From The
Journal of Philosophy XCVII, 7 (2000): 365-386.
Permission to publish this material on this site is
granted to The Philosopher’s Annual
by The Journal of Philosophy, Inc.,
which retains copyright.
Visit The
Journal of Philosophy at:
10
Evolution and the Problem of
Other Minds*
1
The following diagram illustrates two inference problems. First, there is the strictly third-person behavior-to-mind problem[1] in which I observe your behavior and infer that you occupy some mental state. Second, there is the self-to-other problem, in which I notice that I always or usually occupy some mental state when I behave in a particular way; then, when I observe you produce the same behavior, I infer that you occupy the same mental state. This second inference is the subject of the traditional philosophical problem of other minds.
How are these inferences related? Notice that the inputs to behavior-to-mind inference are a subset of the inputs to self-to-other inference. In the first, I consider the behavior of the other individual; in the second, I consider that behavior, as well as my own behavior and mental state. This suggests that if self-to-other inferences cannot be drawn because the evidence available is too meager, the same will be true of behavior-to-mind inferences as well.
I see no reason to think that a purely third-person scientific psychology is impossible. Of course, if one is a skeptic about all nondeductive inference, that skepticism will infect the subject matter of psychology. And if one thinks that scientific inference can never discriminate between empirically equivalent hypotheses, one also will hold that science is incapable of discriminating between such hypotheses when their subject matter is psychological. However, these are not special problems about psychology. What, then, becomes of the problem of other minds? If that problem concerns the tenability of self-to-other inference, it appears to be no problem, if science is able to draw behavior-to-mind inferences.
To see how the self-to-other problem can be detached from the problem of strictly third-person behavior-to-mind inference, we need to distinguish absolute from incremental versions of the self-to-other problem. The absolute problem concerns whether certain input information permits me to infer that the other person occupies mental state M rather than some alternative state A. As I have said, if information about the behavior of others permits me to infer that they are in mental state M, then it is hard to see why this inference should be undermined by adding the premiss that I myself am in mental state M when I produce behavior B. However, the question remains of whether first-person information makes a difference. To whatever degree third-person information provides an indication of whether the other person has M or A, does the addition of first-person information modify this assessment? This incremental version of the self-to-other problem is neutral on the question of whether third-person behavior-to-mind inference is possible. It is this incremental problem that I think forms the core of the problem of other minds, and this is the problem that I want to address here.[2]
Although the problem of other minds usually begins with an introspective grasp of one=s own mental state, it can be detached from that setting and formulated more generally as a problem about ‘extrapolation.’ Thus, we might begin with the assumption that human beings produce certain behaviors because they occupy particular mental states and ask whether this licenses the conclusion that members of other species that exhibit the behavior do so for the same reason. However, none of us knows just by introspection that all human beings who produce a given behavior do so because they occupy some particular mental state. In fact, this formulation of the problem of other minds, in which it is detached from the concept of introspection, is usually what leads philosophers to conclude that inferences about other minds from one=s own case are weak. The fact that I own a purple bow tie should not lead me to conclude that you do too. This point about bow ties is supposed to carry over to the fact that I have a mind and occupy various mental states. I know that I own a purple bow tie, but not by introspection.[3]
Discussion of the problem of other minds in philosophy seems to have died down (if not out) around thirty years ago.[4] Before then, it was discussed as an instance of ‘analogical’ or ‘inductive’ reasoning and the standard objection was that an inference about others based on your own situation is an extrapolation from too small a sample. This problem does not disappear merely by thinking of introspective experience as furnishing you with thousands of data points. The fact remains that they all were drawn from the same urn—your own mind. How can sampling from one urn help you infer the composition of another?
When the problem is formulated in this way, it becomes
pretty clear that what is needed is some basic guidance about inductive
inference. The mental content of self-to-other inference is not what
makes it problematic, but the fact that it involves extrapolation. Some
extrapolations make sense while others do not. It seems sensible to say that
thirst makes other people drink water, based on the fact that this is usually
why I drink. Yet, it seems silly to say that other folks walk down
The first step towards answering this question started to emerge in the 1960=s, not in philosophy of mind, but in philosophy of science. There is a general point about confirmation that we need to take to heart—observations provide evidence for or against a hypothesis only in the context of a set of background assumptions. If the observations do not deductively entail that the hypothesis of interest is true, or that it is false, then there is no saying whether the observations confirm or disconfirm, until further assumptions are put on the table. This may sound like the Duhem/Quine thesis, but that way of thinking about the present point is somewhat misleading, since Duhem[5] and Quine[6] discussed deductive, not probabilistic, connections of hypotheses to observations. If we want a person to pin this thesis to, it should be I.J. Good. Good[7] made this point forcefully in connection with Hempel=s formulation of the ravens paradox.[8] Hempel thought it was clear that black ravens and white shoes both confirm the generalization that all ravens are black. The question that interested him was why one should think that black ravens provide stronger confirmation than white shoes. Good responded by showing that empirical background knowledge can have the consequence that black ravens actually disconfirm the generalization. Hempel replied by granting that one could have special information that would undercut the assumption that black ravens and white shoes both confirm. However, he thought that in the absence of such information, it was a matter of logic that black ravens and white shoes are confirmatory. For this reason, Hempel asked the reader to indulge in a ‘methodological fiction.’ We are to imagine that we know nothing at all about the world, but are able to ascertain of individual objects what colors they have and whether they are ravens. We then are presented with a black raven and a white shoe; logic alone is supposed to tell us that these objects are confirming instances. Good=s response was that a person in the circumstance described would not be able to say anything about the evidential meaning of the observations. I think that subsequent work on the concept of evidence, both in philosophy and in statistics, has made it abundantly clear that Good was right and Hempel was wrong. Confirmation is not a two-place relationship between observations and a hypothesis; it is a three-place relation between observations, a hypothesis, and a background theory.[9] Black ravens confirm the generalization that all ravens are black, given some background assumptions, but fail to do so, given others. And if no background assumptions can be brought to bear, the only thing one can say is—out of nothing, nothing comes.
If we apply this lesson to the problem of other minds, we obtain the following result: The fact that I usually or always have mental property M when I perform behavior B is, by itself, no reason at all to think that you usually or always have M when you perform B. If that sounds like skepticism, so be it. However, the inference from Self to Other can make sense, once additional background assumptions are stated. A nonskeptical solution to the problem of other minds, therefore, must identify plausible further assumptions that bridge the inferential gap between Self and Other.[10]
1
The problem of other minds has been and continues to be important in psychology (actually, in comparative psychology), except that there it is formulated in the first person plural. Suppose that when human beings perform behavior B, we usually or always do so because we have mental property M. When we observe behavior B in another species, should we take the human case to count as evidence that this species also has mental property M? For this problem to be nontrivial, we assume that there is at least one alternative internal mechanism, A, which also could lead organisms to produce the behavior. Is the fact that humans have M evidence that this other species has M rather than A?
Discussion of this problem in comparative psychology has long been dominated by the fear of naive anthropomorphism.[11] This attitude was crystalized by C. Lloyd Morgan,[12] who suggested that if we can explain a nonhuman organism=s behavior by attributing to it a ‘higher’ mental faculty, or by assigning it a ‘lower’ mental faculty, then we should prefer the latter explanation. Morgan’s successors embraced this ‘canon’ because of its prophylactic qualities—it reduces the chance of a certain type of error. However, it is important to recognize that there are two types of error that might occur in this situation:
|
|
O lacks M |
O has M |
|
Deny that O has M |
|
type-2 error |
|
Affirm that O has M |
type-1 error |
|
Morgan=s canon does
reduce the chance of type-1 error, but that is not enough to justify the canon.
By the same token, a principle that encouraged anthropomorphism would
reduce the chance of type-2 error, but that would not justify this liberal
principle, either. Morgan tried to give a deeper defense of his canon; he
thought he could justify it on the basis of
I now want to leave Morgan=s late-nineteenth century ideas about comparative psychology behind, and fast forward to the cladistic revolution in evolutionary biology that occurred in the 1970's and after. The point of interest here is the use of a principle of phylogenetic parsimony to infer phylogenetic relationships among species, based on data concerning their similarities and differences.[14] Although philosophers often say that parsimony is an ill-defined concept, its meaning in the context of the problem of phylogenetic inference is pretty clear. The hypotheses under consideration specify phylogenetic trees. The most parsimonious tree is the one that requires the smallest number of changes in character state in its interior to produce the observed distribution of characteristics across species at the tree=s tips.
Consider the problem of inferring how sparrows, robins, and crocodiles are related to each other. Two hypotheses that might be considered are depicted in Figure 2. The (SR)C hypothesis says that sparrows and robins have a common ancestor that is not an ancestor of crocs; the S(RC) hypothesis says that it is robins and crocs that are more closely related to each other than either is to sparrows. Now consider an observation—sparrows and robins have wings, while crocodiles do not. Which phylogenetic hypothesis is better supported by this observation?
If winglessness is the ancestral condition, then the (SR)C hypothesis is more parsimonious. This hypothesis can explain the data about tip taxa by postulating a single change in character state on the branch with a slash through it. The S(RC) hypothesis, on the other hand, must postulate at least two changes in character states to explain the data. The principle of cladistic parsimony says that the observations favor (SR)C over S(RC). Notice that the parsimoniousness of a hypothesis is assessed not by seeing how many changes it says actually occurred, but by seeing what the minimum number of changes is that the hypothesis requires. (SR)C is more parsimonious because it entails a lower minimum.[15]
Although cladistic parsimony first attracted the attention of biologists because it helps one infer genealogical relationships, there is a second type
Figure 2
of problem that parsimony allows one to address. If you use a set of traits to reconstruct the genealogy that connects several contemporaneous species (as in Figure 2), you can place a new set of traits on the tip species in that inferred tree and use parsimony to infer the character states of their ancestors. This inference is illustrated in Figure 3. Given the character states (1's and 0's) of tip species, the most parsimonious assignment of character states to interior nodes is the one shown.[16]
Parsimony is now regarded in evolutionary biology as a reasonable way to infer phylogenetic relationships. Whether it is the absolutely best method to use, in all circumstances, is rather more controversial. And the foundational assumptions that need to be in place for parsimony to make sense as an inferential criterion are also a matter of continuing investigation.
The reason I have explained the basic idea behind cladistic parsimony is that it applies to the problem of other minds, when Self and Other are genealogically related. Suppose both Self and Other are known to have behavioral characteristic B, and that Self is known to have mental characteristic M. The question is whether Other should be assigned M as well. As before, I will assume that M is sufficient for B, but not necessary (an alternative internal mechanism, A, also could produce B). The two hypotheses we need to consider are depicted in Figure 4. If the root of the tree has the characteristic not-B ( and so has neither-M-nor-A),[17] then the (Same) hypothesis is more parsimonious than the (Diff) hypothesis. It is consistent with (Same) that the postulated similarity linking Self and Other is a homology; it is possible that the most recent common ancestor of Self and Other had M, and that M was transmitted unchanged from this ancestor to the two descendants. The (Same) hypothesis, therefore, requires only a single change in character state, from neither-M-nor-A to M. In contrast, the (Diff) hypothesis requires at least two changes in character state.[18]
Figure 3
Frans De Waal[19] presents this cladistic argument in defense of the idea that parsimony favors anthropomorphism—we should prefer the hypothesis that other species have the same mental characteristics that we have when they exhibit the same behavior.[20] This parsimony argument goes against Morgan=s canon, just as Morgan foresaw. De Waal adds the reasonable proviso that the parsimony inference is strongest when Self and Other are closely related.
There is another proviso that needs to be added to this analysis, which is illustrated in Figure 5. As before, Self and Other are observed to have behavioral trait B. We know by assumption that Self has the mental trait M. The question, as before, is what we should infer about Other—does it have M or A? The new wrinkle is that there are additional species depicted in the tree, ones that are known to lack B. The genealogical relationships that connect these further species to Self and Other entail that the most parsimonious hypothesis is that B evolved twice. It now makes no difference in parsimony whether one thinks that Other has M or A. What this shows is that parsimony favors anthropomorphism about mentalistic properties only when the behaviors in question are thought to be homologous.
Figure 4
This point has implications about a kind of question that often arises in connection with sociobiology. Parsimony does not oblige us to think that ‘slave-making’ in social insects has the same psycho-social causes as slave-making in humans (or that rape in human beings has the same proximate mechanism as ‘rape’ in ducks). The behaviors are not homologous, so there is no argument from parsimony for thinking that the same proximate mechanisms are at work. This is a point in favor of the parsimony analysis—cladistic parsimony explains why certain types of implausible inference really are implausible. Parsimonious anthropomorphism is not the same as naive anthropomorphism.
Just as extrapolation from Self to Other can be undermined by behavioral information about other species, extrapolation also can be undermined by neurophysiological information about Self and Other themselves. If Self has mental state M by virtue of being in physical state P1, what are we to make of the discovery that Other has physical state P2 (where P1 and P2 are mutually exclusive)? If we accept functionalism=s assurance that mental state M is multiply realizable, this information does not entail that Other cannot be in mental state M; after all, P2 might just be a second supervenience base for M. On the other hand, P2 might be a supervenience base for alternative state A, and not for M at all. This inference problem is illustrated in Figure 6. For simplicity, let us suppose that P1 and P2 each suffice for B. P1 produces B by way of mental state M; it is not known at the outset whether P2 produces B by way of mental state M, or by way of alternative state A. We assume that the ancestor at the root of the tree did not exhibit the behavior in question, and that this ancestor had trait P0.
Figure 5
The conclusion we must reach is that the two hypotheses in Figure 6 are equally parsimonious. (Same) says that Self and Other both have M and that P1 and P2 are two of M’s supervenience bases. (Same) requires two changes to account for the characteristics at the tips; either P1 or P2 evolved long ago from an ancestor who had P0, and then either P1 changed to P2, or P2 changed to P1, in a subsequent branch of the tree. Of course, not-B changed to B in the same branch in which P0 changed to either P1 or P2, but this does not count as two changes, since the first entails the second. The (Diff) hypothesis also requires two changes if the root of the tree=s having P0 is to be transformed into Self=s having P1 and Other=s having P2. Of course, (Diff) also will require changes from neither-M-nor-A at the root to M and A at the tips, but these do not count as changes additional to the ones that occur among P0, P1, and P2. Thus, the discovery of relevant neurophysiological differences can provide a context in which anthropomorphism is not sanctioned by parsimony considerations.
As a final exercise in understanding how the machinery of phylogenetic parsimony applies to the problem of extrapolating from Self to Other, let us consider the current controversy in cognitive ethology concerning whether chimps have a theory of mind.[21] It is assumed at the outset that chimps have what Dennett (op. cit.) terms first-order intentionality; they are able to formulate beliefs and desires about the extra-mental objects in their environment. The question under investigation is whether they, in addition, have second-order intentionality. Do they have the ability to formulate beliefs and desires about the mental states of Self and Other? Adult human beings have both first- and second-order intentionality. Do chimps have both, or do they have first-order intentionality only?
Figure 6
Figure 7 provides a cladistic representation of this question. I assume that the ancestral condition is the absence of both types of intentionality, and that second-order intentionality can evolve in a lineage only if first-order intentionality is already in place. The point to notice is that parsimony considerations do not discriminate between the two hypotheses. The (SAME) hypothesis and the (DIFF) hypothesis both require two changes in the tree=s interior—first- and second-order intentionality each must evolve at least once. The principle of parsimony in this instance tells us to be agnostic, and so disagrees with Morgan=s canon, which tells us to prefer (DIFF).
Why does the problem formulated in Figure 7 lead to a different conclusion from the problem represented in Figure 4 in terms of the two states M and A? In Figure 4, we find that parsimony considerations favor extrapolation; in Figure 7, we find that extrapolating is no more and no less parsimonious than not extrapolating. Notice that Figure 4 mentions the behavior B, which internal mechanisms M and A are each able to produce. Figure 7, however, makes no mention of the behaviors that first- and second-order intentionality underwrite. This is the key to understanding why the analyses come out differently.
The point behind Figure 4 is this: if two species exhibit a homologous behavior, each of them must also possess an internal mechanism (M or A) for producing the behavior. It is more parsimonious to attribute the same mechanism to both Self and Other than to attribute different mechanisms to each. This pattern of argument can be applied to the problem depicted in Figure 7 by making explicit the behavioral consequences that first- and second-order intentionality are supposed to have. Suppose that first-order intentionality allows the organisms that have it to exhibit a range of behaviors B1. When a species that has first-order intentionality evolves second-order intentionality, this presumably augments the behaviors that organisms in the species are able to produce; the repertoire expands to B1&B2.[22] If human beings exhibit B1 because they have first-order intentionality, we parsimoniously explain the fact that chimps also exhibit B1 (if they do) by saying that chimps have first-order intentionality. However, there is no additional gain in parsimony to be had from attributing second-order intentionality to them as well, unless we also observe them producing the behaviors in B2.[23]
Figure 7
Cognitive ethologists are trying to find behaviors that chimps will produce if they have second-order intentionality but will not produce if they possess first-order intentionality only.[24] Identifying behaviors of this sort would permit an empirical test to be run concerning whether chimps have a theory of mind. It certainly is desirable that such behaviors should be found. However, even if chimps exhibit B2, the question arises of why we should explain this by attributing second-order intentionality to them, rather than some alternative mechanism A. Of course, A will not be the trait of purely first-order intentionality, but presumably there are more than two options to be considered here. It is at this point that the principle of parsimony makes its entrance; if chimps exhibit behavior B2, this is better explained by the hypothesis that they have second-order intentionality than by the hypothesis that they have alternative mechanism A.
2
Why should we trust the parsimony arguments just described? Is parsimony an inferential end in itself or is there some deeper justification for taking parsimony seriously? In the present circumstance at least, there is no need to regard parsimony as something that we seek for its own sake.[25] My suggestion is that parsimony matters in problems of phylogenetic inference only to the extent that it reflects likelihood.[26] Here I am using the term Alikelihood@ in the technical sense introduced by R.A. Fisher. The likelihood of a hypothesis is the probability it confers on the observations, not the probability that the observations confer on the hypothesis. The likelihood of H, relative to the data, is Pr(Data * H), not Pr(H * Data).
To see how the likelihood concept can be brought to bear on cladistic parsimony, consider Figure 2. It can be shown, given some minimal assumptions about the evolutionary process, that the data depicted in that figure are made more probable by the (RS)C hypothesis than they are by the R(SC) hypothesis. These assumptions are as follows:
heritability: Character states of ancestors and descendants are positively correlated. That is, Pr(Descendant has a wing * Ancestor has a wing) > Pr(Descendant has a wing * Ancestor lacks a wing).
chance: All probabilities are strictly between 0 and 1.
screening-off: Lineages evolve independently of each other, once they branch off from their most recent common ancestor.
What we have here, in its essentials, is a proof that Reichenbach[27] gave in connection with his principle of the common cause.
I think that most biologists would agree that these three assumptions hold pretty generally. The first of them does not say that descendants probably end up resembling their ancestors. The claim is not that stasis is more probable than change—that Pr(Descendant has a wing * Ancestor has a wing) > Pr(Descendant lacks a wing * Ancestor has a wing). Rather, the claim is that if a descendant has a wing, that this result would have been more probable if its ancestor had had a wing than it would have been if its ancestor had lacked a wing.[28] The last assumption, it should be noted, is not exceptionless; after all, there are ecological circumstances in which a lineage=s evolving a trait influences the probability that other contemporaneous lineages will do the same. But even here, the assumption of independence is often true; and when it is false, it usually can be weakened without materially affecting the qualitative conclusions I want to draw.
Not only does likelihood provide a framework for understanding the role of parsimony considerations in phylogenetic inference; it also has implications about the Self and Other problem depicted in Figure 4. If traits M and A obey the assumptions listed, the following inequality is a consequence:
(P) Pr(Self has M * Other has M) > Pr(Self has M * Other has A).
This proposition says that Self and Other are correlated (nonindependent) with respect to the traits M and A. It also says that there is a likelihood justification for anthropomorphism.[29] The observation that Self has M is rendered more probable by the hypothesis that Other has M than by the hypothesis that Other has A. Such differences in likelihood are generally taken to indicate a difference in support—the observation favors the first hypothesis over the second.[30]
The likelihood concept also throws light on De Waal=s proviso—that parsimonious anthropomorphism is on firmer ground for our near relatives than it is for those individuals to whom we are related more distantly. We may translate this into the claim that the two probabilities compared in proposition (P) are more different when Self and Other are closely related and become more similar as the relationship becomes more distant. This thesis is illustrated by Figure 8, in which X and Y are more closely related to each other than either is to Z. The lower case letters in the tree=s interior are path coefficients, which entail that the correlation of X and Y (rXY) and the correlation of X and Z (rXZ) have the values rXY = ab and rXZ = acd. Notice that rXY > rXZ if and only if b > cd. This inequality need not be true, but it does follow from an assumption that often figures in evolutionary discussions. This is the assumption of uniform rates—that a given evolutionary event has the same probability of occurring in different contemporaneous branches of a tree. In the present example, uniform rates means that a=b and d=ac, which suffices to insure that X and Y are more strongly correlated than are X and Z. Thus, De Waal=s proviso[31] is not true unconditionally; but when it is true, it need not be added as an independent constraint on parsimony arguments (which have no machinery for taking account of recency of divergence). The proviso flows from a likelihood analysis.
Figure 8
3
Genealogical relatedness suffices to justify a likelihood inference from Self to Other, if the Reichenbachian assumptions that I described hold true. However, is genealogical relatedness necessary for this extrapolation to make sense? Proposition (P) says that Self and Other are correlated with respect to trait M. What could induce this correlation? Reichenbach argued that whenever there is a correlation of two events, either the one causes the other, or the other causes the one, or the two trace back to a common cause. Considerations from quantum mechanics suggest that this is not always the case,[32] and doubts about Reichenbach’s principle can arise from a purely classical point of view as well.[33] However, if we are not prepared to suppose that mentalistic correlations between Self and Other are brute facts, and if Self=s having M does not causally influence whether Other has M (or vice versa), then Reichenbach’s conclusion seems reasonable, if not apodictic; if Self and Other are correlated, this should be understood as arising from a common cause.
Genealogical relatedness is one type of common cause structure. It can induce the correlation described in (P) by having ancestors transmit genes to their descendants, but there are alternatives that we need to recognize. To begin with, parents exert nongenetic influences on their offspring through teaching and learning. For example, children have a higher probability of speaking Korean if their parents speak Korean than if their parents do not, but this is not because there is a gene for speaking Korean. And there are nongenetic connections between parents and offspring that do not involve learning, as when a mother transmits immunity to her children through her breast milk.
Correlations between Self and Other also can be induced by common causes when Self and Other are not genealogically related. If students resemble their teachers, then students of the same teacher will resemble each other. Here learning does the work that genetic transmission is also able to do. Similarly, Self and Other can be correlated when they are influenced by a common environmental cause that requires no learning. For example, if influenza is spreading through one community, but not through another, then the fact that I have the flu can be evidence that Other does too, if the two of us live in the same community.
I list these alternatives, not because they apply with equal plausibility to the problem of other minds, but to give an indication of the range of alternatives that needs to be considered. Genealogical relatedness is only an example; the fundamental question is whether there are common causes that impinge on Self and Other that induce the correlation described in (P). If there are, then there will be a likelihood justification for extrapolating from Self to Other.[34]
Figure 9
4
What, exactly, does cladistic parsimony and its likelihood analysis tell us about the problem of other minds? When I cry out, wince, and remove my body from an object inflicting tissue damage, this is (usually) because I am experiencing pain. When other organisms (human or not) produce the same set of behaviors, is this evidence that they feel pain? This hypothesis about Other is more parsimonious (if the behaviors are homologous and there are no known relevant neurophysiological differences), and it is more likely (if Reichenbachian assumptions about common causes apply). Does that completely solve the problem of other minds, or does there remain a residue of puzzlement?
One thing that is missing from this analysis is an answer to the question—how much evidence does the introspected state of Self provide about the conjectured state of Other? I have noted that the likelihoods in proposition (P) become more different as Self and Other become more closely related, but this comparative remark does not entail any quantitative benchmarks. It is left open whether the observation strongly favors one hypothesis over the other, or does so only weakly.
Another detail that I have not addressed is how probable it is that Other has M. If I have M when I produce behavior B, and Other exhibits B, is the probability greater than 2 that Other has M as well? The previous discussion helps answer that question, in that principle (P) is equivalent to the claim that Self’s having M raises the probability that Other has M. Whatever the prior probability is that Other has M, the posterior probability is greater. Whether additional information can be provided that allows the value of that posterior probability to be estimated is a separate question.
A genealogical perspective on the problem of other minds helps clarify how that problem differs from the behavior-to-mind problem discussed at the outset. At first glance, it might appear that it does not matter to the problem of other minds whether the other individual considered is a human being, a dog, an extra-terrestrial, or a computer. In all these cases, the question can be posed as to whether knowledge of one’s own case permits extrapolation to another system that is behaving similarly. We have seen that these different formulations receive different answers. I share ancestors with other human beings, and I share other, more remote, ancestors with the nonhuman organisms found on earth. However, if there are creatures on other planets that evolved independently of life on earth, then I share no common ancestors with them. In this case, extrapolation from Self to Other does not have the genealogical justification I have described. I would go further and speculate that it has no justification at all. This does not mean that we should never attribute mental states to such creatures. What it does mean is that we must approach such questions as instances of the purely third-person behavior-to-mind inference problem. Similar remarks also may apply to computers. When they behave similarly to us (perhaps by passing an appropriate Turing test or by playing a competent game of chess), we may ask what causes them to do so. We have no ancestors in common with them; rather, we have constructed them so that they produce certain behaviors. Does this fact about the design of computers provide a reason to think that the proximate mechanisms behind those behaviors resemble those found in human beings? Arguably not. For extra-terrestrials and (arguably) for computers, extrapolation from one=s own case will not be justified. Indeed, this point applies to organisms with whom we do share ancestors, if the behaviors we have in common with them are not homologies (Figure 5). And even when Self and Other do exhibit a behavioral homology, if Self and Other are known to deploy different neural machinery for exhibiting that behavior, the extrapolation of M from Self to Other is undermined (Figure 6).
A probabilistic representation of the problem of other minds shows that the usual objection to extrapolating from Self to Other is in fact irrelevant, or, more charitably, that it rests on a factual assumption about the world that we have no reason to believe. The question is not whether introspected information about one=s own mind provides lots of data or only a little. Rather, the issue is how strong the correlation is between Self and Other. Consider two urns that are filled with balls; each ball has a color, but the frequencies of different colors in the urns are unknown. The urns may be similar or identical in their compositions, or they may be very different. If I sample one ball from the first urn and find that it is green, does this provide substantial evidence concerning the second urn=s composition? If I sample a thousand balls from the first urn, does this allow me to say any more about the second? Everything depends on how the two urns are related. If they are independent, then samples drawn from the first, whether they are small or large, provide no information about the second. But if they are not independent, then even a small sample from the first may be informative with respect to the second. There is nothing wrong with asking whether knowledge of Self supports a conclusion about Other. But the skeptical assertion that it does not involves a factual claim about the world. A claim of independence is no more a priori than a claim of correlation. If the relevant mental states of Self and Other are joint effects of a common cause (with the properties that Reichenbach described), then the skeptical assertion is false.
The argument I have presented is intended to show how certain propositions are justified; I have not addressed the question of whether people are justified in believing this or that proposition. My own mental state can be an indicator of the mental states of others, whether or not I know that this is true, or understand why it is true. But what is required for people to be justified in extrapolating from Self to Other? Must they know the fine points of phylogenetic parsimony or of likelihood analysis? Or does it suffice that a given extrapolation is sanctioned by parsimony or likelihood considerations? This is the epistemological thicket in which internalist and externalist views of justification are in contention.[35] I won=t try to evaluate these different approaches to the concept of justification, nor to say in any detail how they are related to the ideas I have developed in this paper, but I do want to explore one line of questioning that arises from an internalist point of view.
Is it possible for me to figure out that proposition (P) is true, if I do not already know whether Other has M or A? Surely I can. I can tell whether M is heritable by looking at still other individuals who are known to have either M or A and see how they are related genealogically. But is it possible to determine that (P) is true without knowing anything about which individuals (other than one=s self) have M and which have A? Without this information, how can I tell whether the traits are heritable?[36] Well, knowledge in some strong philosophical sense is probably unnecessary, but perhaps judgments about heritability require one to have reasonable opinions, however tentative, about which individuals have which traits. Even so, the solution to the problem of other minds that I have suggested would not be undermined. The incremental version of the problem, recall, asks whether knowledge of my own case should make a difference in the characteristics I attribute to others. It is not required that I conceive of myself as beginning with no knowledge at all concerning the internal states of others.
We learned from I.J. Good that there is no saying whether a black raven confirms the generalization that all ravens are black unless one is prepared to make substantive background assumptions. The mere observation that the object before you is a black raven is not enough. The same point, applied to the problem of other minds, is that the mere observation that Self and Other share B and that Self has M is not enough. Further assumptions are needed to say whether these observations confirm the hypothesis that Other has M. Recognizing this point in the ravens paradox does not lead inevitably to skepticism, and it should not have that effect in the case of self-to-other inference. The problem of other minds should not be shackled with the ‘methodological fiction’ that Hempel imposed on the ravens paradox. When the fetters are broken, the problem of other minds turns into the problem of searching out common causes.