*[[This paper is largely based on material in other papers. The first three sections and the appendix are drawn with minor modifications from Chalmers 2002c (which explores issues about phenomenal concepts and beliefs in more depth, mostly independently of questions about materialism). The main ideas of the last three sections are drawn from Chalmers 1996, 1999, and 2002a, although with considerable revision.]]
The classic statement of the knowledge argument against materialism has been given by Frank Jackson (1982). Mary knows everything that can be stated in physical terms about the physical processes that are in any way relevant to color vision. But Mary has never experienced colors other than black, white, and shades of grey. It seems that Mary has complete physical knowledge, but she does not have complete phenomenal knowledge: in particular, she does not know what it is like to see red. Jackson argues that Mary knows all the physical facts, but does not know all the facts. When she sees red for the first time, she learns a new fact concerning what it is like to see red. So there are facts over and above the physical facts, and materialism is false. In particular, phenomenal facts - facts about the character of conscious experience - are nonphysical facts, and phenomenal properties are nonphysical properties.
In this paper, I will analyze some issues that are relevant to the knowledge argument. In particular, I will address issues about phenomenal concepts and phenomenal belief. The knowledge argument crucially turns on Mary's lack of knowledge, and on her acquisition of new beliefs. To fully understand the situation, we need to understand the nature of the knowledge that Mary lacks, and the nature of the beliefs that Mary gains. This requires in turn an analysis of the concepts that are involved in such beliefs. I will bring this analysis to bear on the knowledge argument itself.
In what follows, I will assume phenomenal realism: roughly, the view that Mary acquires new factual knowledge when she sees red for the first time. This excludes views on which merely gains a new ability, or on which she gains no knowledge at all. It is compatible with views on which Mary gains knowledge of an old fact in a new way. The important aspect of this view is that it allows an epistemic gap between physical truths and phenomenal truths, in the sense that phenomenal truths are not entailed a priori by physical truths. The view is silent on whether or not there is an ontological gap. As such, the view excludes "type-A materialist" views such as those of Dennett and Lewis, which deny an epistemic gap, but it includes "type-B" materialist views such as those of Loar and Tye, which allow an epistemic gap but denying an ontological gap, as well as including many non-materialist views.
A note on modality: Because I am assuming phenomenal realism but not property dualism, references to necessity and possibility in the first part of this paper should be taken as invoking conceptual necessity and possibility. Similarly, talk of possible worlds can be taken as invoking conceivable worlds, corresponding to the epistemically constructed scenarios of Chalmers (forthcoming). If one accepts a certain sort of link between conceptual and metaphysical possibility (e.g. the thesis that a posteriori statements have a netaphysically contingent epistemic intension), then these references can equally be taken as invoking metaphysical possibility and necessity.
Phenomenal properties are properties characterizing what it is like to be a subject, or what it is like to be in a mental state. Phenomenal beliefs are beliefs that attribute phenomenal properties. I will be especially concerned with first-preson phenomenal beliefs, such as "I am now having a red experience". These phenomenal properties are attributed using phenomenal concepts. To understand phenomenal beliefs, we need to understand phenomenal concepts.
I look at a red apple, and visually experience its color. This experience instantiates a phenomenal quality R, which we might call phenomenal redness. It is natural to say that I am having a red experience, even though of course experiences are not red in the same sense in which apples are red. Phenomenal redness (a property of experiences, or of subjects of experience) is a different property from external redness (a property of external objects), but both are respectable properties in their own right.
I attend to my visual experience, and think I am having an experience of such-and-such quality, referring to the quality of phenomenal redness. There are various concepts of the quality in question that might yield a true belief.
We can first consider the concept expressed by 'red' in the public language expression 'red experience', or the concept expressed by the public language expression 'phenomenal redness'. The reference of these expressions is fixed via a relation to red things in the external world, and ultimately via a relation to certain paradigmatic red objects that are ostended in learning the public language term 'red'. A language learner learns to call the experiences typically brought about by these objects 'red' (in the phenomenal sense), and to call the objects that typically bring about those experiences 'red' (in the external sense). So the phenomenal concept involved here is relational, in that it has its reference fixed by a relation to external objects. The property that is referred to need not be relational, however. The phenomenal concept plausibly designates an intrinsic property rigidly, so that there are counterfactual worlds in which red experiences are never caused by red things.
One can distinguish at least two relational phenomenal concepts, depending on whether reference is fixed by relations across a whole community of subjects, or by relations restricted to the subject in question. The first is what we can call the community relational concept, or red_C. This can be glossed roughly as the phenomenal quality typically caused in normal subjects within my community by paradigmatic red things. The second is what we can call the individual relational concept, or red_I. This can be glossed roughly as the phenomenal quality typically caused in me by paradigmatic red things. The two concepts red_C and red_I will corefer for normal subjects, but for abnormal subjects they may yield different results. For example, a red/green-inverted subject's concept red_C will refer to (what others call) phenomenal redness, but his or her concept red_I will refer to (what others call) phenomenal greenness.
The public language term 'red' as a predicate of experiences can arguably be read as expressing either red_C or red_I. The community reading of 'red' guarantees a sort of shared meaning within the community, in that all uses of the term are guaranteed to corefer, and in that tokens of sentences such as 'X has a red experience at time t' will have the same truth-value whether uttered by normal or abnormal subjects. On the other hand, the individual reading allows a subject better access to the term's referent. On this reading, an unknowingly inverted subject's term 'red' will refer to what she think it refers to (unless the inversion was recent), while on the community reading, her term 'red' may refer to something quite different, and her utterance 'I have had red experiences' may even be unknowingly quite false.[*] In any case, we need not settle here just what is expressed by phenomenal predicates in public language. All that matters is that both concepts are available.
*[[These cases may not be entirely hypothetical. Nida-Rumelin (1996) gives reasons, based on the neurobiological and genetic bases of colorblindness, to believe that a small fraction of the population may actually be spectrum-inverted with respect to the rest of us. If so, it is natural to wonder just what their phenomenal expressions refer to.]]
Phenomenal properties can also be picked out by demonstratives. When seeing the tomato, I can refer demonstratively to a visual quality associated with it, using a concept I might express by saying 'this quality' or 'this sort of experience'. The relevant concept here is the phenomenal demonstrative that we might call E, which functions roughly by picking out whatever quality the subject is currently ostending. Like other demonstratives, it has a "character", which fixes reference in a context roughly by picking out whatever quality is ostended in that context; and it has a distinct "content", corresponding to the quality that is actually ostended - this case, phenomenal redness. The demonstrative concept E rigidly designates its referent, so that it picks out the quality in question even in counterfactual worlds in which no-one is ostending the quality.
The three concepts red_C, red_I, and E may all refer to the same quality, phenomenal redness. All of them fix reference to phenomenal redness relationally, characterizing it in terms of its relations to external objects or acts of ostension, and all of them designate this quality rigidly.
There is another crucial phenomenal concept in the vicinity, one that does not pick out phenomenal redness relationally, but rather picks it out directly, in terms of its intrinsic phenomenal nature. This is what we might call a pure phenomenal concept.
To see the need for the pure phenomenal concept, consider the knowledge that Mary gains when she learns for the first time what it is like to see red. She learns that seeing red has such-and-such quality. Mary learns (or reasonably comes to believe) that red things will typically cause experiences of such-and-such quality in her, and in other members of her community. She learns (or gains the cognitively significant belief) that the experience she is now having has such-and-such quality, and that the quality she is now ostending is such-and-such. Call Mary's "such-and-such" concept here R.
Mary's concept R picks out phenomenal redness, but it is quite distinct from the concepts red_C, red_I, and E. We can see this by using cognitive significance as a test for difference between concepts. Mary gains the belief red_C=R - that the quality typically caused in her community by red things is such-and-such - and this belief is cognitively significant knowledge. She gains the cognitively significant belief red_I = R in a similar way. And she gains the belief E=R - roughly, that the quality she is now ostending is such-and-such.
Mary's belief E=R is as cognitively significant as any other belief in which the object of a demonstrative is independently characterized: e.g. my belief I am David Chalmers, or my belief that object is tall, or my belief that shape is roundness. For Mary, E=R is not a priori. No a priori reasoning can rule out the hypothesis that she is now ostending some other quality entirely, just as no a priori reasoning can rule out the hypothesis that I am David Hume, or that the object I am pointing to is short. Indeed, nothing known a priori entails that R is ever instantiated in the actual world.
So the concept R is quite distinct from red_C, red_I, and E. Unlike the other concepts, the pure phenomenal concept characterizes the phenomenal quality as the phenomenal quality that it is.
The concept R is difficult to express directly in language, since the most natural terms, such as 'phenomenal redness' and 'this experience', arguably express other concepts such as red_C and E. Still, one can arguably discern uses of these terms that express pure phenomenal concepts; or if not, one can stipulate such uses. For example, Chisholm suggests that there is a "noncomparative" sense of expressions such as 'looks red'; this sense seems to express a pure phenomenal concept, whereas his "comparative sense" seems to express a relational phenomenal concept.[*] And at least informally, demonstratives are sometimes used to express pure phenomenal concepts. For example, the belief that E=R might be informally expressed by saying something like "this quality is this quality".
*[[The distinction also roughly tracks Nida-Rumelin's (1995; 1997) distinction between "phenomenal" and "nonphenomenal" readings of belief attributions concerning phenomenal states. "Phenomenal" belief attributions seem to require that the subject satisfies the attribution by virtue of a belief involving a pure phenomenal concept, while "nonphenomenal" attributions allow that the subject can satisfy the attribution by virtue of a belief involving a relational phenomenal concept.]]
It may even be that there is a broad sense in which R can be regarded as a "demonstrative" concept. I will not regard it this way: I take it that demonstrative concepts work roughly as analyzed by Kaplan (1977), so that they have an epistemic reference-fixing "character" that leaves their referent open. This is how E behaves: its content might be expressed roughly as "this quality, whatever it happens to be". R, on the other hand, is a substantive concept that is tied a priori to a specific sort of quality, so it does not behave the way that Kaplan suggests that a demonstrative should. Still, there is an intimate relationship between pure and demonstrative phenomenal concepts that I will discuss later in the paper; and if someone wants to count pure phenomenal concepts as "demonstrative" in a broad sense, there is no great harm in doing so, as long as the relevant distinctions are kept clear. What matters for my purposes is not the terminological point, but the more basic point that the distinct concepts E and R exist. For a phenomenal realist, there can be no doubt about this.[*]
The relations among these concepts can be analyzed straightforwardly using the two-dimensional framework for representing the content of concepts. A quick introduction to this framework is given in an appendix; more details can be found in Chalmers (2002b).
According to the two-dimensional framework, when an identity A=B is a posteriori, the concepts A and B have different epistemic (or primary) intensions. If A and B are rigid concepts and the identity is true, A and B have the same subjunctive (or secondary) intensions. So we should expect that the concepts red_C, red_I, E, and R have different epistemic intensions, but the same subjunctive intension. And this is what we find. The subjunctive intension of each picks out phenomenal redness in all worlds. The epistemic intension of red_C picks out, in a given centered world, roughly the quality typically caused by certain paradigmatic objects in the community of the subject at the center of the world. The epistemic intension of red_I picks out roughly the quality typically caused by those objects in the subject at the center.
As for the demonstrative concept E: to a first approximation, one might hold that its epistemic intension picks out the quality that is ostended by the subject at the center. This characterization is good enough for most of our purposes, but it is not quite correct. It is possible to ostend two experiences simultaneously and invoke two distinct demonstrative concepts, as when one thinks that quality differs from that quality, Ostending two different parts of a symmetrical visual field (see Austin 1990). Here no descriptive characterization such as the one above will capture the difference between the two concepts. It is better to see E as a sort of indexical, like I or now. To characterize the epistemic possibilities relevant to demonstrative phenomenal concepts, we need centered worlds whose centers contain not only a "marked" subject and time, but also one or more marked experiences; in the general case, a sequence of such experiences.[*] Then a concept such as E will map a centered world to the quality of the "marked" experience (if any) in that world. Where two demonstrative concepts E1 and E2 are involved, as above, the relevant epistemic possibilities will contain at least two marked experiences, and we can see E1 as picking out the quality of the first marked experience in a centered world, and E2 as picking out the quality of the second. Then the belief above will endorse all worlds at which the quality of the first marked experience differs from the quality of the second. This subtlety will not be central for the purposes of this paper.
*[[In the experience-based framework: if experiences do not map one-to-one to instances of phenomenal properties, then instances of phenomenal properties should be marked instead.]]
The epistemic intension of R is quite distinct from all of these. It picks out phenomenal redness in all worlds. I will analyze this matter in more depth shortly; but one can see intuitively why this is plausible. When Mary believes roses cause R experiences, or I am currently having an R experience, she thereby excludes all epistemic possibilities in which roses cause some other quality (such as G, phenomenal greenness), or in which she is experiencing some other quality: only epistemic possibilities involving phenomenal redness remain.
The cognitive significance of identities such as red_C=R, red_I=R, and E=R is reflected in the differences between the concept's epistemic intensions. The first two identities endorse all epistemic possibilities in which paradigmatic objects stand in the right relation to experiences of R; these are only a subset of the epistemic possibilities available a priori. The third identity endorses all epistemic possibilities in which the marked experience at the center (or the ostended experience, on the rough characterization) is R. Again, there are many epistemic possibilities (a priori) that are not like this: centered worlds in which the marked experience is G, for example. Once again, this epistemic contingency reflects the cognitive significance of the identity.[*]
We can now complicate the situation by introducing another thought-experiment on top of the first one. Consider the case of Inverted Mary, who is physically, functionally, and environmentally just like Mary, except that her phenomenal color vision is red/green inverted. (I will assume for simplicity that Inverted Mary lives in a community of inverted observers.) Like Mary, Inverted Mary learns something new when she sees red things for the first time. But Inverted Mary learns something different from what Mary learns. Where Mary learns that tomatoes cause experiences of (what we call) phenomenal redness, Inverted Mary learns that they cause experiences of (what we call) phenomenal greenness. In the terms given earlier, Mary acquires beliefs red_C=R, red_I=R, and E=R, while Inverted Mary acquires beliefs red_C=G, red_I=G, and E=G (where G is the obvious analog of R). So Mary and Inverted Mary acquires beliefs with quite different contents.
This is already enough to draw a strong conclusion about the irreducibility of content. Recall that Mary and Inverted Mary are physical/functional and environmental twins, even after they see red things for the first time. Nevertheless, they have beliefs with different contents. It follows that belief content does not supervene conceptually on physical/functional properties. And it follows from this that intentional properties are not conceptually supervenient on physical/functional properties, in the general case.
This is a nontrivial conclusion. Phenomenal realists often hold that while the phenomenal is conceptually irreducible to the physical and functional, the intentional can be analyzed in functional terms. But if what I have said here is correct. then this irreducibility cannot be quarantined in this way. If the phenomenal is conceptually irreducible to the physical and functional, so too is at least one aspect of the intentional: the content of phenomenal beliefs.
At this point, there is a natural temptation to downplay this phenomenon by reducing it to a sort of dependence of belief content on reference that is found in many other cases: in particular in the cases that are central to externalism about the content of belief. Take Putnam's case of Twin Earth. Oscar and Twin Oscar are functional duplicates, but they inhabit different environments: Oscar's contains H2O as the clear liquid in the oceans and lakes, while Twin Oscar's contains XYZ (which we count not as water but as twin water). As a consequence, Oscar's water concept refers to water (H2O), while Twin Oscar's analogous concept refers to twin water (XYZ). Because of this difference in reference, Oscar and Twin Oscar seem to have different beliefs: Oscar believes that water is wet, while Twin Oscar believes that twin water is wet. Perhaps the case of Mary and Inverted Mary is just like this?[*]
*[[This sort of treatment of phenomenal belief is suggested by Francescotti (1994).]]
The analogy does not go through, however. Or rather, it goes through only to a limited extent. Oscar and Twin Oscar's water concepts here are analogous to Mary and Inverted Mary's relational phenomenal concepts (red_C or red_I), or perhaps to their demonstrative concepts. For example, the relational concepts that they express with their public language expressions 'red experience' will refer to two different properties, phenomenal redness and phenomenal greenness. Mary and Inverted Mary can deploy these concepts in certain beliefs, such as the beliefs that they express by saying 'Tomatoes cause red experiences', even before they leave their monochromatic rooms for the first time. Because of the distinct referents of their concepts, there is a natural sense (Nida-Rumelin's "nonphenomenal" sense) in which we can say that Mary believed that tomatoes caused red experiences, while Inverted Mary did not; she believed that tomatoes caused green experiences. Here the analogy goes through straightforwardly.
The pure phenomenal concepts R and G, however, are less analogous to the two water concepts than to the chemical concepts H2O and XYZ. When Oscar learns the true nature of water, he acquires the new belief water = H2O, while Twin Oscar acquires an analogous belief involving XYZ. When Mary learns the true nature of red experiences, she acquires a new belief red_C = R, while Inverted Mary acquires an analogous belief involving G. That is, Mary and Inverted Mary's later knowledge involving R and G is fully lucid knowledge of the referents of the concepts in question, analogous to Oscar and Twin Oscar's knowledge involving the chemical concepts H2O and XYZ.
But here we see the strong disanalogy. Once Oscar acquires the chemical concept H2O and Twin Oscar acquires XYZ, they will no longer be twins: their functional properties will differ significantly. By contrast, at the corresponding point Mary and Inverted Mary are still twins. Even though Mary has the pure phenomenal concept R and Inverted Mary has G, their functional properties are just the same. So the difference between the concepts R and G across functional twins is something that has no counterpart in the standard Twin Earth story.
All this reflects the fact that in standard externalist cases, the pairs of corresponding concepts may differ in reference, but they have the same or similar epistemic or notional contents. Oscar and Twin Oscar's water concepts have different referents (H2O vs. XYZ), but they have the same epistemic contents: both intend to refer to roughly the liquid around them with certain superficial properties. Something like this applies to Mary's and Inverted Mary's relational phenomenal concepts, which have different referents but the same epistemic content (which picks out whatever quality stands in a certain relation), and to their demonstrative concepts (which pick out roughly whatever quality happens to be ostended).
In terms of the two-dimensional framework, where epistemic contents correspond to epistemic intensions: Oscar's and Twin Oscar's water concepts have the same epistemic intension but different subjunctive intensions. A similar pattern holds in all the cases characteristic of standard externalism. The pattern also holds for Mary's and Inverted Mary's relational phenomenal concepts, and their demonstrative phenomenal concepts.
But Mary's concept R and Twin Mary's concept G have different epistemic contents. In this way they are analogous to Oscar's concept H2O and Twin Oscar's concept XYZ. But again, the disanalogy is that R and G are possessed by twins, and H2O and XYZ are not. So the case of Inverted Mary yields an entirely different phenomenon: a case in which epistemic content differs between twins.
This can be illustrated by seeing how the concepts in question are used to constrain epistemic possibilities. When Oscar confidently believes that there is water in the glass, he is not thereby in a position to rule out the epistemic possibility that there is XYZ in the glass (unless he has some further knowledge, such as the knowledge that water is H2O). The same goes for Twin Oscar's corresponding belief. For both of them, it is equally epistemically possible that the glass contains H2O and that it contains XYZ. Any epistemic possibility compatible with Oscar's belief is also compatible with Twin Oscar's belief: in both cases, these will be roughly those epistemic possibilities in which a sample of the dominant watery stuff in the environment is in the glass.
Epistemic content reflects the way that a belief constrains the space of epistemic possibilities, so Oscar's and Twin Oscar's epistemic contents are the same. Something similar applies to Mary and Inverted Mary, at least where their pairwise relational and demonstrative phenomenal concepts are concerned. When Mary confidently believes (under her relational concept) that her mother is having a red experience, for example, she is not thereby in a position to rule out the epistemic possibility that her mother is having an experience with the quality G. Both Mary's and Inverted Mary's beliefs are compatible with any epistemic possibility in which the subject's mother is having the sort of experience typically caused in the community by paradigmatic red objects. So their beliefs have the same epistemic contents.
But Mary's and Inverted Mary's pure phenomenal concepts do not work like this. Mary's concept R and Inverted Mary's concept G differ not just in their referents but in their epistemic contents. When Mary leaves the monochromatic room and acquires the confident belief (under her pure phenomenal concept) that tomatoes cause red experiences, she is thereby in a position to rule out the epistemic possibility that tomatoes cause experiences with quality G. The only epistemic possibilities compatible with her belief are those in which tomatoes cause R experiences. For Inverted Mary, things are reversed: the only epistemic possibilities compatible with her belief are those in which tomatoes cause G experiences. So their epistemic contents are quite different.
Again, the epistemic situation with R and G is analogous to the epistemic situation with the concepts H2O and XYZ. When Oscar believes (under a fully lucid chemical concept) that the glass contains H2O, he is thereby in a position to rule out all epistemic possibilities in which the glass contains XYZ. For Twin Oscar, things are reversed. This is to say that H2O and XYZ have different epistemic contents. The same goes for R and G.
So in the case of the pure phenomenal concepts, uniquely, we have a situation in which two concepts differ in their epistemic content despite the subjects being physically identical. So phenomenal concepts seem to uniquely give a case in which epistemic content is not conceptually supervenient on the physical.
Using the two-dimensional framework: the epistemic intension of a concept reflects the way it applies to epistemic possibilities. We saw above that the epistemic intensions of Oscar's and Twin Oscar's water concepts are the same, as are the epistemic intensions of Mary's and Inverted Mary's relational and demonstrative phenomenal concepts. But R and G differ in the way they apply to epistemic possibilities, and their epistemic intensions differ accordingly: the epistemic intension of R picks out phenomenal redness in all worlds, and the epistemic intension of G picks out phenomenal greenness in all worlds. When Mary thinks I am having an R experience now, the epistemic intension of her thought is true at all and only those worlds in which the being at the center is having an R experience.
Something very unusual is going on here. In standard externalism, and in standard cases of so-called "direct reference", a referent plays a role in constituting the subjunctive content (subjunctive intension) of concepts and beliefs, while leaving the epistemic content (epistemic intension) unaffected. In the pure phenomenal case, by contrast, the quality of the experiences plays a role in constituting the epistemic content of the concept and of the corresponding belief. One might say very loosely that in this case, the referent of the concept is somehow present inside the concept's sense, in a way much stronger than in the usual cases of "direct reference".[*]
*[[Martine Nida-Rumelin suggests in a forthcoming paper that we can call this sort of concept a super-rigid concept. A rigid concept is one whose subjunctive intension picks out the same referent at all worlds; a super-rigid concept is a rigid concept whose epistemic intension picks out the same referent at all worlds.]]
This analysis of phenomenal concepts bears on the knowledge argument in a numbre of ways. For a start, it gives us a better characterization of Mary's new knowledge. The crucial new beliefs that Mary gains do not involve just any phenomenal concept that refers to phenomenal redness, but rather involve a pure phenomenal concept such as R. These beliefs have the form this experience is R, I am experiencing R, red things typically cause R experiences, other people experience R when they look at tomatoes, R is instantiated, and so on. The content of R is tied to the phenomenal property R in a very direct way: both the epistemic and subjunctive intensions of the concept pick out instances of R in all possible worlds. And this content appears to be determined directly by the instantiation of R itself in Mary: if Inverted Mary instantiates G instead, she will have a quite different pure phenomenal concept G, quite different resulting beliefs.
Nothing I have said so far entails that the knowledge argument is sound. So far, what I have said can be embraced in principle by a type-B materialist who holds that phenomenal properties are identical to physical properties, but that phenomenal concepts are distinct from physical concepts. The type-B materialist can take on board everything so far as a point about the distinctive behavior of phenomenal concepts, and hold that no non-materialist ontological consequences follow.
Nevertheless, what I have said so far is already enough to see that certain materialist responses to the knowledge argument fail. It is common for materialists to respond to the knowledge argument by invoking specific claims about the nature of phenomenal concepts, or analogies with other concepts. For example, some (e.g. Ismael 1999, Perry 2001) hold that phenomenal concepts are indexical concepts, so that the epistemic gap here can be assimilated to the epistemic gap between objective knowledge and indexical knowledge more generally. Others (e.g. Hawthorne 2001, Loar 1997) hold that phenomenal concepts are demonstrative or recognitional concepts, so that the epistemic gap can be assimilated to that that holds between theoretical knowledge and demonstrative or recognitional knowledge more generally.
>From the discussion above, it is clear that Mary's crucial phenomenal concept is not an indexical or a demonstrative concept. Mary does have a concept of this form: it is her demonstrative phenomenal concept E. That concept behaves roughly as the accounts above suggest that phenomenal concepts should. But Mary's important new knowledge involves not E but R, and R is not an indexical or demonstrative concept at all. Rather, it is a pure phenomenal concept, a concept characterizing the phenomenal property in question directly in terms of its phenomenal character.
One can also make a direct case against any analysis of phenomenal knowledge as indexical or demonstrative knowledge, as follows. In the indexical case, any epistemic gaps disappear from an objective perspective. Say that I am physically omniscient, but do not know whether I am in the USA or Australia (let's imagine that there are appropriate qualitative twins in both). Then I have a certain indexical ignorance, and discovering that I am in the USA will constitute new knowledge. But if someone else is watching from the third-person point of view and is also physically omniscient, they will have no corresponding ignorance: they will know that A is in Australia and that B is in the USA, and that is all. There is no potential knowledge that they lack: from their perspective, they know everything there is to know about my situation. So my ignorance is essentially indexical, and evaporates from the objective viewpoint. The same goes for indexical ignorance concerning what time it is, for demonstrative ignorance concerning what this is, and so on. In all these cases, the ignorance disappears from the objective viewpoint: an objectively omniscient observer can know everything there is for them to know about my situation, and there will be no doubts for them to settle.
Now consider Mary's ignorance. From her black-and-white room, she is ignorant of all sorts of facts: what it will be like for her to see red for the first time, what it is like for others to see red, and so on. Only the first of these looks even apparently indexical, so let us focus on that. In this case, a physically omniscient observer may have precisely analogous ignorance: even given his complete physical knowledge, he may have no idea what it will be like for Mary to see red for the first time. So this ignorance does not evaporate from the objctive viewpoint. The same goes even more strongly for knowledge of what it is like for others to see red. For any observer, regardless of their viewpoint, there will be an epistemic gap between complete physical knowledge and this sort of phenomenal knowledge. This suggests very strongly that phenomenal knowledge is not a variety of indexical or demonstrative knowledge at all. Rather, it is a sort of objective knowledge of the world, not essentially tied to any viewpoint.
If this is right, then any analysis of phenomenal concepts as indexical or demonstrative concepts fails, and any attempt to explain Mary's epistemic gap in terms of the epistemic gap for indexical or demonstrative concepts fails.
There is more to say here about the knowledge argument here. But to explore these issues, it is useful to first set out the two-dimensional analysis of the knowledge argument.
In my view, the knowledge argument is most powerful when augmented by the two-dimensional semantic framework. The two-dimensional framework draws a connection betweeen the epistemic and metaphysical modalities, and allowing one to draw ontological conclusions from epistemic premises.
For current purposes, the crucial claim of the two-dimensional framework is the following: when a sentence S is a posteriori, S has a metaphysically contingent epistemic intension. That is, the epistemic intension of S is false in some centered metaphysically possible world. This applies, for example, to all standard Kripkean a posteriori necessities. The epistemic intension of 'water is H2O' is false in a Twin Earth centered world. The epistemic intension of 'Hesperus is Phosphorus' is false in centered world where the evening star near the center is distinct from the morning star near the center. And so on. All these worlds are metaphysically poaaible. (From now on, talk about worlds should be interpreted as invoking metaphysically possible worlds, not merely epistemically possible scenarios.) The claims above are quite compatible with Kripke's claim that these sentences are necessary: Kripke's claim is in effect the claim that the sentences have necessary subjunctive intensions.
A first approximation to a two-dimensional analysis of the knowledge argument can be put as follows. Let P be the complete microphysical truth about the world; let T be a statement asserting that the world is a minimal world satisfying P; let I be a conjunction of indexical truths about a speaker's identity ('I am A', where A is an identifying description relative to the information in P&T), the current time ('now is B'), and any other essentially indexical claims that are not settled by objective information and the previous indexical claims. Let PTI be the conjunction of these claims. Let Q be a phenomenal truth involving a pure phenomenal concept. 'PTI -> Q' is a material conditional from PTI to Q.
(1) 'PTI -> Q' is a posteriori.
(2) If 'PTI->Q' is a posteriori, 'PTI->Q' has a metaphysically contingent epistemic intension.
(3) If 'PTI->Q' has a metaphysically contingent epistemic intension, then either 'PT->Q' has a metaphysically contingent subjunctive intension or panprotopsychism is true.
(4) If 'PT->Q' has a metaphysically contingent subjunctive intension, materialism is false.
----
(5) Materialism is false or panprotopsychism is true.
The argument is valid. The first premise is a statement of phenomenal realism, corresponding to the claim that Mary gains new knowledge of truths not deducible from physical truths and indexical truths. The second premise is an instance of the general two-dimensional claim above. The fourth premise states a straightforward commitment of materialism, given that P states the complete microphysical truth about the world and Q is a truth.
The third premise requires discussion. One would be simpler to assert that if 'PTI->Q' has a contingent epistemic intension, it has a contingent subjunctive intension. But here loopholes arise from the possibility that the expressions involved in 'PTI->Q' are have distinct epistemic and subjunctive intensions. This is not a problem for Q, since we have seen that pure phenomenal concepts have the same epistemic and subjunctive intensions (modulo indexicality). The same goes for T. But there remain worries about I (and associated worries about indexicality) and about P.
On the first of these points: an epistemic intension is indexical (a function over centered worlds) where a subjunctive intension is not. But if 'PTI->Q' is false at a centered world, we can say nonindexically that (PT->Q)@I is false at that world, where this is the uncentered intension evaluated in an uncentered world by first picking out a center according to the identifying descriptions in I, and evaluating the epistemic intension of 'PT->Q' there. From here we can infer that the subjunctive intension of 'PT->Q' is false at that world, subject only to worries about the epistemic and subjunctive intension of P.
On this second point, the epistemic and subjunctive intensions of microphysical concepts may differ in that the former plausibly pick out whatever properties play a certain causal role in a world, while the latter arguably pick out in all worlds the intrinsic properties that play this role in the actual world. So even if 'PTI->Q' does not have a necessary epistemic intension, 'PT->Q' could still have a necessary subjunctive intension, as long as the underlying intrinsic properties in our world necessitate Q but the causal structure of the world does not. This is the Russellian position that I have elsewhere called "panprotopsychism" (see also Maxwell 1979, Lockwood 1989). This position is not ruled out by the knowledge argument (see Stoljar 2001 here). It is arguable whether or not this is a form of materialism, but it is clearly a distinctive and unusual form.
These two points together establish the third premise. The only way to deny that 'PT->Q' has a necessary subjunctive intension, given that 'PTI->Q' has a contingent epistemic intension, is to embrace panprotopsychism. Given the plausibility of the other three premises, the conclusion is established.
We can use the argument above to analyze various responses that have been made to the knowledge argument.
(i) The ability analysis: According to this response (Lewis 1991, Nemirow 1991), Mary does not gain new factual knowledge, but merely gains an ability. Proponents of this response will deny that there are truths that Mary cannot know from in her room, and so will deny either premise (1) or the claim that Q is a truth. The same goes for other positions (e.g. Dennett 1992) according to which Mary gains no factual knowledge. The analysis above has no special force against this position, as the discussion here takes for granted that Mary gains new factual knowledge. Any reasons for rejecting these positions rest on separate grounds.
(ii) The indexical analysis: According to this analysis (Bigelow and Pargetter, Ismael, Perry), Mary's new knowledge is likened to indexical knowledge. Proponents of this position will presumably also deny premise (1) above. They will accept that 'PT->Q' is a posteriori, but deny that 'PTI->Q' is a posteriori; they hold that Q is itself indexical knowledge, so if PTI contains full indexical knowledge it will entail Q. (Mere indexical knowledge of her identity and the current time will obviously not help Mary to know what it is like to see red, but a proponent might suggest that I build in further indexical claims, perhaps identifying the referent of indexical phenomenal concepts such as E.) This positions can be rejected for the reasons discussed earlier: although it is plausible that there are indexical phenomenal concepts, Q does not involve any such concepts: pure phenomenal concepts are cognitively distinct from indexical concepts and do not exhibit the epistemic behavior of indexicals.
(iii) The old-fact/new-way analysis: According to the most popular response to the knowledge argument, Mary gains knowledge of a fact she already knew, under a different mode of presentation. In the standard version, this is held to be analogous to someone who knew Hesperus is a planet and who learns that Phosphorus is a planet, or someone who knew that Superman can fly and who learns that Clark Kant can fly. Each pair of items of knowledge arguably involves a single fact (about a single property instantiated by a single individual), under distinct modes of presentation. Proponents of this response hold that analogously, Mary's new phenomenal knowledge is knowledge of a fact she knew already (about the instantiation of a physical property), nuder a different mode of presentation. Mary has distinct physical and phenomenal concepts with a common referent, just as with Hesperus and Phosphorus, or Superman and Clark Kent.
As it stands, the standard response is unable to deal with the argument above. Nothing in the standard cases involving Hesperus, Superman, and the like gives any reason to deny any of the premises of the argument. In particular, none of these cases give any reason to deny premise 2. A posteriori identities such as Hesperus is Phosphorus have necessary subjunctive intensions, but they have contingent epistemic intensions, just as the premise predicts. So the argument shows that this response is inadequate.
One can put the moral without invoking the two-dimensional framework. The basic point is that where there are distinct modes of presentation of a common referent, the distinct modes of presentation are associated with distinct properties of the individual. (The exception is cases involving indexicality, but we have seen that indexicality is irrelevant here.) As such, when one gains new knowledge equating the referents of the two concepts, one gains knowledge of new facts connecting the modes of presentation. In particular, one gains knowledge that a single individual has both associated properties. For example, when one learns that Hesperus is Phosphorus, one learns the new fact that the brightest object visible in the evening is also visible in the morning. When one learns that Superman is Clark Kent, one learns the new fact that the guy with the cape works at the Daily Planet. And so on. Note that this does not require the descriptivist claim that the associated properties are used to fix reference; the general point is independently plausible in any given case. (The one exception involves cases of indexicality, but we have seen that indexicality is irrelevant to Mary's case.) Applying this pattern to Mary: even if Mary acquires knowledge of an old fact under a new mode of presentation, she will also acquire knowledge of a new fact connecting those modes of presentation. So the physical facts do not exhaust all the facts.
(iv) Loar's analysis. For an old-fact/new-way analysis to have any hope of succeeding, it must treat Mary's new knowledge as disanalogous to standard cases of coreference, and in particular it must give reason to think that premise 2 fails in this case. This is an uphill battle, since it is plausible that premise 2 holds in all familiar cases: in particular, all familiar a posteriori identities appear to involve distinct epistemic intensions over centered worlds.
A sophisticated attempt in this diretcion is made by Loar (1997). Loar isolates the "semantic premise" of the knowledge argument:
(Semantic premise) A statement of property identity that links conceptually independent concepts is true only if at least one concepts picks out the property it refers to by connoting a contingent property of that property. (Loar 1997, p. 600):
Here, a concept connotes a property when it uses that property to pick out its referent. This is very much akin to a property associated with a mode of presentation in the sense above. If we make the translation, the claims above can be seen as closely akin to the semantic premise: on view above, the new fact connecting two modes of presentation of an individual is plausibly contingent, so at least one of the properties ascribed to that individual must be possessed contingently. The semantic premise is also closely related to premise 2 of the formal argument above (where connoted properties map onto epistemic intensions across possible worlds). There are some minor differences, but these do not matter too much here.
Loar allows that the semantic premise applies to standard a posteriori identities, but denies that it holds in the physical-phenomenal case. He suggests (p. 602) that this can be explained by the fact that phenomenal concepts are recognitional concepts that pick out physical properties without a contingent associated mode of presentation. Here, a recognitional concept is a type-demonstrative: "one of that kind". And a concept that lacks a contingent mode of presentation can be seen as one with the same epistemic and subjunctive intensions (Loar 1999 makes the latter claim explicitly). Loar suggests that the fact that phenomenal concepts are recognitional concepts explains the aposteriority of a phenomenal-physical identity: identities involving recognitional and theoretical concepts are always a posteriori. And he suggests that the lack of a contingent associated mode of presentation explains why the semantic minor premise fails in this case, so that no ontologically distinct properties are involved in the identity, and the epistemic gap is compatible with physicalism.
I think there is good reason to reject this account.[*] Most importantly, as we have seen earlier, there is good reason to believe that the phenomenal concepts crucial to Mary's argument are not type-demonstratives. Mary's demonstrative phenomenal concepts are type-demonstratives, but her pure phenomenal concepts are quite distinct from these, and it is pure phenomenal concepts that are crucial to the argument. So an account based on an appeal to type-demonstratives cannot succeed here.
*[[I discuss some related problems for Loar's account in Chalmers 1999.]]
Second, one can argue that if phenomenal concepts were recognitional concepts (type-demonstratives), they could not have distinct epistemic and subjunctive intensions (i.e., they could not pick out their properties by a contingent mode of presentation). All demonstrative concepts have distinct epistemic and subjunctive intensions, and no non-circular reason has been given for supposing that phenomenal concepts could be different. Anticipating such a response (p. 602), Loar protests that antiphysicalists themselves hold that phenomenal concepts lack a contingent associated mode of presentation (they have the same epistemic and subjunctive intension), and the physicalist is entitled to the same claim. That seems correct, but it does nothing at all to resolve the conflict between the two claims that phenomenal concepts are recognitional and that their intensions coincide. Insofar as we have reason to believe that the intensions coincide, we also have reason to believe that phenomenal concepts are not recognitional. The arguments for the distinctness of pure and demonstrative phenomenal concepts earlier in this paper rested on largely the same considerations as the arguments for the coincident intensions of pure phenomenal concepts. Of course there are recognitional phenomenal concepts, and there are phenomenal concepts that pick out their referents without contingent modes of presentation, but there is no reason to think that these classes overlap.
(Loar says (p. 602) "On both metaphysical views, phenomenal concepts differ from other recognitional concepts; phenomenal concepts are a peculiar sort of recognitional concept on any account, and that can hardly count against physicalism." But this is incorrect: the nonphysicalist should simply deny that phenomenal concepts are recognitional concepts.)
Loar might suggest that he is using "recognitional concept" and "type-demonstrative" more broadly than I am, so that what I am calling pure phenomenal concepts also count as recognitional or as demonstrative. But in this case, analogies with standard recognitional and demonstrative concepts can no longer do any work. In particular, one can no longer appeal to the fact that recognitional concepts typically corefer with theoretical concepts even when they are cognitively distinct. This feature arises from the distinctive epistemic and semantic character (indexicality, distinct epistemic and subjunctive intensions) of standard recognitional concepts, and the considerations do not transfer to pure phenomenal concepts. Once again, the distinctive nature of pure phenomenal concepts (nonindexicality, coincident epistemic and subjunctive intensions) stands in strong tension with the feature in question. In every other case, concepts of this sort do not support substantive a posteriori necessities (setting aside any slack due to distinct intensions for theoretical concepts, which leads only to panprotopsychism), and no reason has been given to think that pure phenomenal concepts are any different.
I conclude that Loar's attempt to reconcile the distinctive epistemic behavior of phenomenal concepts with physicalism fails. There is no reason to believe that the relevant phenomenal concepts have the features he suggests, and there is good reason to deny this claim.
(5) Other analyses. The materialist might hope that some other account of phenomenal concepts can be given, such that their distinctive epistemic behavior can be reconciled with their reference to a physical property. I am skeptical that any such account can be given. I think the only remote chance is to attempt to deny premise 2 (and related claims about properties associated with modes of presentation), but as I have argued in Chalmers (2002c), there is reason to believe that this premise is a deep (nontrivial) conceptual truth. On my view, all other accounts that attempt to deny this premise fall prey to problems that are akin to the problem of Loar's account.
Nevertheless, there is room for fruitful further debate on this topic. And whatever the consequences for the truth of materialism, a deeper understanding of phenomenal concepts is likely to have deep consequences for our understanding of consciousness.
Austin, D.F. 1990. What's the Meaning of "This"? Ithaca, NY: Cornell University Press.
Chalmers, D.J. 1996. The Conscious Mind: In Search of a Fundamental Theory. Oxford University Press.
Chalmers, D.J. 2002a. Consciousness and its place in nature. In (S. Stich & F. Warfield, eds) The Blackwell Guide to the Philosophy of Mind. Blackwell. [consc.net/papers/nature.html]
Chalmers, D.J. 2002b. The components of content. In (D. Chalmers, ed) The Philosophy of Mind: Classical and Contemporary Readings. Oxford University Press. [consc.net/papers/content.html]
Chalmers, D.J. 2002c. The content and epistemology of phenomenal belife. In (Q. Smith & A. Jokic, eds) Consciousness: New Philosophical Essays. Oxford University Press.
Chalmers, D.J. (forthcoming). The nature of epistemic space. [consc.net/papers/espace.html]
Chisholm, R. 1957. Perceiving: A Philosophical Study.
Francescotti, R.M. 1994. Qualitative beliefs, wide content, and wide behavior. Nous 28:396-404.
Hawthorne, J. (forthcoming). Advice to physicalists. Philosophical Studies.
Ismael, J. 1999. Science and the phenomenal. Philosophy of Science.
Jackson, F. 1982. Epiphenomenal qualia. Philosophical Quarterly 32:127-136.
Kaplan, D. 1989. Demonstratives. In (J. Almog, J. Perry, and H. Wettstein, ed.) Themes from Kaplan. New York: Oxford University Press.
Loar, B. 1997. Phenomenal states (second version). In (N. Block, O. Flanagan, & G. Güzeldere, eds) The Nature of Consciousness. MIT Press.
Nida-Rumelin, M. 1995. What Mary couldn't know: Belief about phenomenal states. In (T. Metzinger, ed) Conscious Experience. Ferdinand Schoningh.
Nida-Rumelin, M. 1996. Pseudonormal vision: An actual case of qualia inversion? Philosophical Studies 82:145-57.
Nida-Rumelin, M. 1997. On belief about experiences: An epistemological distinction applied to the knowledge argument. Philosophy and Phenomenological Research.
Perry, J. 2001. Knowledge, Possibility, and Consciousness. MIT Press.
What follows is a brief and simplified introduction to the two-dimensional semantic framework as I understand it. See also Chalmers (2002b; forthcoming).
Let us say that it is epistemically possible in the broad sense that S if the hypothesis that S is not ruled out a priori. Then there will be a wide space of epistemic possible hypotheses (in the broad sense; I omit the qualifer in what follows). Some of these will conflict with each other; some of them will be compatible with each other; and some will subsume each other. We have a systematic way of evaluating and describing epistemic possibilities that differs from our way of evaluating and describing subjunctive counterfactual possibilities. It is this sort of evaluation and description that is captured by the first dimension of the two-dimensional framework.
It is epistemically possible that water is not H2O, in the broad sense that this is not ruled out a priori. And there are many specific versions of this epistemic possibility: intuitively, specific ways our world could turn out such that if they turn out that way, it will turn out that water is not H2O. Take the XYZ-world, one containing superficially identical XYZ in place of H2O. It is epistemically possible that our world is the XYZ-world. When we consider this epistemic possibility - that is, when we consider the hypothesis that our world contains XYZ in the oceans, and so on - then this epistemic possibility can be seen as an instance of the epistemic possibility that water is not H2O. We can rationally say "if our world turns out to have XYZ in the oceans (etc.), it will turn out that water is not H2O". The hypothesis that the XYZ-world is actually rationally entails the belief that water is not H2O, and is rationally inconsistent with the belief that water is H2O.
Here, as with subjunctive counterfactual evaluation, we are considering and describing a world, but we are considering and describing it in a different way. In the epistemic case, we consider a world as actual: that is, we consider the hypothesis that our world is that world. In the subjunctive case, we consider a world as counterfactual: that is, we consider it as a way things might have been, but (probably) are not. These two modes of consideration of a world yield two ways in which a world might be seen to satisfy a sentence or a belief. When the XYZ-world is considered as actual, it satisfies 'water is XYZ'; when it is considered as counterfactual, it does not.
In considering a world as actual, we ask ourselves: what if the actual world is really that way? In the broad sense, it is epistemically possible that Hesperus is not Phosphorus. This is mirrored by the fact that there are specific epistemic possibilities (not ruled out a priori) in which the heavenly bodies visible in the morning and evening are distinct; and upon consideration, such epistemic possibilities are revealed as instances of the epistemic possibility that Hesperus is not Phosphorus.
When we consider worlds as counterfactual, we consider and evaluate them in the way that we consider and evaluate subjunctive counterfactual possibilities. That is, we acknowledge that the character of the actual world is fixed, and say to ourselves: what is the world had been such-and-such a way? When we consider the counterfactual hypothesis that the morning star might have been distinct from the evening star, we conclude not that Hesperus would not have been Phosphorus, but rather that at least one of the objects is distinct from both Hesperus and Phosphorus (at least if we take for granted the actual-world knowledge that Hesperus is Phosphorus, and if we accept Kripke's intuitions).
Given a statement S and a world W, the epistemic intension of S returns the truth-value of S in W considered as actual. (Test: is the epistemic possibility that W is actual an instance of the epistemic possibility that S?) The subjunctive intension of S returns the truth-value of S in W considered as counterfactual. (Test: if W had obtained, would S would have been the case?.) We can then say that S is primarily possible (or 1-possible) if its epistemic intension is true in some world (i.e. if it is true in some world considered as actual), and that S is secondarily possible (or 2-possible) if its subjunctive intension is true in some world (i.e. if it is true in some world considered as counterfactual). Primary and secondary necessity can be defined analogously.
For a world to be considered as actual, it must be a centered world -- a world marked with a specified individual and time - as an epistemic possibility is not complete until one's "viewpoint" is specified. So a epistemic intension should be seen as a function from centered world to truth-values. For example, the epistemic intension of 'I' picks out the individual at the center of a centered world; and the epistemic intension of 'water' picks out, very roughly, the clear drinkable (etc.) liquid in the vicinity of the center. No such marking of a center is required for considering a world as counterfactual, or for evaluating subjunctive intensions.
Epistemic and subjunctive intensions can be associated with statements in language, as above, and equally with singular terms and property terms. The intension of a statement will be a function from worlds to truth-values; the intension of a term will be a function from worlds to individuals or properties within those worlds. (In some cases, intensions are best associated with linguistic tokens rather than types.)
Epistemic intensions can also be associated in much the same way with the (token) concepts and thoughts of a thinker, all of which can be used to describe and evaluate epistemic possibilities as well as subjunctive counterfactual possibilities. In "The Components of Content" I argue that the epistemic intension of a concept or a thought can be seen as its "epistemic content" (a sort of internal, cognitive content), and that the subjunctive intension captures much of what is often called "wide content".
A crucial property of epistemic content is that it reflects the rational relations between thoughts. In particular, if a belief A entails a belief B by a priori reasoning, then it will be epistemically impossible (in the broad sense) for A to be true without B being true, so the epistemic intension of A entails the epistemic intension of B. Further, if an identity a=b is a posteriori for a subject, then it is epistemically possible for the subject that the identity is false, and there will be an epistemic possibility in which the referents of the two concepts involved differ, so the subject's concepts a and b will have distinct epistemic intensions. So epistemic intensions behave something like Fregean senses, individuating concepts according to cognitive significance at least up to the level of a priori equivalence.
Insofar as the two-dimensional framework is used in this paper, it is being applied rather than being discussed or justified in its own right. The discussion here indicates important distinctions among phenomenal concepts whose analysis requires the idea of epistemic content. And importantly, there are epistemological distinctions that turn on these distinctions in content. This reflects a more general phenomenon: the sort of possibility that is most crucial in epistemology is epistemic possibility, and the sort of content that is correspondingly most crucial is epistemic content.