scc0905s1

ANTISKEPTICAL CONDITIONALS

Theodore J. Everett

SUNY-Geneseo

Empirical knowledge exists in the form of antiskeptical conditionals, which are propositions like [if I am not undetectably deceived, then I am holding a pen]. Such conditionals, despite their seeming triviality, have the same essential content as the categorical propositions that we usually talk about, and can serve the same functions in science and practical reasoning. In this paper, I sketch out two versions of a general response to empirical skepticism using these conditionals. For the first, I argue that even if our ordinary, categorical knowledge statements are all false as skeptics say, they can easily be replaced by statements using antiskeptical conditionals so as to get around the skeptical problem, given that such conditionals are all that we really think we know. For the second, I argue that the objects of our ordinary knowledge attributions can also be analyzed as antiskeptical conditionals, on the grounds that such conditionals are all that we really intend to convey in making ordinary knowledge claims. This would allow us to refute, not just evade, the most common sort of skeptical argument. I argue that both versions of the new theory compare favorably to the best-known current approaches to skepticism, matching most of their intuitive strengths while avoiding their main weaknesses. I try to show both versions to be plausible independently, too, in part by distinguishing levels of belief and meaning. At the end, I suggest the possibility that neither version of the theory is exclusively correct, and that the truth is sometimes indeterminate between the two.

A preliminary word on diction. I need a reasonably neutral way of referring to the antecedents of antiskeptical conditionals – that is, the denials of standard skeptical hypotheses – as things that underlie our ordinary statements and beliefs. "Assumption" and "presupposition" are the two obvious candidates, and as we often use these two words interchangeably, it should not matter very much which term I use. But the word "assumption" tends to suggest a mere hypothesis that one is conscious of as such, while "presupposition" implies a positive belief that one takes thoughtlessly for granted. With some regimentation of ordinary speech, we might even say that where A is a presupposition of B, an utterance of "B" really denotes a proposition of the form (A & B), and that where A is merely an assumption underlying B, an utterance of "B" denotes a proposition of the form (A ® B). Since I will be arguing for the second sort of understanding of the objects of empirical knowledge, I prefer to use the word "assumption" in this paper, where some other writers on this topic have been using "presupposition". Nothing essential to my argument hangs on this choice of terms.

1. The skeptical problem.

Much recent writing about skepticism starts with Cohen’s (1988) neat formulation of the skeptical problem as three jointly inconsistent propositions (technically, proposition schemas):

(A) S does not know that not-H.

(B) If S does not know that not-H, then S does not know that O.

Resolving this paradox would seem to require the denial of at least one of the three statements.[1] If we substitute me for S, the standard skeptical hypotheses that I am a brain in a vat for H, and for O my claim to know that I am holding a pen, we get:

(a) I do not know that I am not a brain in a vat.

(b) If I do not know that I am not a brain in a vat, then I do not know that I am holding a pen.

Premises (a) and (b) both seem plainly to be true. But then the negation of (c) follows immediately by modus ponens – so my claim to know that I am holding a pen is instantly falsified. This is the most common form of argument for skepticism with regard to empirical knowledge, which DeRose (1995) has sometimes called the Argument from Ignorance (or AI), and which is taken by some to show that we have no empirical knowledge at all. Few philosophers these days are skeptics, though, and most will say that many instances of (C) are plainly true: we obviously do know some empirical things. But affirming this will evidently force us to accept either the negation of (A), which would require our knowing something that appears impossible to know, or the negation of (B), which would make the possibility of our senses being radically deceptive irrelevant to all the knowledge that depends on those senses. Thus, it looks like any way of resolving the problem will require either a good deal of subtlety or a substantial sacrifice of intuitions. Some current philosophers accept the skeptic's argument as sound. Others attack it head-on. The rest try to finesse the problem by carving out some kind of space for restricted sorts of knowledge to exist, or at least for the word "knows" to be correctly ascribed some of the time.

Unger’s (1971) one-time infallibilist approach to the problem affirms premises (A) and (B), plus the skeptical conclusion not-(C). He also posits an explanation of why we think and say that we have knowledge when in fact we do not, by means of a semantic analysis of the concept of knowledge. He claims that it is an “absolute concept” like flatness or emptiness, so that if the property is not there completely, it is not there at all. Ordinary attributions of empirical knowledge might be approximately true, but can never be entirely true, because we can never be absolutely certain in empirical matters. So our common belief that we know things such as that I am holding a pen is literally false, although our thinking and saying such things casually is pretty harmless, in the same way that we can say loosely, but never quite truly, that a wooden board is flat. What actual empirical knowledge would require is the strict impossibility of error – and this is simply unavailable to creatures like ourselves, although we do have relative degrees of certainty, and these matter in practical life. Most epistemologists have found Unger's demand for perfection unreasonable, including, ultimately, Unger (1984) himself.[2] Ordinary usage seems to guarantee that something less than perfect certainty is what we mean when attributing empirical knowledge, since perfect certainty is plainly unattainable, yet such talk is plainly sensible. But it is far from obvious what more modest, fallibilist criterion of knowledge would be satisfactory.

Some philosophers have followed Moore (1959) in denying proposition (A), relying on common sense and common usage to rule skeptical doubts out of bounds. They argue that the falsehood of skeptical hypotheses is entailed by the self-evident, foundational, or unquestionably central status of our least-doubted empirical beliefs, such as that we have hands. Since instances of (C) are obviously true, and (B) goes without saying, (A) must take the fall: a modus tollens reversal of AI. One main problem with this sort of view is that it renders the force of AI mysterious, since the skeptic's hypotheses should strike us as silly rather than perplexing if our ordinary knowledge is so clearly secure. But it is plainly conceivable that people should be deceived in ways that they cannot detect. Indeed, advancing technology makes it seem inevitable that we will one day enjoy entire "virtual reality" experiences internally indistinguishable from real life. If such things were not readily imaginable, then popular movies like Total Recall and The Matrix would be impossible to follow, not to mention Plato's cave analogy and Descartes's First Meditation.

Other recent writers also want to count many ordinary ascriptions of knowledge as true simpliciter, but only if they are uttered in some kind of normal setting. The two best known approaches of this general type are the relevantism of Dretske (1970, 1971, 1981) and Nozick (1981), and the semantic contextualism of Cohen (1987, 1988), DeRose (1992, 1995), and Lewis (1996). Dretske and Nozick affirm both (A) and (C) while denying (B). They defend this choice by arguing against the closure principle on which (B) depends, to the effect that if one knows a proposition P, and knows that P entails Q, then one must also know that Q. Most philosophers find this principle highly intuitive, if not downright self-evident. And it clearly implies that, in order to for me know such things as that I am holding a pen right now, I must be able to rule out such alternative states of affairs as that I am only dreaming, a brain in a vat, etc. But according to Dretske, for me to state truly that I know I am holding a pen, I do not need to rule out every alternative state of affairs consistent with my evidence, only those that are relevant to my circumstances and the statement that I am making. In Dretske’s well-known example, I can truly claim to know that I am looking at a zebra in the zoo when I am actually doing so, despite my inability to rule out the possibility that I am actually looking at a mule that has been painted to look like a zebra. It is true that my knowledge here depends on the fact of my not having been deceived in this way, he says, but it does not depend on my being able to establish this fact, or even being aware of it. In the same way, on Dretske’s view I can truly be said to know such things as that I am holding a pen, while admitting that I don’t know such irrelevant consequences of this claim as that I am not a brain in a vat. My knowing that I am holding a pen is conditioned on the external fact that I am not a brain in a vat, but not on my internal knowledge of this fact.

This is an attractive approach to solving the skeptical problem in that it seems to capture ordinary ascriptions of knowledge in a common-sense way, i.e. by ruling out possibilities of error of the sort that interest only philosophers, in order to get at what we really mean in practical discussion. Yet it demands that we give up not just the closure principle, but also the common-sense idea that we can find out what it is that we know. This presents a danger of circularity for the relevantist. In Dretske's zebra case, if there had actually been a rash of replacement pranks in local zoos, then this would have made the mule hypothesis a relevant alternative to the straight zebra theory, and I would not have known that I was really looking at a zebra until I had ruled out that alternative. But there hasn't been, as far as I know, so I can still know, as far as I know, that I am looking at an actual zebra. Similarly, if there were millions of people's brains actually sparking away in vats, then this skeptical alternative to my empirical beliefs might be a relevant one, something that I would have to eliminate before claiming to know anything empirical. As things actually stand, though, there are no functioning brains in vats at all, as far as I know, so it is not a real alternative, as far as I know. Surely, I can't just walk around (or seem to walk around) looking for a brain-vat with my name on it; if there are actually lots of brains in vats, they are liable to be somewhere that is not empirically accessible to me. So I am in no position to discover how many brains in vats there really are, hence to evaluate the probability that I am such a brain myself. Therefore, Dretske's approach will do me no good in actually arguing against a real or imagined skeptic. I am left insisting on a claim that I have no way to back up, viz.:

(D1) ME: I know that I am holding a pen.

SK: How do you know that you know that, given that you might be a brain in a vat?

ME: I don't. It is true, though, because I'm not in fact a brain in a vat.

SK: But how do you know that you are not in fact a brain in a vat?

ME: I don't.

This makes no better sense than:

(D2) Sandra: I am going to be a big movie star.

Doris: How do you know that, given that so few actresses become stars?

Sandra: I don't. It's true, though, because I'm going to get a lot of lucky breaks.

Doris: But how do you know that you will get a lot of lucky breaks?

Sandra: I don't.

Here, Sandra might be right about becoming a big star, and she might have the correct explanation, too, in terms of lucky breaks. But she cannot reasonably assert these things, because she has no accessible reason to believe them. Similarly, Dretske may have shown that empirical knowledge is objectively possible despite the skeptic, but he has given up on establishing any non-question-begging claims to such knowledge in the face of AI.[3] The best he really has is a conditional like this: If I am lucky (in its not being objectively likely that I am a brain in a vat), then the possibility of such deception is irrelevant, in which case I know that I am holding a pen. This position is, in fact, even less reasonable than Sandra's, in that I can never even find out whether I am lucky in that way.

Semantic contextualists accept the relevantists’ general idea that the truth of our knowledge claims is sensitive to circumstances, but they focus on the linguistic context of the claim, not the external epistemic situation of the speaker or subject. In terms of Cohen's skeptical paradox, they support the closure principle and affirm statement (B), then split their response into two parts. With respect to the context of ordinary discourse, they assert both the antiskeptical conclusion (C) and the negation of (A). With respect to philosophical discussions about knowledge, however, they assert statement (A) and accept the skeptical conclusion not-(C). Thus they agree with either the skeptics or the Mooreans, depending on context. Ordinary knowledge claims are true in ordinary contexts of discussion, they say, because a low standard of knowledge is in place for all such conversations, in that we need not rule out every possibility of mistake in correctly attributing knowledge to somebody, only those alternatives that are ordinarily considered relevant. But when a skeptical hypothesis is bought to our attention, this is enough by itself to change the context into a philosophical or “skeptical” one, which raises the standard of knowledge to infallibility, since in discussing skeptical hypotheses we must consider all conceivable alternatives, such as one’s being a brain in a vat, to be relevant ones. By this higher standard, any ordinary knowledge attribution will be false.

This approach provides a better account for much of our linguistic behavior than does simple relevantism. We say most of the time that we know we have hands, hold pens, and similar things, but we also tend to retract such claims when faced with arguments like AI. Then again, after that skeptical discussion is over, we seem to revert to the first way of speaking with little or no hesitation, and continue to speak that way until skeptical doubts are raised again. This odd, alternating pattern of assertion and denial is the chief symptom of what Lewis (1996) calls the “elusiveness" of knowledge. It suggests to many philosophers that different things are somehow being said using the same verb "to know". Contextualists are able to account for our conflicting intuitive judgments in this way by invoking different epistemic standards. And they do so without sacrificing closure: in the ordinary context, we can say truly (a la Moore) both that we know we have hands and that we know we are not brains in vats; in the skeptical context, we cannot say truly either one. But the contextualist approach has some intuitive problems of its own. For one thing, it seems absurd to say that we can ever really know by any reasonable standard that we are not brains in vats. For another, it is very odd that, though we usually have such knowledge on the contextualist account, we can never truly assert (or even think) that we do, for this would instantly place us in a context where such knowledge claims are always false. For a third, our ordinary concept of knowledge is of a cohesive, fairly stable thing that we possess prior to any discussion and throughout any shifts in topic. The sort of temporary, low-class empirical "knowledge" state that contextualists seem to be attributing to us during ordinary conversations does not fit that central concept at all – unlike the high-grade, philosophical stuff that they concede we lack. It is not easy to see, then, how the contextualists improve on Unger's view that empirical knowledge does not really exist, though we may use the term successfully in an approximate way.

Here is the basic structure of these four responses to Cohen's paradox, with respect to its three jointly inconsistent propositions:

skepticism Mooreanism relevantism contextualism

(A) T F T T/F*

(B) T T F T/T*

* high/low epistemic standard

The view that I want to develop in this paper originates in a pair of non-contextualist responses to Dretske. Yourgrau (1983) and Sanford (1991) point out that even if our ordinary knowledge claims can be interpreted only relative to some sort of context, our "absolute" (Yourgrau) or "proper" (Sanford) knowledge, such as it is, can always be expressed non-contextually, in the form of an appropriate conditional. In Sanford's example, he might make a simple claim to know that Dretske is seated at a certain table, which he might know only with respect to a relevant-alternatives set appropriate for people at a philosophical convention. But alternatively, he says, “[w]hat I know properly I can express conditionally: if that is a member of the usual philosophical crowd, and not someone in disguise, then it is Fred Dretske.” In general, statements of the form "S knows that P" can be replaced by ones of the form "S knows that if C, then P", where C stands for conditions that one is not claiming to know, but has assumed in order to assert the increment of knowledge that P represents. This sort of conditionalization, he says, “shows how a great many of our ordinary attributions of knowledge…are immune from standard skeptical challenge. For what one intends to attribute by these statements is narrower than the original formulation reveals…Having smaller scope, it exposes less flank to skeptical attack.” (1991, pp. 50-51) Sanford grants (as does Yourgrau) that the standard skeptical arguments still work against the genuinely categorical claims that are its usual target, such as his claim to know that he is sitting in his office and not being a brain in a vat. But he says that the argument should no longer be taken to defeat all attributions of empirical knowledge.

Dretske responds to Sanford by insisting that one cannot legitimately bring the background conditions of an utterance into its meaning – what Yourgrau calls the exportation of pragmatic context into semantic content. He argues (1991, p. 189) that we cannot validly infer in general from "if C, then S knows that P" to "S knows that if C, then P", and in particular where C is the external conditions or contextual factors required for empirical knowledge. Our ordinary knowledge is of categorical propositions, though it depends on conditions that we may or may not be aware of. So again, if I am not in fact being tricked by means of a mule painted in stripes or something equally unexpected, then I know simpliciter that the zebra-looking thing in the pen marked "Zebra" is a zebra. This is conditional knowledge, says Dretske, but it is not knowledge of a conditional.

I think that both forms of analysis are sometimes appropriate. There are cases where Dretske's externalist account, which gives the knowledge operator only narrow scope, is plainly correct, such as:

(E1) If my dog has any brains, he knows that Barbara's Weimaraner wants to kill him.

Dogs have no idea what brains are (except, perhaps, as foodstuff), so they can hardly be expected to know conditionals with antecedents that refer to brains. Hence, the condition here (my dog having brains) is clearly external to the content of the knowledge being attributed (that Barbara's Weimaraner wants to kill him).[4] But there are other cases that clearly require a wide-scope reading of the knowledge operator, like:

(E2) Tommy knows that if his sister gets there first, she will eat the whole pie.

Obviously, Tommy's knowledge today cannot be said to depend on his sister's early arrival tomorrow, so all that can be reasonably attributed here is the entire conditional.

To add confusion, we often use the narrow-scope form in English to express such wide-scope knowledge attributions, as in:

(E3) If his sister gets there first, Tommy knows that she will eat the whole pie,

which is ordinarily identical in meaning to (E2). And there are some examples that can be intended either way, like:

(E4) If Christine is smart enough, she knows that they will let her into Swarthmore.

So, it is not as if the matter can be settled instantly, as Dretske seems to suggest, based on intuitions about all knowledge attributions in conditional form.[5]

It seems to me that Dretske's narrow-scope understanding of the conditionality of knowledge is most plausible when knowledge is being attributed to people, animals, or artifacts thought of as knowing in the broad, external sense of having reliably-produced true beliefs (what Sosa 1997 calls animal knowledge), and that a Sanford-style wide-scope reading is correct when the knowledge is attributed to persons qua reflective subjects (Sosa's reflective knowledge). Thus, we are inclined to read the conditional (E4) as like (E1) if we are concerned with Christine's ability to grasp the fact that Swarthmore is admitting her, but like (E3) if we think of her as self-consciously worried about whether she is smart enough to get in.

Dretske wants to deal with skepticism by insisting always on a narrow-scope understanding of the conditionality of knowledge, and this may well be right with respect to knowledge in the animal, reliably-believing sense of the word. But that conception of knowledge is not really what AI is aiming to defeat. Skeptics do well to grant that we have plenty of knowledge in the external sense that we apply to dogs and toddlers – but this is only a semantic concession. It is knowledge in the reflective sense that has always been the skeptic's main target. Unless it can be shown, then, that such knowledge is never what we mean to attribute in empirical discourse, the skeptical problem survives.[6]

2. Conditional substitutes for ordinary knowledge attributions.

I want to extend Sanford's idea of replacing or interpreting some categorical knowledge statements as wide-scope conditionals to cover all ordinary attributions of reflective empirical knowledge. I want to argue that, however successful the skeptical argument is against our ordinary, categorical claims to empirical knowledge, it fails to falsify a corresponding set of claims to knowledge of conditionals, to the effect that if one is not being deceived in some undetectable way, then this or that empirical fact exists. Though we cannot rule out every conceivable alternative to our initial knowledge claims, we can in principle rule out all the empirically accessible ones, and condition our beliefs on the falsehood of the inaccessible ones. Thus, I may not really know that I am holding a pen right now, because I might be dreaming, a brain in a vat, or something just as bad, but I can in principle know that if I am not dreaming, a brain in a vat, or something just as bad, then I am holding a pen right now. Such globally antiskeptical conditionals represent a set of defensible increments of knowledge larger and more systematic than the local conditionals that Sanford talks about.

Look at this discussion between me and an empirical skeptic.

(D3) ME: I am holding a pen.

SK: Do you know that you are holding a pen?

ME: Yes, of course I do. Here is the pen.

SK: Well, is it conceivable that you are dreaming right now?

ME: What?

SK: Can you prove that you are not dreaming right now?

ME: No, I suppose I can't. So what?

SK: If you are dreaming, then you're probably lying in bed, right? In which case you are probably not holding a pen.

ME: Okay.

SK: So, since you might be dreaming, hence asleep in bed, it is possible that you aren't holding a pen. And since you might be mistaken about holding a pen, you can't actually know that you are holding a pen.

ME: Uh….

It suddenly occurs to me to use an antiskeptical conditional:

ME: How about this, then? What I know is that if I am not dreaming, then I am

holding a pen.

The skeptic has no means of batting this away directly. He must try another hypothesis:

SK: Okay, so assume that you are awake. What if you have been given a hallucinogenic drug, so that all you see and feel right now is figmentary? If that's possible, then you still don't know that you are holding a pen.

But I can reply with an expanded conditional:

ME: Fine. I'll say that what I know is that if I am neither dreaming nor hallucinating, then I hold the pen.

And the skeptic is free to keep responding in form:

SK: Then what if you are just a brain in a vat, being fed all your sensations through electrodes?

Now, I can keep hitting lob shots in return, adding a new condition for each specific hypothesis the skeptic produces, or I can put the point away at once:

ME: Look. What I really know is that if I am not dreaming, not hallucinating, not stuck in a vat, and not in any other way that I cannot detect deceived, then I am holding a pen.

What can the skeptic say to this reply? All of his famous hypotheses are on the same theme, that one's senses are being tricked, either by nature or intelligent design, in such a way that one cannot find out about the trick, so one cannot affirm for certain anything that one believes based on the workings of those senses. But we can always bundle these hypotheses together, and exclude them through conditionalization. The term "undetectably deceived" seems to capture the whole range of such skeptical possibilities, so I will use it formulaically, without intending to suggest that speakers ordinarily have such a term in mind. I say that whenever we say that person S knows an empirical fact O, and any standard skeptical challenge is made to that claim, we may respond by substituting for "S knows that O" this conditional:

S knows that if S is not undetectably deceived, then O.

Such attributions of conditional knowledge may be logically weaker than the statements they replace, but it should be clear that they are all impervious to AI. So, it turns out that we retain considerable empirical knowledge – in fact the same amount, proposition for proposition, as we have always thought we had – just in a different form.[7]

Here, then, in outline, is the first version of my theory about skepticism and empirical knowledge:

New Theory One (NT1)

(1) We do not know that the standard skeptical hypotheses are false.

(2) The closure principle is true.

(3) Ordinary utterances of the form "S knows that O" are false.

(4) Ordinary utterances of the form "S knows that O" attribute knowledge of categorical propositions.

(5) But antiskeptical conditionals can be substituted for the objects of our ordinary knowledge attributions at little practical cost.

(6) Skepticism is true with respect to categorical empirical propositions.

(7) Skepticism is false with respect to the antiskeptical conditionals that can be substituted for the objects of our ordinary knowledge attributions.

And here is how it deals with Cohen's skeptical paradox, in contrast to the other theories I have mentioned:

skepticism Mooreanism relevantism contextualism ___ NT1

(A) T F T T/F* T/T**

(B) T T F T/T* T/F**

* high/low epistemic standard

** categorical assertion/conditional replacement

NT1 resembles semantic contextualism in that it gives two parallel lines of analysis. With respect to categorical knowledge attributions as such, NT1 is equivalent to Unger's infallibilist skepticism, as is contextualism with respect to knowledge attributions in the skeptical context. With respect to antiskeptical conditionals, though, NT1 behaves like Dretske's theory, not like Moore's. Thus, NT1does not require us to say we know that we are not being undetectably deceived. But also, unlike relevantism, NT1 requires no violation of the closure principle, because the proposition being attributed to any person S as knowledge is not the categorical O but the substitute conditional:

(O') If not-H, then O,

so that Cohen’s premise (B) expands to:

(B') If S does not know that not-H, then S does not know that if not-H, then O.

Therefore, when I replace my claim to know that I am holding a pen with one about the corresponding antiskeptical conditional, the paradox instantiates to this:

(a) I do not know that I am not undetectably deceived.

(b') If I do not know that I am not undetectably deceived, then I do not know that if I am not undetectably deceived, then I am holding a pen.

(c') I know that if I am not undetectably deceived, then I am holding a pen.

We can reject the second statement here without needing to reject the closure principle, because (b') is false on its face. The conditional proposition [if I am not undetectably deceived, then I am holding a pen] does not even appear to entail [I am not undetectably deceived].

Some readers may object to NT1 on the grounds that antiskeptical conditionals amount to mere tautologies – which, if true, would render the whole theory trivial. For doesn’t being undeceived amount to nothing more than being correct in our judgments? So, when I say that if I am undeceived, then I am holding a pen, don’t I just mean that if I am right in thinking that I am holding a pen, then I am holding a pen? This is no better than saying that if I am holding a pen, then I am holding a pen, or that if I am right, then I am right. Such statements are devoid of content.

This objection is mistaken. Antiskeptical conditionals posit no undetectable deception, not no deception or mistake at all, so they are far from tautologous. The antecedents of these conditionals are just negations of skeptical hypotheses, but their consequents are formed by, and entirely dependent on, contingent facts of our experience. What I have been claiming to know, after all, is that if I am not undetectably deceived, then it is a pen that I am holding in my hand. I cannot sincerely claim to know that if I am not undetectably deceived, then I am holding a cranberry muffin, or a bloody dagger, or a Coupe de Ville, only a pen. For this pen-holding presentation is a feature of the only way the world appears to me right now. If I am making an ordinary empirical error here – say, by mistaking a mechanical pencil for a pen – then the alternative conditional claim will be as false as the categorical one that it replaces. Were a tautology actually needed here, it would have to be something like this: if I am totally undeceived, and if it seems to me that I am holding a pen, then I am holding a pen. But that is not what antiskeptical conditionals assert.[8]

It is possible to argue separately, though, that at least one important class of antiskeptical conditionals is trivially true. Suppose that I really am a brain in a vat, or deceived in some similar way. In that case, the antecedents of any antiskeptical knowledge claims I make will be false, which will make each conditional claim as a whole automatically true, on the truth-functional material interpretation of conditionals. This is because material conditionals are true whenever either their consequents are true or their antecedents are false. It won't matter whether I claim to know that I am holding a pen (as it seems to me that I am) or that I am holding the Taj Mahal (as it seems I am not); these will be equally true claims, provided only that I am thoroughly deceived. This is a bizarre result – really the opposite of common sense. But I think that this problem arises not from NT1, but from the artificial simplicity of material conditionals themselves, about which philosophers have been complaining for years. It is equally unintuitive, for example, to suppose that a poor student in the process of failing a test speaks the truth when he whispers to himself, "If I pass this test, I will one day be President of the United States". Most people working on indicative conditionals reject the material interpretation in favor of something that works better with false antecedents. There are several proposed alternatives, of which the subjunctive analysis of Stalnaker (1968) is perhaps the most widely accepted. On Stalnaker's theory, a conditional with a false antecedent is true just in case the consequent would have been true if the antecedent had been true. This analysis would falsify the desperate hopes of the unrealistic student above, and any knowledge claims made by a brain in a vat that are unjustified from his own point of view.[9] I believe that most other reasonable non-truth-functional theories of indicative conditionals would also do the job.

It can still be objected that antiskeptical conditionals, even if they represent a real form of empirical knowledge, are still too weak to be of much use either in science or in daily life. As Dretske says of Sanford's incremental conditionals, “Immunity [from skeptical challenge] is secured, yes, but for a much less desirable commodity.” (1991, p. 189) If this is true, then even if NT1 must be counted formally as a solution to the skeptical problem, it will not be an intellectually satisfying one. My response is to insist that antiskeptical conditionals, understood as systematically connected, are for all practical purposes equivalent to the corresponding categoricals. If we can construct the whole breadth and complexity of empirical knowledge out of categorical statements, then we can do so just as easily with antiskeptical conditionals. The only interesting principle we need is the equivalence of (A ® B) & (A ® C) to A ® (B & C). Consider these propositions:

(P1) If I am not undetectably deceived, then I am holding a pen.

(P2) If I am not undetectably deceived, then I am wearing glasses.

The conjunction of (P1) and (P2) entails

(P3) If I am not undetectably deceived, then I am holding a pen and I am wearing glasses.

Now conjoin this to the proposition

(P4) If I am not undetectably deceived, then the sun is millions of miles away,

and we can derive

(P5) If I am not undetectably deceived, then I am holding a pen, I am wearing glasses, and the sun is millions of miles away,

and similarly for the antiskeptical counterpart of every other ordinary empirical claim that we would like to add, articulated in the antiskeptical conditional form. The antecedent is redundant for each new case, all the complexity accumulating on the other side of the resulting conditional. Thus, for any coherent conjunction of categorical empirical propositions T, there is a proposition of the form

(P6) If I am not undetectably deceived, then T,

which is as knowable in principle as any of the single propositions T comprises. So, there is no greater problem in making arbitrarily large and elaborate knowledge structures out of antiskeptical conditionals than there is in making them out of categoricals. They could, in fact, be viewed as identical structures, except in that one of them includes a meta-proposition like [this whole structure of empirical beliefs assumes that I am not undetectably deceived], which would implicitly transform the separate categoricals in that structure into antiskeptical conditionals.

Even if they are substantially equivalent in content to the categoricals we usually have in mind, though, someone could still object that these conditionals are not altogether antiskeptical, inasmuch as they do not allow me to distinguish between my own conditional beliefs and those of my imaginary “internal twin”, who really is a brain in a vat. If I am counting these beliefs as empirical knowledge for myself, then I have to say that my twin has exactly the same knowledge. I believe that this objection is correct. Let me concede that my conditional beliefs are weak in just this way. For I can grant to the skeptic that I really do not know whether I am the twin inside the vat or the twin outside the vat, and still insist that each of us knows the conditional that if he is the one outside the vat, then he has a pen in his hand. Therefore, regardless of which twin I turn out to be, I will retain that piece of knowledge. This may seem too much of a concession to the skeptic, but I think that it gives up no more of value than either common sense or the competing theories do. None of the other theories that I have mentioned claims that internal twins are inconceivable, or that there would be any way for me to discover which one is me, assuming that such twins exist. They only claim, at best, that I can know (in some sense of the word) which twin I am, despite my having no way to find out which twin I am.

We should remember that these twins of ours are among the very luckiest of all imaginary brains in vats. They are not being fed mere random sensory noise, or dreamlike incoherencies, or patches roughly stitched together from different possible existences, but rather whole lives of connected experience that they can interpret, remember, and examine rationally. There is only one big problem for them, namely whether it is all mainly real or mainly fake. This is the reflective epistemic situation our internal twins are in. It is also just the situation we are in, if AI with respect to categorical empirical beliefs is sound.

There is nothing really strange about this situation. We often know many things about an object without knowing whether it is real or not. In physics, for example, we can learn a great deal about the properties of, say, black holes or Higgs bosons well in advance of knowing whether such things exist. More to the point, we can often read and understand a story without knowing whether it is a piece of fiction. This happens frequently with elements of historical novels such as War and Peace: we may know perfectly well what is being said by whom (Napoleon, for instance), and how all the story's events fit together, but not know quite where the historical facts leave off and the author's invention begins. Even non-fiction history and journalism can be full of error, not to mention the occasional deliberate lie. What we know directly from written accounts is never that such-and-such events occurred, but only that if the authors are reliable, then those events probably occurred. The same is true for testimonial knowledge in general. We learn most of what we know through testimony one way or another, including all that anybody knows about events that took place before his own early childhood, but it all depends on our assuming that others have been telling us the truth. And the same is true for information coming through our senses. The possibility of error is obvious for any bit of empirical information from any particular source, if less and less vivid for broader and more systematic possible deceptions – but still easily imaginable, as much successful literature, from the Republic on, requires. The skeptic’s global hypotheses are really not so different from more ordinary sources of mistake, just further away from our daily concerns. There is not much to choose, after all, between the dreadful epistemic situations of the characters in 1984 and those in The Matrix. If we are to be rational in interpreting experience, then, we must always leave open the possibility, if only in theory, that we are being misled, deliberately or not, in any number of possible ways, including ways that we cannot detect.

I think that we are generally rational, and that we do in principle accept the possibility of undetectable deception in our empirical beliefs. This is why it seems to be so easy to get people (unscientifically surveyed, I admit) to accept conditional replacements for their knowledge claims once they have been confronted with AI. In an important sense, the antiskeptical conditionals are all that most of us really believe, despite their not being what we usually say. I say "really" believe, for there is a more superficial sense of "believe" in which we do typically believe just what we say in making ordinary claims to categorical empirical knowledge, as in:

(D4) ME: I know that I am holding a pen.

YOU: Is that what you believe?

ME: Of course it is. I just said that I know it, didn't I?

But this is just what might be called the shallow or unreflective sense of "believe". It is the sort of thing that we tend to keep presently in mind, that we are liable to express spontaneously if someone merely asks us what we think. We often speak as if the contents of all our beliefs are fully and directly known to us, but this is merely a convenient assumption, not a psychological fact. What we blurt out in conversation (“I think I left the oven on”) is frequently not the best indicator of our real beliefs, but only of those possible states of affairs that we are most concerned about at present. Our everyday conversations may bring certain ideas forcefully to mind, but our actual beliefs, in their full structure, are sometimes obscure to us, and can only be evoked through further discussion. It is not just what we say we think spontaneously, but also how we answer questions and objections, that reveals our deep or considered beliefs – what we really believe, when all is said and done – and not only to others, but sometimes to ourselves as well. In some cases, it may take discussions with our teachers, psychologists, or other advisors for us to discover (or, perhaps, “recover”) what is really there, as in:

(D5) Joseph: I believe my parents did the best they could. They were both very busy.

Dr. Siegel: Is that really what you believe? Do you really believe that they couldn't have permitted you to play with friends or join the Little League?

Joseph: No…I guess I do know that they could have done a better job. I just don't want to admit it, even to myself.

In the epistemological case at hand, I suggest that it is philosophical discussions about skepticism that can bring out the full structure of our everyday empirical beliefs.

There are good reasons for the conditionality of everyday beliefs to be somewhat hidden from us psychologically. For one thing, it is the consequents, not the antecedents, of antiskeptical conditionals that affect our lives in practice, so these are cognitively stored in an accessible and vivid way. The more remote alternatives appear, the less intensely they are liable to impinge on our occurrent thoughts. And the hypotheses that skeptics make the most of – that we are dreaming, that we are brains in vats, that we are otherwise undetectably deceived – are of almost no practical interest at all, given that we cannot know whether they are true or false. So they make little impression on our conscious minds, unless our minds are focused at the moment on epistemology. For another, since these deepest antecedents are all in any case essentially the same, to the effect that we are mainly reliable in our empirical judgements, there is no reason to expect that they be stored in memory redundantly, once in connection to each consequent. That would take a huge number of identical representations. It is surely more efficient cognitively to detach this uniform component of all of our empirical beliefs, and to use the bulk of our mental capacity for things that make a difference in practical reasoning. Thus, what appears at first glance, even introspectively, to be a set of separate categorical beliefs ought to be seen instead as the efficient storage of conditional ones that have identical, practically uninteresting antecedents.[10]

Psychological factors can create confusion about what we believe, but philosophy can straighten us out. In my discussion with the skeptic (D3), I discovered in the face of AI that my initial, categorical claim to know that I am holding a pen is false, and all I really know is its conditional counterpart, that if I am not undetectably deceived, then I am holding a pen. I now see, too, that I never even really thought that I knew this sort of thing in a completely categorical way. The conditional is all that I could have believed in the first place, I realize, because of course I cannot know that I am not being undetectably deceived, and this of course entails that I cannot know unconditionally that I am holding a pen. So, when I think about it, all that I would really want to commit myself to philosophically is the conditional. That is my deep belief. If this is right, then NT1's replacement scheme ought to be seen as more than just a way to get around AI. It is also a way of showing that, however much we might confuse ourselves with the shortcuts of everyday thinking and speech, the skeptical argument does little or no damage to our most serious empirical beliefs. We just need to look at these beliefs in terms of AI's implications, and perceive that their conditionality is harmless to practical reasoning.

To the extent that we were willing to adopt the antiskeptical conditional replacements for our ordinary knowledge claims, we could prevent the skeptical paradox from ever coming up. It would be nice for us if we could all agree to speak this way consistently. If all of our empirical knowledge claims attributed only the proper antiskeptical conditionals, then empirical skeptics would have no complaints against us, and this longstanding problem in epistemology would disappear. We can at least imagine people who habitually think and speak this way, always alert to the limits of their senses and other faculties, always careful not to claim what they cannot establish as a fact. It might take a lot of trouble for them to cover these concerns explicitly whenever speaking or thinking about empirical matters – too much trouble, perhaps, for busy people like ourselves. They would have to be rather like the patient and philosophical Houynynymmhs in Gulliver's Travels, for whom the constant hedging that completely antiskeptical discourse requires would seem less awkward than it does to us, and among whom "saying the thing which is not" would be considered more shameful. There would be nothing wrong with people like this, no intellectual function that they could not perform. It would just take them longer to express themselves.[11]

3. The conditional analysis of ordinary knowledge attributions.

A bolder new theory is possible. It might be claimed that knowledge of antiskeptical conditionals is not just something that we ought to be attributing to people instead of categorical knowledge. In fact, it is all that we actually are attributing. For the conditionals are not just workable substitutes, and not just more complete expressions of our real beliefs; they are all that we really even mean when using categoricals in ordinary knowledge statements. No replacement statements would be needed, in this view, to get around the skeptical problem – indeed, the problem could be solved directly in this way, not just avoided. We would need only to insist on the correct conditional analysis of ordinary knowledge attributions.

In (D3) I claimed to know that I was holding a pen, and took it for granted that this statement was as categorical as it appeared to be – but did I really mean it as a claim to knowledge that is absolutely unconditional? I certainly did not say I meant it unconditionally, when I could easily have done so – though this would have made my statement obviously false. I could also have said that I knew it conditionally (in the wide-scope sense discussed above), or even spelled out the conditions, and I did not do that either. I just said that I knew it, and I did not say which way I knew it until the skeptic pushed me into a more explicit statement. Now, where does this leave the argument? I think the situation is like this. I said something that could have been taken two ways, categorically or conditionally. The skeptic argued that my claim was wrong, on the assumption that I meant it to be understood in the categorical way, and I responded by replacing its object with an explicitly conditional one. Looking back, though, I could just as easily have responded that I meant my statement to be understood the other way in any case:

(D6) …

SK: If you are dreaming, then you're probably lying in bed, right? In which case you are probably not holding a pen.

ME: Okay.

SK: So, since it's possible that you are dreaming, hence in bed, it is possible that you aren't holding a pen. And since you might be mistaken about holding a pen, it can't be true that you know that you are holding a pen.

ME: Uh….

This time, it strikes me that I never actually claimed to have ruled out such possibilities as that I am presently dreaming. I have just been assuming that I am not dreaming, as we all do in practical discussions.

…

ME: Listen. When I said I know that I am holding a pen, I didn't mean to say that I know I'm holding a pen even if I'm only dreaming. How could anybody know a thing like that? I just meant, you know, assuming that I'm awake, and in my right mind, and so on, like everybody does, then I am holding a pen. That is all I actually think I know, and all I really ever meant to convey.

Is this not a perfectly reasonable response to skeptics who inject AI into our ordinary conversations?

Philosophical discussions are a different matter, not because this sort of linguistic context raises our standards of knowledge, but just because things are more literal and spelled-out. If you like, our standards of clarity are higher. My discussion with the skeptic might continue:

…

SK: So, in the end, what do you want to say? Do you know that you are holding a pen, or don't you?

ME: Well, do you mean categorically or conditionally?

SK: Just, do you know that you are holding a pen? Yes or no answer.

ME: Then I will take it that you mean to ask whether I know the categorical proposition, "I am holding a pen", to be true in that form. And my answer is no. But that is not what I was really claiming earlier, and not the sort of thing I ordinarily intend.

Now, this sort of answer will not silence every skeptic, or satisfy every anti-skeptic. Those who still seek to justify reflective knowledge of genuinely categorical empirical propositions will have to look elsewhere. But most people writing on this topic have already given up on such a total refutation of the skeptical position. What remains for us is to establish the most generally satisfactory account of such knowledge as is not defeatable by arguments like AI, which means that we must give up something or other that we like. I am claiming here that if we give up the surface, categorical interpretation of our ordinary knowledge attributions in favor of the antiskeptical-conditional one, then pretty much everything else can be preserved.

Hidden conditional forms are not really unfamiliar, and not always very deeply hidden. In fact, we use categorical sentences to express conditional propositions all the time in ordinary discourse, including almost any time that we directly answer a hypothetical question. Consider the following pair of brief discussions:

(D7) Andrea: If I get fired, I am going to sue the company.

(D8) Beth: What are you going to do if you get fired?

Andrea: I am going to sue the company.

Andrea’s statement in (D8) is plainly equivalent in meaning to her statement in (D7), despite their different surface forms.[12] They both express the same conditional proposition, although the sentence Andrea uses in (D8) is a categorical one. And this intuitive conditional analysis seems not to be affected when we apply the knowledge operator:

(D9) Beth: Do you know what you will do if you get fired?

Andrea: Yes. I know that I will sue the company.

Plainly, Andrea's response in (D9) amounts to saying only that she knows what she will do if she gets fired, not what she will do regardless. There is no confusion about what is actually meant in these examples, because the hypothesis that Andrea is to be fired has just been raised, and plainly stands before both speakers' minds.

But other conversational assumptions ride somewhat further back, and may catch us off guard when someone questions them, for example:

(D10) Vinny: After the job, we'll meet up at Tony's place.

Lenny: You mean, even if we get caught?

Vinny: Uh…no. If we get caught, we'll probably meet again in jail. I meant that if we don't get caught, we'll meet up at Tony's. Got it?

We could say, I suppose, that Vinny's first statement is unjustified, and that he only got it right the second time. If he had wanted to tell the truth up front, then he should have explicitly conditioned his first statement on the proposition that the gang will not get caught. But this is surely too much to expect of him, since it is clear to everyone involved (except for one slow guy) that the conditional is all he really meant.

Ordinary discourse is shot through with such suppressed conditionality.[13] Without background assumptions, most of our discussions would be painfully inefficient, requiring massive hedging in advance of any interesting statement. And the further in the background these assumptions lie, the harder it can be to dredge them up when we are called on to make a fully analyzed or fully justified statement. Some such assumptions are common to virtually all empirical discussions, e.g. that the participants are mainly sincere, that it is better to believe true things than false, and that the laws of nature are not subject to arbitrary change. But the most basic one, hence the furthest from the surface of our ordinary thoughts, is the assumption that we are not so profoundly and systematically deceived as to make empirical discussion pointless in the first place.[14] We assume our own basic empirical reliability – fallible, no doubt, but not too fallible – as a condition of our having anything interesting to say.

Here, then, in outline, is the second, stronger variant of my approach:

New Theory Two (NT2)

(1) We do not know that the standard skeptical hypotheses are false.

(2) The closure principle is true.

(3) Ordinary utterances of the form "S knows that O" are often true.

(4) In some non-ordinary contexts such as philosophical discussions, utterances of the form "S knows that O" attribute knowledge of categorical propositions.

(5) But in ordinary contexts of discussion, utterances of the form "S knows that O" typically attribute knowledge of antiskeptical conditionals.

(6) Skepticism is true with respect to categorical empirical propositions.

(7) Skepticism is false with respect to the antiskeptical conditionals that are the actual objects of our ordinary knowledge attributions.

And here is a (very slightly) revised chart contrasting this revised response to Cohen's paradox with the four existing theories that I have mentioned above:

skepticism _Mooreanism _relevantism _contextualism NT2

(A) T F T T/F* T/T**

(B) T T F T/T * T/F**

* high/low epistemic standard

** categorical/conditional interpretation

The only difference here from NT1 is in the key for **, where I have replaced two different statements with two interpretations of one statement. But NT2 presents a fuller, more direct response to Cohen’s paradox, with greater intuitive advantages over the other four approaches. Unlike Unger’s skepticism, the new theory allows for ordinary knowledge attributions to be true, even if infallibility is required. Unlike Mooreanism, it allows as true what seems obvious to most of us, that we do not know that we are not being undetectably deceived, while still defending ordinary knowledge claims. Unlike relevantism, it retains the very plausible closure principle while supporting our attributions of reflective knowledge, which are what skeptics really target. With respect to knowledge claims as we use them in ordinary conversation, NT2 agrees with Dretske and Nozick in affirming both (A) and the antiskeptical conclusion (C), while rejecting the major premise (B). NT2 does not need to deny the closure principle that underlies the plausibility of (B), though, only the relevance of that principle to the correct analysis of (B). This is because NT2, like NT1 regarding antiskeptical replacements of our ordinary knowledge claims, expands the major premise of AI to instances like:

(b’) If I do not know that I am not undetectably deceived, then I do not know that if I am not undetectably deceived, then I am holding a pen,

which are straightforwardly false.

The conditional-interpretive approach of NT2 also has clear intuitive advantages over semantic contextualism, despite its similar distinction between ordinary and philosophical contexts of discussion. First, as noted, contextualists have to accept the Moorean idea that we actually know (albeit only by the low standards of ordinary discussion) that we have not been undetectably deceived, despite this knowledge being absolutely inaccessible to us, while NT2 rejects this awkward position. Second, as with NT1 above, the new theory does not require an account of shifting epistemic standards, with which contextualism effectively splits the concept of knowledge in two, but only the familiar idea of hidden logical form. This allows knowledge itself to remain the single, generally stable, context-independent stuff of common sense. In NT2, it is non-contextually true that one knows certain empirical things in the form of antiskeptical conditionals, and non-contextually false that one knows the corresponding categoricals, while contextualists cannot confirm that anything empirical is known simpliciter, outside of real or hypothetical discussions.

These are just quick points, of course. I am not trying to argue that contextualists and other theorists cannot solve their major intuitive problems in ways consistent with the motivations for their views. There are plenty of arguments in play, and much ingenuity on all sides. My main claim here is that NT2 resolves the skeptical paradox as clearly as these other theories, in a reasonable (if not, perhaps, independently compelling) way, without having to solve these major problems.

NT2 does have some problems of its own, however – two that stand out as potentially fatal objections. The first is what some readers will consider the complete intuitive implausibility of its central analysis. How can it be correct to say that antiskeptical conditionals are all that we express in ordinary knowledge attributions? Such attributions are commonly understood and treated as genuinely categorical ones. If they were just ellipses for conditionals, why wouldn't this be obvious to us? Why is the conditional analysis something new, and not just the automatic, common-sense response to skepticism?

The second, more specific objection points out that statements of the form "S knows that p" can be true on NT2's analysis while p itself is false. This seems to violate the almost universally accepted truth condition for knowledge:

(TCK) Necessarily, if S knows that p, then p is true.

Suppose I say in some ordinary context that I know that I am holding a pen. According to TCK, this statement is false whenever I am not actually holding a pen. NT2's analysis of my statement should also make the statement false, then, if that analysis is to be plausible intuitively. Here is the problem. The analysis says that what I ordinarily mean by "I know that I am holding a pen" is that I know that if I am not undetectably deceived, then I am holding a pen. But if it turns out that I actually am undetectably deceived, but that if I weren't deceived I would be holding a pen (other things being equal), then NT2's analysis will make the statement true, not false. And this is not just a technical error of some sort – having such statements come out true is one of NT2's main goals. So it looks like either NT2 or TCK will have to be abandoned – and TCK is about as well-established as a philosophical principle can be.

A quick response is that NT2's analysis only appears to violate TCK. After all, the theory says that ordinary utterances of the form "S knows that p" do not really claim that S knows p, but only that S knows the antiskeptical conditional [if S is not undetectably deceived, then p]. The truth of p itself is not entailed by this analyzed statement, then, only the truth of the conditional. Thus, the statement that I actually make in uttering "I know that I am holding a pen" is not falsified when I turn out not to be holding a pen, provided the reason is only that I am undetectably deceived. So we just have to keep clear about what is really being said, and we will avoid all such apparent violations of TCK. The problem with this quick response is that it fails to respect our intuitions about when people's ordinary knowledge claims are true and false. If I claim to know that I see a zebra at the zoo, and it turns out to have been only a zebra-painted mule, then I was just wrong. It doesn't matter whether I was thinking I could rule out such oddball possibilities or not. What I said was that I knew it was a zebra – and it wasn't, so I didn't. Similarly, if I claim to know that I am holding a pen right now, and I find out later that the pen was only a hallucination, I will recant my prior claim to knowledge. I wasn't holding a pen, so I didn't know that I was holding a pen, so my claim to know that I was holding a pen was false. Thus, the TCK objection, though it can be countered in a superficial way, still serves to point out just how implausible NT2's conditional analysis of knowledge claims can be.

This is a powerful intuitive objection, but it is not conclusive in itself. For we know that philosophical analysis will often bend linguistic practice and pre-philosophical belief to some extent in order to achieve a theory that is more intuitive in other ways, or simply better. The connections of meaning that philosophers seek are sometimes very subtle ones of which the ordinary speaker is not aware, and it can take considerable thought and argument to see that two things treated differently in ordinary thought and discourse must at bottom be the same, or that two things thought of as the same must be different. The classic example is Russell's (1905) analysis of definite descriptions, which attempts to solve the problem of reference to non-existent things, many have thought successfully, by exposing a complex, initially counter-intuitive logical form beneath our ordinary, simple but problematic utterances. A sentence like "The Dalai Lama is a man" seems intuitively to have the same subject-predicate form as "This is a man", but Russell would argue that the meaning of the former must be analyzed as [There is one and only one thing that is a Dalai Lama, and that thing is a man]. Now, it may take quite a while to convince people that this is the sort of thing they really mean in using sentences with definite descriptions, just as it can take a while to convince people that by "I have nothing in my pockets" they mean only the negation of the proposition [I have something in my pockets], or that by "All men are mortal" they mean to say of each thing in the universe that if it is a man, then it is mortal. But these analyses are widely accepted by philosophers. NT2's conditional analysis of knowledge attributions can be viewed as the same kind of analytical effort – intuitively odd at first, but worth getting used to if it helps to solve a long-standing problem. So, I might concede that the objects of empirical knowledge claims are not ordinarily treated as conditionals – including with respect to TCK – any more than statements containing definite descriptions are treated as quantified conjunctions. But this does not entail that they do not turn out to be conditionals on the correct analysis.[15]

A good criterion of meaning ought to take dialectical factors into account. Instead of assuming that our meaning always has essentially the surface structure of our utterance, we might think of a statement's meaning as whatever would be counted as the best explication of that statement, after appropriate discussion and reflection by all concerned. What I intend here by the "meaning" of a statement is of course what Grice (1957) calls speaker's meaning, namely whatever proposition one intends to convey on a particular occasion of utterance, as opposed to the objective sentence meaning, which is fixed in the language. I take it for granted that the sentence meaning of "S knows that O" is just [S knows that O]; it is what the speaker usually means by saying "S knows that O" that NT2 is analyzing in conditional terms. We should distinguish further, in parallel with the two levels of belief distinguished above, between spontaneous or shallow meaning (what we casually suppose we mean) and considered or deep meaning (what we would say we really mean, upon appropriate consideration). A speaker's shallow meaning tends to match sentence meaning more closely, for obvious reasons, while deep meaning can require analysis that results in something that looks very different, just like deep belief. So, if I claim to know that I am holding a pen, and someone asks me if I mean by this that I cannot be wrong about that fact, I might well thoughtlessly say "yes". But if I am then asked whether I have claimed to know that I am holding a pen regardless of whether I am dreaming at the moment, and I reply "of course not, not if I am only dreaming", then I should be counted as having really meant something else. If the explanation is suggested that I actually intended to claim knowledge of a conditional, to the effect that if I am not dreaming (or in general if I am not being undetectably deceived), then I am holding a pen, I might respond, after considering the matter, "yes, that is all I really meant". In this case, I ought to be understood as having really meant only the conditional claim. To the extent that our pre-philosophical intuitions and linguistic practice seem to tell us otherwise, this is because they represent only the shallow meanings of everyday discussion. The special problem with TCK stems from the pragmatic fact that ordinary discourse trades in categorical, objective statements, which saves us the trouble of framing every empirical claim in terms of cumbersomely autobiographical, though more defensible,