1. Introduction

Are psychological properties multiply realizable? I gather that the answer to this question is “yes.” There appears to be no logical contradiction in the possibility that the same psychological property, for example, being hungry, might be realized in various physical ways. Moreover, there is no obvious nomological constraint on multiple realization (but see Shapiro Reference Shapiro2004). A more interesting question is whether psychological properties are multiply realized. The relationship between psychology and neuroscience and, a fortiori, the extent to which these sciences might inform and constrain each other depends on how pervasive multiple realization is.

Philosophers of mind have typically assumed that if some psychological properties are multiply realized, then many or all are. Thus, although Putnam (Reference Putnam, Capitan and Merrill1967), in his groundbreaking discussion of multiple realizability,Footnote ¹ mentions only pain and hunger as examples of mental properties that are likely to be realized differently in mammals and mollusks, he clearly believes multiple realization to extend to many other psychological properties as well. This would be unremarkable if it were only the multiple realizability of psychological states at issue, but Putnam took himself to be offering something much more, namely, that psychological properties are actually multiply realized. Even if Putnam had presented evidence for his claim that pain and hunger are multiply realized, which he did not, I think it would be too hasty to conclude from these two cases that all or even many psychological properties are multiply realized.

Just as the number of psychological properties that are multiply realized may range from none to all, so too the number of different ways to realize a psychological property may vary from property to property (see Polger Reference Polger2002). Indeed, it seems unlikely that if pain is realized in 2,384 different ways, then every psychological property must be realized in 2,384 different ways. Some psychological properties might be “easier” to realize than others. As with the proportion of psychological properties that are multiply realized, the armchair is not the right place from which to decide the extent to which any given psychological property is multiply realized.

But rejection of the armchair is no guide to how one should approach the empirical questions about multiple realization that I have raised above. How does one test whether a given psychological property is multiply realized? Below, I discuss some of the conceptual difficulties that confront the attempt to test for multiple realization.

2. Testing

Some general remarks about testing are in order. The first point is familiar. Testing a hypothesis involves drawing predictions from it, the accuracy of which must be checked against observations. To test the hypothesis that it was Sophia who stole the cookies from the cookie jar, one must consider what differences observable to the world Sophia's crime will have produced. But, as Quine (Reference Quine1951) reminded us, the hypothesis that Sophia stole the cookies does not by itself entail any observations. Various auxiliary assumptions must be added to the hypothesis. Is Sophia tidy or is she careless? If the latter, the hypothesis might predict that the lid will have been left off the cookie jar. Does Sophia like to eat in her bed or at the kitchen counter? If the former, the hypothesis might predict that there will be crumbs in her bed. We are in a position to test the hypothesis only if we assume some things about Sophia.

A second point is this. The observations used to test a hypothesis cannot assume the truth of the hypothesis. The observation that the lid has been left off the jar does not assume that Sophia stole the cookies. Perhaps some other thief is careless. However, consider the observation statement “Sophia left the lid on the counter.” If this statement could be true only if Sophia stole the cookies, then the observation cannot be used to test the hypothesis. It is an observation that already rules out any alternative hypothesis about why the cookies are missing.

One final point. As Sober (Reference Sober1999) has emphasized, testing is a contrastive exercise. This is so for two reasons. To see the first, consider the hypothesis that Sophia's sister Thalia stole the cookies and left behind evidence that would incriminate Sophia. So skilled was Thalia at framing her sister that any evidence supporting the hypothesis that Sophia stole the cookies also supports the hypothesis that Thalia has framed her. Here we have a case in which Sophia's guilt is difficult to determine by the empirical means available. The hypothesis that Sophia stole the cookies is untestable relative to the hypothesis that Thalia framed her. However, relative to some other hypothesis, for example, that Sophia's ravenous father took some cookies, the hypothesis might be testable. This shows that whether a hypothesis is testable depends on its competitors. Prior to articulating a hypothesis against which another hypothesis is to be tested, we cannot say whether the hypothesis is testable.

The second reason to regard testing as contrastive follows from the frequent difficulty in assigning a value to the probability that some particular observations will be made given the truth of a hypothesis. How probable is it that the cookie jar lid will be left on the counter or that there will be crumbs in Sophia's bed if Sophia stole the cookies? Maybe Sophia only sometimes leaves the lid off the cookie jar. Maybe the cookies she is eating are sticky and leave no crumbs. The connection between hypotheses and the observations they predict is probabilistic and often indeterminate. This means that the most one can often hope to determine is whether one hypothesis makes some observations more probable than another hypothesis would. If the hypothesis about Sophia is being tested against the hypothesis that Sophia's perpetually dieting mother took the cookies, we do not need to calculate precise probabilities in order to judge that the probability of the jar's lid being left on the counter is higher given the former hypothesis than the latter.

In summary, testing a hypothesis requires attention to the following principles:

1. 1. The hypothesis must be combined with auxiliary assumptions.

2. 2. The truth of observation statements should not assume the truth of the hypothesis.

3. 3. The hypothesis must be tested against another.

In the next section we will see how these principles apply to the question of multiple realization.

3. Testing for Multiple Realization

Consider a psychological property such as pain. If pain is multiply realized, we should predict that the realizer of pain in one organism differs from the realizer of pain in another organism. But in this context the third point above is relevant. Putnam assumes that pain is realized differently in mammals and mollusks. Similarly, Block and Fodor (Reference Block and Fodor1972) (and many others) claim that the brain's plasticity demonstrates that various psychological properties are multiply realized. But the observation statements “mammals and mollusks realize pain differently” and “the same brain can realize the same psychological properties in different ways” assume the truth of multiple realization. An observation statement whose truth assumes the truth of a hypothesis, as these observation statements do, cannot serve in a test of that hypothesis. Putnam, Block, and Fodor are, in effect, reasoning as follows:

1. 1. Hypothesis: Psychological properties are multiply realized.

2. 2. Observation: Psychological properties are multiply realized.

3. 3. $P( O\vert H) =1$.

4. 4. Therefore, psychological properties are multiply realized.

Clearly, this reasoning is not the sort by which hypotheses should be tested.

To help see which observations do count as evidence for multiple realization, it is necessary to consider multiple realization against a competing hypothesis. Let RU be the hypothesis that the realizers of psychological properties are uniform and MR the hypothesis that psychological properties are multiply realized. Let MM be the observation statement that pain is multiply realized in mammals and mollusks. Here is the defective version of Putnam's argument:

As I pointed out above, MM already assumes that MR is correct and that RU is false. Putnam needs to characterize his evidence in a way that does not beg the question.

I suggest that Putnam adopt some other observation statement, that is,

DIF avoids the problem of circularity that MM introduces. However, it faces other challenges. Which differences are those that matter: which differences count as evidence for MR and against RU? Intuitively, the fact that the brains of mammals and mollusks differ in location, or in color, or in the number of atoms they contain should not be relevant. These are the sorts of observations that might be expected given either hypothesis, and so if they support one, they support the other. Grounding this intuition is an appreciation for the kind of explanations that are common in cognitive neuroscience. Cognitive neuroscientific explanations abstract from details such as color, mass, and number of constituent atoms. These explanations decompose the brain into assortments of information-processing mechanisms, which are then identified with particular structures in the brain (see chapters in Bechtel et al. Reference Bechtel, Mandik, Mundale and Stufflebeam2001 for numerous examples of this sort of explanation). Because analyses of how neural structures perform information-processing tasks do not mention properties such as color, mass, and so on, differences in these properties should play no role in comparing MR to RU. If the neural mechanisms identified as those responsible for pain in the mammal could not or do not process information as the neural mechanisms in the mollusk do, then this would appear to be evidence for MR and against RU.

But intuition, as well as an appeal to practice in cognitive neuroscience, cannot settle the issue of relevance and so cannot provide all the answers about what evidence will count in favor of MR and against RU. A case that I have considered in previous work (Reference Shapiro2004) shows the problem. Simplifying, Sur and his group have succeeded in rewiring the brains of ferrets so that information that is normally processed in visual cortex receives processing in auditory cortex instead (Sharma, Angelucci, and Sur Reference Sharma, Angelucci and Sur2000; von Melchner, Pallas, and Sur Reference von Melchner, Pallas and Sur2000). The case is a promising one for proponents of multiple realization because the mechanism by which normal ferrets process visual information appears to differ from the mechanism by which rewired ferrets process visual information. The rewired ferret brains seem to provide evidence in favor of MR and against RU.

However, Shapiro Reference Shapiro2004 argued that multiple realization requires both a difference at the level of the realizing properties and a similarity at the level of realized properties; but it is not obvious that the rewired ferret brains do differ from normal ferret brains in the relevant respects, nor that the visual properties of the rewired ferrets are similar to the visual properties of normal ferrets in the relevant respects. In fact, the auditory cortex of the rewired ferrets comes to resemble the visual cortex in its structure: two-dimensional spatial maps appear; columns of orientation-sensitive cells, characteristic of visual cortex, emerge; and so on. On the other hand, the visual properties of the rewired ferrets are not exactly the same as those of the normal ferrets. Rewired ferrets have poorer visual acuity than normal ferrets do.

The ferret case illustrates the difficulty involved in testing MR against RU. The rewired ferrets ought to count in favor of one or the other hypothesis, but which? Until some auxiliary assumptions are added to the hypotheses, it is not possible to say. Indeed, it is illuminating to reconstruct disputes over the possibility and extent of multiple realization in terms of disagreements over auxiliary assumptions.

Suppose one accepted the following auxiliary assumption:

As Heil (Reference Heil, Walker and Heckmann2003) has pointed out, AA1 is standard fare in discussions of multiple realization. To say that pain is a higher-level property is to say that being in pain is having some property or other that has the causal powers of pain. For instance, mammals and mollusks could share the higher-level property of being in pain but can differ in the properties in virtue of which they possess this higher-level property. But an attribution of the same higher-level property to different entities requires that these entities possess the same property, that is, that they are exactly similar in some respect. Thus, if one is to accept AA1 and agrees with Putnam that mammals and mollusks can both possess the higher-level property of being in pain, then one must also be prepared to accept that a mammal's pain is exactly similar to a mollusk's pain. One would also have to reject that visual properties are multiply realized in normal and rewired ferrets because, in fact, these visual properties differ across the two kinds of ferret. Thus, MR combined with AA1 does not receive support from Sur's work on ferrets.

But AA1 is not the only option. Heil (Reference Heil, Walker and Heckmann2003) endorses

Motivating AA2 is the recognition that not all predicates name properties. This is clearly true of predicates such as ‘is phlogiston’ and ‘is a witch’, but, Heil thinks, this is also true of predicates used to pick out higher-level properties. The claim that ‘has pain’, ‘is hungry’, ‘has a desire for chocolate’, and so forth do not name properties is remarkable and obviously wanting of defense.

One reason that Heil rejects higher-level properties is a suspicion that entities never are exactly similar in some respect, as attributions of higher-level properties require. Is the mammal's pain exactly similar to the mollusk's pain? This seems unlikely. Similarly, are various red objects exactly similar with respect to their redness? Far more plausible, Heil thinks, is that ‘is in pain’ or ‘is red’ picks out families of similar properties. We might use the predicate ‘is in pain’ to describe a skunk and an octopus, and application of this predicate to each might be completely proper, but none of this entails that a skunk and an octopus actually share a pain property. Rather, all that is necessary for correct application of the predicate is that a skunk and an octopus have similar properties, that is, properties that confer similar powers.

I have argued that when combined with AA1, MR predicts that the normal and rewired ferrets do not exemplify multiple realization. The differences in the ferrets’ visual properties show that there is no higher-level visual property that the ferrets share. Thus, there is no property that is being multiply realized. How does MR fare in combination with AA2?

Heil seems to think that MR is true given AA2:

However, contrary to Heil's sentiment, I think that AA2's truth makes MR false. The attraction of AA1, according to which there exist higher-level properties, is that it supposes the existence of something that can be multiply realized. If pain is a property, then we can understand what it means to say that pain is multiply realized in mammals and mollusks. But if the predicate ‘pain’ names no property, then there is nothing that mammals and mollusks multiply realize when each is said to be in pain. We can grant that the predicate still applies, where this means only that each has a property that disposes them to behave in similar ways, but, of course, the predicate is not multiply realized.

The criticism above does not, of course, show AA2 to be false. Rather, it shows that AA2 is inconsistent with MR. This means that if Heil's “ontologically serious” conception of properties is correct, then MR is false. In this case, the most that could be said about the ferrets is that rewired ferrets have visual properties that are similar to those of normal ferrets. Whether they are similar enough to warrant labeling them with the same predicates is another matter, and one not relevant to the question of multiple realization.

I have been investigating various auxiliary assumptions that might render MR and RU testable. The strategy has been to look at some evidence that ought to favor one or the other of MR and RU. As it turns out, on either auxiliary assumption, the evidence does not support MR. But notice that RU does no better than MR. The reason is that RU too depends on the existence of higher-level properties. The hypothesis that realization is uniform assumes that there is some property that is common across all organisms that, for instance, are in pain. But if there is such a property, it is not a higher-level property, as AA1 assumes. Nor, as AA2 assumes, are there properties whose realization can be uniform in the first place.

Especially distressing for discussions of multiple realization is that AA1 and AA2 seem to exhaust the space of possibilities. Either predicates such as ‘is in pain’ name higher-level properties or they do not. If they do, then every organism to which the predicate applies is exactly similar in some respect, which seems false. If they do not, then there is nothing to be multiply realized in the first place. Our effort to test MR has taken a disappointing turn.

4. A New Recipe and a New Auxiliary Assumption

I have so far treated the problems with the ferret case as stemming from the auxiliary assumptions that must be combined with MR to render it testable. Assumptions AA1 and AA2 seem exhaustive: there are higher-level properties or not. Either way, however, MR runs into trouble. Perhaps, however, the trouble should not be traced to the auxiliary assumptions, but to the characterization of MR I have assumed. I have supposed that MR requires sameness at one level and difference at another. Suppose, instead of sameness and difference, MR requires only similarity and difference. Adoption of this new recipe allows one to adhere to AA1 without ill results.

With this new recipe in hand, one can still hold a distinction between higher- and lower-level properties, contrary to what AA2 stipulates. One can also still hold that two entities share a higher-level property if and only if they are exactly similar in some respect. However, because MR is no longer taken to require that the same property has different realizations, but only that similar properties have different realizations, concerns about exact similarity lose their edge. Now the normal and rewired ferrets might constitute evidence for MR because, although the visual properties of these ferrets are not the same, they are similar.

Unfortunately, this suggestion faces an obvious problem that the ferret case makes clear. We return to our old question: do the normal and rewired ferrets provide evidence for MR? Suppose we judge that the normal and rewired ferrets have similar visual properties; that is, their visual properties are much more similar to each other than either's is to the visual properties of any other organism. Moreover, the brains of normal and rewired ferrets differ from each other. But, we are now back to the problem of difference. Are the normal and rewired brains different enough? What kinds of difference are relevant?

Here is what we do not want to conclude. We do not want to say, on the one hand, that ferret vision is multiply realized, but on the other hand that the similarities in the visual properties of normal and rewired ferrets are the result of similarities in the brains of normal and rewired ferrets. To see why this is so, let $V_{n}$ stand for the visual properties of the normal ferret and $V_{r}$ stand for the visual properties of the rewired ferret. Suppose that $V_{n}$ is realized by neural properties $N_{1}$, $N_{2}$, $N_{3}$, and $N_{4}$. Suppose also that $V_{r}$ is realized by neural properties $N_{1}$, $N_{2}$, $N_{3}$, and $N_{5}$. Further suppose that the similarities in $V_{n}$ and $V_{r}$ are a result of the fact that the realizers of these properties share $N_{1}$–$N_{3}$ and that the differences between $V_{n}$ and $V_{r}$ are a consequence of the fact that among $V_{n}$'s realizers is $N_{4}$ whereas among $V_{r}$'s realizers is $N_{5}$. Now it appears that $V_{n}$ and $V_{r}$ are similar insofar as their realizers are the same, and they are dissimilar insofar as their realizers differ.

As so described, I think the differences between the brains of normal and rewired ferrets do not count as evidence for the multiple realization of ferret visual properties. The right thing to say is that the visual properties of the two kinds of ferret are similar insofar as their brains are similar, and they differ insofar as their brains differ. Consequently, the new recipe for MR—sort of the same but sort of different—does not work. In what follows I will offer an auxiliary assumption that, when combined with the new recipe for MR, seems to get the right results. I will then consider in the next section a case in which failure to adopt the auxiliary assumption leads to the wrong results.

The auxiliary assumption I have in mind is this:

I believe the motivation for DD is intuitively appealing. Consider three watches. Watches S and F are analog and watch D is digital. Suppose that watch S runs a bit slow and watch F runs a bit fast. Watch D keeps perfect time. I shall refer to these tendencies as the timekeeping properties of the watches. How many different (types of) realizations of timekeeping properties do these watches represent? All three watches exhibit similar timekeeping properties. However, the mechanisms within the three watches do not differ to the same degree. Specifically, the mechanisms that realize timekeeping properties in watches S and F are much more similar to each other than either is to the mechanism that realizes timekeeping properties in watch D. Moreover, suppose it were true that explaining why S and F differ in their timekeeping properties is the fact that their mechanisms differ in just a single respect: the spring in S is longer than the spring in F. Accordingly, the variance in timekeeping properties between S and F is fully explained by the variance in spring length in S and F. It seems natural to say in this case that timekeeping properties are realized in the same way in S and F, but that because the realizers differ slightly in S and F, S and F have slightly different timekeeping properties.

The fact that watch D has slightly different timekeeping properties than either S or F, on the other hand, is not a consequence of D's possessing a realizer whose difference from the realizers of timekeeping in S and F causes only a difference in timekeeping properties. In fact, the realizer of timekeeping in D differs in many ways from the realizers of timekeeping in S and F, and not all these different realizing properties account for differences in timekeeping properties. Indeed, some of the realizing properties of timekeeping in D account for the similarities in timekeeping that D shares with S and F. This is even clearer if we suppose that D and S are exactly similar with respect to their timekeeping properties. In such a case, S and D have the same timekeeping properties, and so there is no difference between the realizers of timekeeping that causes a difference in timekeeping properties, but there are differences in the realizing properties nevertheless. This is the kind of evidence we want for MR.

With DD in place, the ferrets do not support MR. The reason is that the differences in ferret brains (I am assuming) explain nothing more than differences in ferrets’ visual properties. If neural plasticity is to provide the kind of evidence that supports MR, it must be plasticity of a special sort. It must be plasticity that reorganizes the brain in a way such that differences between the modified brain and the normal brain are not limited to just those differences that explain differences in the realized psychological property.

5. Getting It Wrong: Bechtel and Mundale

In an influential article, Bechtel and Mundale (Reference Bechtel and Mundale1999) argue that psychological properties are not multiply realized. Their argument begins with the observation that neuroscientists have been very successful in mapping the brain, that is, in identifying the neural structures that realize psychological properties. The second premise in their argument is that brain mapping has been possible only through a comparative analysis of brains—an analysis that crosses species lines. But, premise three is “it is the very similarity (or more precisely homology) of brain structures which permits us to generalize across certain species” (Reference Bechtel and Mundale1999, 178; their italics). Bechtel and Mundale thus conclude that “in the context of neuroscientific research, they are not multiply realized” (178).

Before I consider the merit of this argument, Bechtel and Mundale's conclusion needs clarification. They seem to be saying that brain structures are not multiply realized, but this cannot be what they mean. The issue of importance in discussions of multiple realization, and the issue that Bechtel and Mundale intend to address, is whether psychological properties (or states, or capacities) are multiply realized. Thus, I believe a charitable reading of their conclusion is not that brain structures are not multiply realized, but that psychological properties are not multiply realized. In a nutshell, then, Bechtel and Mundale are arguing that psychological properties are not multiply realized because analyses of homologous brains show that the same or similar psychological properties are realized in the same or similar neurological structures.

The troubling feature of Bechtel and Mundale's argument against MR is the use it makes of homologies (see Kim Reference Kim2002 for a similar line of criticism). Homologous structures in different species owe their similarity to the fact that the species descended from a common ancestor. Naturally, because the species diverged over time, one should expect that homologous structures are not exactly similar. Thus, for instance, the wings of hummingbirds and eagles are not exactly similar despite being homologues. But, and this is the crucial point, the extent to which homologous traits are similar owes to their common ancestry.

Suppose now that hummingbirds and eagles exhibit different flight properties, including properties such as gliding time, wing beats per minute, distance covered per wing beat, and so on. Owing to their common ancestry, the wings of hummingbirds and eagles will be similar; but owing to divergence over time, the wings will differ. My claim is that the differences that hummingbirds and eagles display in flight properties are due to differences in their wings, but that these differences cause only the differences in flight properties. In other words, hummingbirds and eagles share flight properties in virtue of possessing wings that are similar, and whatever differences there are in their wings explain only differences in their flight properties.

The claim becomes clearer when comparing the hummingbird's flight properties to the bat's. The wings of bats and hummingbirds evolved independently. Whatever similarities exist between the flight properties of bats and hummingbirds are not the result of common ancestry. The differences in flight properties, we can suppose, are the result of differences in the wing properties. However, the bat's wings differ from the hummingbird's in other respects that do not cause differences in flight properties. Whereas differences in the wings of hummingbirds and eagles account only for the differences in their flight properties, some of the differences in the wings of hummingbirds and bats will not produce differences in flight properties.

Because Bechtel and Mundale focus on homologous brain structures, they overlook potentially genuine cases of multiple realization. Macaque vision and human vision might be very similar, but whatever difference there is between the visual properties of macaques and human beings is possibly due to the only differences that are present between macaque and human brains. Ideally, to test whether MR holds true of visual properties, one should be comparing brains that differ in respects other than those that make a difference in visual properties. One should be looking at different brains that realize similar visual properties despite their differences. But this is precisely the kind of evidence for MR that Bechtel and Mundale choose to ignore.

6. Conclusion

I have been considering some methodological issues concerning the testability of multiple realization. Insofar as MR is construed as an empirical hypothesis, observations counting as evidence on its behalf cannot exclude the possibility of alternative hypotheses. MR must be tested against a competitor. Finally, MR must be combined with auxiliary hypotheses. But which auxiliary hypotheses should be conjoined with MR to render it testable? I have suggested DD as an auxiliary assumption that clarifies the kind of evidence that can be used in support of MR. According to DD, differences in realizers that do not only cause differences in realized properties, but cause similarities as well, are the right sort of differences. These are the differences that provide evidence for MR and against RU.