical contextualism, it fails to explain why speakers should be led to retract). Moreover, the tension between invariantism and our intra-contextual judgments is rather more stark than that between contextualism and any judgments of cross-contextual inconsistency: the invariantist must regard one of the speakers (as well as those of us considering her claim) as mistaken about the truth-conditions of a claim she herself makes, and not merely those of a claim made in a distinct context. (Here relativists claim a distinct advantage over both contextualism and invariantism; see section 6.4.) It is more charitable to speakers to attribute to them an error about the truth-conditions of a claim made in a context they are considering at a distance than it is to attribute to them an error about the truth-conditions of their own claims.

A number of other claims to the effect that our linguistic behavior is not as we should expect it to be given contextualism have been discussed in the literature:

Propositional Anaphora. Stanley 2005, 52–53 says that various uses of the locution “what I said” that should be true given contextualism are infelicitous. Suppose it is raining outside and I say, “It is raining here.” Moments later when I move inside, I can say, “What I said before is true, but it’s not raining here.” Could Louise felicitously say of Thelma, “What she said is true, but Lena doesn’t know that John was in”? It seems not; Stanley claims that this undermines the claim that “know” is context sensitive.

Intra-Discourse Switching. Stanley 2005, 57 says that, for a broad range of context-sensitive terms, “multiple occurrences of that element in a discourse should be able to take on differing values.” For example, one can felicitously say, “That butterfly is large, but that elephant isn’t large” even when the elephant is much larger than the butterfly. Yet felicitous readings of “Sam knows that his car is in the driveway, but he doesn’t know that it hasn’t been stolen” are, according to Stanley, unavailable.

Clarification Techniques. Hawthorne 2004, 104–106 says that, for most context sensitive terms, various clarification techniques are available for elaborating what we meant by using the term when challenged. For example, if I say, “The glass is empty” and you challenge me by pointing out that it has air in it, I can respond by clarifying: “All I was meant was that it was empty of vodka.” But, Hawthorne claims, we have very few such techniques for clarifying what’s meant by a “know”-involving utterance in response to a challenge.

In each case, the contextualist can either dispute the objector’s data or resist the anti-contextualist inference from the data. As an example of the first kind of response, consider Ludlow 2005’s catalogue of natural techniques culled from Google searches for clarifying knowledge claims; Ludlow’s many examples stand in prima facie tension with Hawthorne’s observation. As an example of the second kind of response, note that, since “S knows that P” is true only if P is true, the contextualized version of the knowledge norm of assertion (section 3.2) implies that a High context speaker should not assert that a Low context knowledge attribution was true. This explains why a Louise can’t felicitously utter, “What Thelma said was true, but Lena doesn’t know that John was in,” even if such an utterance would (as contextualism is held to predict) be true. For a different and contextualist-friendly approach to the data that underwrite the intra-discourse switching objection see DeRose 2008, sec. 6.

Recent work in “experimental philosophy” has focused on the attitudes of experimental subjects towards cases like those involved in the argument from Low-High pairs. Some of the results have been taken to undermine the case for contextualism. For example, Schaffer and Knobe, surveying the data, say that “the results suggest that people simply do not have the intuitions they were purported to have” by contextualists and their opponents, suggesting that “the whole contextualism debate was founded on a myth” (Schaffer and Knobe 2012, 675). If correct, this observation seriously threatens the argument from Low-High pairs, and the truth of contextualism itself.

However, more recent experimental data supports, rather than undermines, the argument from Low-High pairs. That argument makes an abductive inference to contextualism from a claim about the intuitive truth of knowledge attributions and denials:

(T-T) The knowledge attributions made by speakers in Low context vignettes are natural and intuitively true, and the corresponding knowledge denials made by speakers in High context vignettes are also natural and intuitively true.

Three studies have tested (T-T) directly. Feltz and Zarpentine 2010 presented subjects with Low and High vignettes modeled on Stanley’s versions of the Bank Cases and asked them to agree or disagree that “What [the speaker] said was true.” They found that subjects tended to be neutral; that is, that they tended neither to agree nor disagree. While this result does not help the case for contextualism, it does not clearly undermine it, either (DeRose 2011, 91–94). More recent work has been more decisive. Hansen and Chemla “confirmed DeRose’s prediction that speakers would find both ‘I know that p’ in the Low context and ‘I don’t know that p’ in the High context true” (Hansen and Chemla 2013, 203). And Buckwalter 2014 also confirmed (T-T) using surveys based on the Bank Cases. Buckwalter concludes that “contextualism has been shown to be compatible with the relevant knowledge behaviors” of ordinary subjects.¹

The bulk of the experimental work on knowledge claims and Low-High pairs concerns the knowledge claims that experimental subjects themselves make concerning those characters. A central interest has been to identify what factors can produce a shift in judgments about whether a character knows. Some argue that manipulation of the perceived stakes of the character’s situation is responsible for the shift (Pinillos 2012, Pinillos and Simpson 2014, and Sripada and Stanley 2012), while others maintain that shifts are brought about by the salience of error possibilities (Buckwalter 2014, Buckwalter and Schaffer 2013, May et al. 2010, and Schaffer and Knobe 2012). This debate is relevant to a number of issues of interest to contextualists. For example, it should help guide our understanding of how and why the truth-conditions of knowledge claims shift (section 4.1). And while contextualism per se makes no prediction about what factors produce the shift, the outcome of the debate may indirectly impact contextualism’s plausibility. For example, evidence for the stakes hypothesis bolsters the case for subject-sensitive invariantism, an alternative to contextualism (section 6.3), and Pinillos 2012 argues that the results of his interesting “evidence-seeking” experiments favor subject-sensitive invariantism over contextualism. So contextualists cannot afford to be insouciant about this growing literature.

The traditional competitor to contextualism is invariantism; the primary challenge for the invariantist is to respond to the argument from Low-High pairs. In this section I’ll describe three invariantist strategies for accounting for the Low-High data, and close by discussing the relativist alternative.

The invariantist may attempt to explain the variability exhibited in Low-High pairs in terms of the conversational propriety or communicative effects of knowledge claims. An assertion whose content is true may nonetheless be misleading or otherwise improper, while a false assertion may be informative and proper. Invariantists who attempt to explain the divergence in intuitive truth-conditions exhibited in Low-High pairs in such terms offer pragmatic accounts of the data. In the literature such accounts are often referred to as examples of what DeRose 2002 calls a warranted assertability maneuver, or WAM (section 3.2). Black 2005, Brown 2006, Hazlett 2007, Pritchard 2010, Rysiew 2001, Rysiew 2007, and Schaffer 2004b are prominent examples.

Some attempt to account for the apparent truth of one of the knowledge claims in a Low-High pair in terms of Grice 1989’s notion of conversational implicature. Hazlett, for example, says that since a High context speaker who made a true knowledge attribution would falsely implicate that the subject can eliminate contextually salient error possibilities, High context speakers instead assert false knowledge denials in order to communicate truly that the subject can’t eliminate such possibilities (Hazlett 2007, 682–687; Brown 2006, 419–428 and Black 2005). One common objection to implicature-based accounts is that in canonical examples of implicature we seem capable of distinguishing the falsehood that we assert from the truth we thereby implicate, yet the knowledge claims in Low-High pairs seem true, not false but informative.

Rysiew’s proposal relies not on implicature but on “earlier” pragmatic effects such as Recanati 1989’s strengthening and Bach 1994’s impliciture (Rysiew 2007, 641–648, esp. footnote 31; DeRose 2009, 118–124 responds to Rysiew’s proposal on behalf of the contextualist). The key difference, for our purposes, between these effects and implicature is that they do not depend upon the speaker’s intending to communicate the literal content of the uttered sentence. Rysiew’s finely crafted view is clearly opposed to indexical contextualism, but since he agrees that in one sense “what is said” by speakers who make knowledge attributions varies with epistemically relevant features of the context of utterance, pragmatic contextualists should find it congenial, as well (Rysiew 2001, 485–487).

Davis 2007 pursues a different sort of pragmatic strategy. Davis endorses a form of skeptical invariantism, but treats knowledge attributions made in Low contexts as examples of loose use (Conee 2005, 52–53 and Fantl and McGrath 2012 also discuss this approach). Loose talk is ubiquitous. When there are only a few grounds of coffee left in the jar, it may be strictly false, but informative and appropriate, to say that there is no coffee left (e.g., when there is not enough coffee left to brew a full pot). Similarly, Davis suggests, in Low contexts “know” is used loosely; the knowledge attribution made in Low is close enough to being true for practical purposes. DeRose 2012 agrees with Davis that loose talk is ubiquitous, but thinks that such talk is often truthful; the standards of precision in a context, on DeRose’s suggestion, are truth-conditionally relevant. Given DeRose’s suggestion, Davis’s proposal yields a version of contextualism, not invariantism.

Focusing on unusual error possibilities or high stakes may lead you to modify your beliefs. Since knowledge requires belief, a subject who has lost her belief that P can truly deny that she knows that P. If a subject in a High context loses her belief that P, she can truly deny that she knows that P, even if her epistemic position is sufficient for knowing P. Bach says that, in a High context, the speaker’s “threshold for (confidently) believing” goes up, so that she “demands more evidence than knowledge requires” before she is willing to believe the relevant proposition (Bach 2005, 77). In a series of rich and empirically well-informed papers, Jennifer Nagel develops a sophisticated account along these lines (Nagel, 2008  2010a, 2010b). Nagel says that subjects in high-stakes situations experience a high “need for closure”; that is, they require more information before forming settled beliefs than do subjects in low-stakes situations (Nagel 2008). She calls the force underlying this phenomenon “epistemic anxiety” (Nagel 2010a). Experiencing heightened epistemic anxiety, speakers in High contexts refrain from forming settled beliefs and so truly deny that they know.

One objection to the Bach-Nagel strategy is that we seem able to suppose that the speakers in High contexts have settled beliefs without making their knowledge denials any less natural or intuitively true; indeed, descriptions of High cases can include a stipulation that the speakers do have the relevant belief (though, see Nagel 2008, 290–293 for doubts about the effectiveness of such stipulations). Additionally, the strategy does not account directly for the intuitive truth of third-person knowledge denials like Louise’s. Since Lena is not part of the conversation with the police, the high stakes haven’t led her to lose her belief. Bach and Nagel both explain third-person High context knowledge denials in terms of an error on the part of the High context speaker. Nagel 2010b argues that due to a bias called “epistemic egoism,” High context speakers project their own epistemic anxiety on to third parties; Bach 2005, p. 76–77 provides a similar explanation.

Williamson offers a different psychological explanation for High context knowledge denials. He says that when we focus on unusual possibilities of error, we tend to give them undue weight in our judgments about knowledge, and so experience an “illusion of epistemic danger” (Williamson 2005, 225–226). High context subjects thus mistakenly judge that their epistemic position is insufficient for knowledge. Following Vogel 1990 and Hawthorne 2004, Williamson suggests that Kahneman and Tversky’s “availability heuristic” might explain the illusion. Nagel 2010b criticizes Wliiamson’s appeal to the availability heuristic, but the core of his suggestion—that speakers in High contexts are vulnerable to a kind of illusion—may be salvageable. One advantage of Williamson’s proposal over the Bach-Nagel strategy is that it does not require positing a distinct factor to explain the apparent truth of a third-person denial: the illusion of epistemic danger should affect first-person and third-person denials equally. Pynn 2014 criticizes Williamson’s illusion solution but argues that our tacit acceptance of the principle that those who know are in a position to assert leads High context speakers to make a similar error. For another proposal in the Williamsonian vein, see Gerken 2013’s discussion of “epistemic focal bias.”

Traditional invariantists hold that, for any two subjects in the same epistemic position with respect to P, either both subjects are in a position to know that P, or neither are. This reflects a commitment to what Stanley 2005 calls intellectualism about knowledge (Fantl and McGrath 2009 call the same idea purism). Anti-intellectualists hold that whether a subject is in a position to know can depend on other aspects of her situation, and in particular on her practical circumstances: for example, when the costs to the subject of being wrong about P are high, a stronger epistemic position is required for her to know that P than when the costs are low. Combining invariantism with anti-intellectualism yields a view known as subject-sensitive invariantism (Hawthorne 2004; Stanley 2005 calls the view interest-relative invariantism).

Subject-sensitive invariantism fits nicely with the Low-High pairs originally used to motivate contextualism. For example, in DeRose 1992’s Bank Cases, the subject in both cases is Keith. In the Low case, Keith claims to know that the bank will be open on Saturday; in the High case, Keith denies that he knows that the bank will be open on Saturday. Subject-sensitive invariantists can respect the truth of both claims: the practical difference between Keith’s situation in Low and his situation in High means that Keith’s epistemic position is sufficient to know in the former but not in the latter.

However, subject-sensitive invariantism does not explain the Low-High pair discussed in section 2.1, since Lena’s practical situation is the same with respect to both cases. Subject-sensitive invariantists appeal to strategies like those discussed in 6.1 or 6.2 to explain third-person cases like those involving Lena. Stanley, for example, proposes that, in considering whether Lena knows, Louise is actually concerned with whether Lena would know if she were in her own (i.e., Louise’s) practical situation. And by the subject-sensitive invariantist’s lights, Lena wouldn’t know if she were in Louise’s situation; according to Stanley this provides “a perfectly intuitive explanation of the intuitions” about Louise’s knowledge denial (Stanley 2005, 102; Fantl and McGrath 2009, 56 suggest a similar explanation).

While subject-sensitive invariantism is a competitor to contextualism, the debate about intellectualism is orthogonal to the debate about contextualism. Anti-intellectualist contextualism is a coherent position (Fantl and McGrath 2009, 34–37). Nonetheless contextualists are generally intellectualists, and treat their ability to account for the sensitivity of knowledge attributions to practical concerns without rejecting intellectualism as an advantage of their view (DeRose 2009, 185–193).

We can distinguish the context in which a claim is made from the context in which the claim is assessed for truth or falsity. Relativists think that a claim’s truth-value can vary with the context in which it is assessed. Applied to knowledge attributions, relativism is usually understood to be the view that a given knowledge claim may be true relative to a context of assessment with a low epistemic standard, but false relative to a context of assessment with a high epistemic standard (Richard 2004, MacFarlane 2005, MacFarlane 2014, ch. 8; Rysiew 2011b provides a helpful overview of the relationship between contextualism and relativism). Relativists can readily accommodate the data from Low-High pairs: relative to a context of assessment with a low standard, Thelma’s knowledge attribution is true and Louise’s denial is false, whereas relative to a context of assessment with a high standard, Louise’s knowledge denial is true and Thelma’s knowledge attribution false.

Given relativism, an assessor in the High context should regard the attribution made in the Low context as false, since that attribution is false relative to the high standard in her context of assessment. For this reason, relativism’s proponents claim that it outperforms contextualism on the issues concerning disagreement and retraction canvassed in section 5.3 (MacFarlane 2005, 219; MacFarlane 2014, 188). As we saw in that section, however, the disagreement and retraction data do not underwrite a straightforward challenge to contextualism. But even if they do, relativism does not clearly fit those data much better. Consider Williamson’s case again. The relativist’s advantage over the contextualist is that she can respect John’s subsequent judgment that his earlier claim was false, and inconsistent with his knowledge denial. (As we saw, the contextualist can respect this judgment as well, but it takes some footwork; for the relativist John’s judgment of falsity is true simply in virtue of relativism itself.) Yet John’s attitude towards his earlier judgment is not merely that it had a false content, but that it was the result of an error. It would be natural for him to retract by saying, “I was mistaken—I shouldn’t have said that I knew.” But the relativist’s advantage over the invariantist consists in her holding that John’s initial attribution and subsequent denial exhibit “joint reflexive accuracy”: each is true as assessed from its context of utterance (MacFarlane 2014, 130). And John seems quite unprepared to accept that his earlier claim was true as assessed from any context. If the semantics of knowledge claims are relativistic, John seems not even to be tacitly aware that this is the case.

In a related vein, Montminy 2009b argues that the relativist must impute to ordinary speakers a kind of semantic error in their cross-contextual judgments not unlike that which contextualists are often thought by their opponents to require. MacFarlane 2014 (196–200) responds to Montminy’s argument. Stanley 2005, ch. 7 presents a number of theoretical objections to MacFarlane’s relativism, and Cappelen and Hawthorne 2009 give a more general critique of relativism in semantics. On the other side, MacFarlane 2014 develops a wide-ranging and finely honed case for the relativist program, offering detailed responses to many criticisms and application to a number of specific bits of language, including knowledge claims. Relativism is an important emerging paradigm in philosophical semantics, and future work may reveal distinctive advantages over the more well-entrenched contextualist and invariantist alternatives.