TYPE-SELECTION VERSUS TOKEN-SELECTION 1

Transcription

1 TYPE-SELECTION VERSUS TOKEN-SELECTION 1 1 There are many controversies about the nature of natural selection. I argue that many of them can be traced back to a simple and basic distinction about whether the theory of natural selection is about trait types or trait tokens. I argue that much of the controversy about selection is due to the fact that this difference has not been acknowledged. The aim of this paper is to trace the consequences of the difference between these two ways of thinking about selection. If we take selection to be typeselection, we cannot explain why individual organisms have the traits they have with reference to selection processes. In turn, if we take selection to be token-selection, we have to reinterpret or even give up the concept of fitness. These considerations, taken together, should persuade us to think of selection as token-selection. I. Introduction Here is an ambiguity in the way the theory of natural selection is formulated. The units of selection, whatever they are (Lloyd 2001), have properties or traits. And when a unit of selection is competing with another one for survival, what decides which one will survive is the difference between their traits. The question is how these traits are to be referred to by a theory of natural selection. When we talk about traits as being selected, we face a type-token ambiguity: are trait types or trait tokens being selected? Similarly, when we say that what decides which unit of selection will survive is the difference between their traits, we face the same type-token ambiguity: difference between trait types or difference between trait tokens? In short, regardless of what we take to be the units of selection, we need to be able to tell whether the theory of natural selection is about their trait types or their trait tokens. The aim of this paper is to argue that depending on whether we take the theory of natural selection to be about trait types or trait tokens, we get very different conceptions of selection, which I call type-selection and token-selection respectively. I will then compare these two conceptions of selection and conclude that we may be better off using the latter one. We need to be careful when formulating the distinction between type-selection and tokenselection. Why couldn t we just say that a theory of selection needs to talk about both types and tokens? Why couldn t we say that we should not pit these two conceptions of selection against one another: that they should work hand in hand? I do not want to deny that a full account of the theory of natural selection will have to talk about both trait types and trait tokens. But the question I raise is more specific: when the theory of natural selection talks about the difference between traits of the units of selection (whatever they may be) that are responsible for (or, that explain) which units of selection will get to be selected, then we need to be able to tell whether these traits are to be taken as trait types or as trait tokens. (When I talk about whether selection theory is about trait types or trait tokens in what follows, I use the phrase in this more specific sense.) The main point is that typeselection explains why a certain type of unit of selection is selected with the help of the property- 1 I am grateful for comments by Peter Godfrey-Smith, Richard Lewontin, John Dupré, Karen Neander, David Sanford, Marc Lange, David Papineau, Roberta Balarin, Jason Rhein, Kris McDaniels, Ben Bradley, Kevan Edwards and James Young on various earlier versions of this paper. This paper grew out of the discussion at my PhD seminar at (deleted) University. I m also very grateful for comments I received from the referees of BJPS.

2 types these units of selection all have instantiations of. Token-selection, in contrast, explains why a certain specific unit of selection is selected with the help of the property-instance of this specific unit of selection. 2 A simpler, but somewhat misleading way of putting this dilemma is this: does selection operate on trait types or on trait tokens? Although this way of rephrasing the question has some intuitive pull, some will object to my use of the phrase that selection operate on traits, especially as it has been argued recently that selection is a statistical and not a causal process, so it does not operate on anything. So a more neutral way of posing this dilemma is to ask whether a theory of selection need to talk about trait-types or about trait-tokens. 3 I argue that there is a serious trade-off between the advantages and disadvantages of these two different ways of thinking about selection. The aim of this paper is to trace the various consequences of these two ways of thinking about selection and point out their advantages and disadvantages. I will only compare them at the end of the paper and conclude that we are better off thinking about selection as token-selection. II. Type-selection versus token-selection I said that a theory of evolution by natural selection needs to be able to talk about the traits of the units of selection (whatever they may be) that are responsible for (or, that explain) which units of selection get to be selected. And the question is whether these traits are to be understood as trait types or as trait tokens. If they are trait types, we talk about type-selection. If they are trait tokens, we talk about token-selection. In order to make this distinction clear, I need to say more about the trait type/trait token distinction. It is important to emphasize that the difference between token-selection and type-selection is not a difference of the general metaphysical vocabulary used to describe the selection process. More precisely, it is not the difference between trope metaphysics and the metaphysics of universals. The term property is ambiguous. It can mean universals: properties that can be present in two (or more) distinct individuals at the same time. But it can also mean tropes: abstract particulars that are logically incapable of being present in two (or more) distinct individuals at the same time (Williams 1953, Campbell 1990, Bacon 1995, Schaffer 2001). Suppose that the color of my neighbor s black car and my black car are indistinguishable. They still have different tropes. The blackness trope of my car is different from the blackness trope of my neighbor s car. These two tropes are similar but numerically distinct. Thus, the blackness of my car and the blackness of my neighbor s car are different properties. If, in contrast, we interpret properties as universals, then the two cars instantiate the same property-type: blackness. Thus, depending on which notion of property we talk about, we have to give different answers to the question about whether the color-property of the two cars is the same or different. If by property we mean trope, then my car has a different (but similar) colorproperty, that is, color-trope, from my neighbor s. If, however, by property we mean propertytype, then my car has the very same property, that is, property-type, as my neighbor s. 2 2 Biologists call the properties of organisms traits. I use the terms trait and property interchangeably in the present context. 3 I am grateful to Mohan Matthen for reminding me of the partiality of the terminology of operating on.

3 It has been argued that the disagreement between those who take properties to be tropes and those who take properties to be universals is a merely verbal one: all claims about (and even arguments for the existence of) tropes can be rephrased in terms of instantiations of property-types (Daly 1997). Conversely, as we can think of property-types as resemblance classes of tropes, all claims about property-types can be rephrased in terms of tropes. Whether or not these claims about ontological equivalence are correct (see Nanay 2009 for an argument against), it is important that the difference between type-selection and token-selection is not the same difference as the one between universals and tropes. Both token-selection and type-selection can be formulated with reference to tropes as well as with reference to universals. If we accept trope nominalism, then this difference will be between the claim that selection theory is about tropes and that it is about resemblance classes of tropes. If we accept realism, it will be between the claim that selection theory is about universals and that it is about specific instantiations of universals. Both type-selection and token-selection are compatible with either metaphysical framework. In other words, the difference between type-selection and token-selection then has little to do with the general metaphysical debates about tropes and universals. It has to do with whether the units of selection are selected in virtue of having a certain trait type or in virtue of having a certain trait token. The type-selection view seems to be the standard way of thinking about selection. After all, evolutionary theory talks about genotypes and phenotypes: about property-types. As Elliott Sober and Richard Lewontin say in one of the most famous papers on natural selection, selection theory is about genotypes not genotokens (Sober-Lewontin 1982, p. 172, see also Sober-Lewontin 1983, p. 649). Or, as William Wimsatt says, if evolution had to depend upon the passing on of gene-tokens, it could not have happened. Genotokens and phenotokens are not inherited, but genotypes and phenotypes may be (Wimsatt 1980, n. 2, p. 174). Although many biologists and philosophers of biology assume that selection has to be typeselection, Elliott Sober gives an explicit statement and detailed defense of this view. As he says: Charlie the Tuna is not a particularly interesting object of study, but tuna dorsal fins are (Sober 2001, p. 310, see also Sterelny-Kitcher 1988). And again: to understand what it means to talk about the selection of genes, organisms, or groups, one must quantify over properties (Sober 1981, p. 162). Or, even more explicitly, evolutionary theory provides reason for thinking that properties [that is, property-types] exist (Sober 1981, p. 148, see also Sober 1984a). 4 Some of these quotes seem to imply that selection theory is indeed about property-types; some others may be interpreted as making a somewhat weaker claim: that it is about tokens qua instantiations of types this would be a similar claim as one that is popular in the causation literature: particular events cause other particular events qua, or in virtue of, having some propertytypes. For the purposes of the argument I will present, this difference is not a significant one: both of these ways of thinking about selection I will consider to be instances of what I call type-selection Sober makes it clear that what he means by properties is what I refer to property-types on pp of Sober He is attacking Quine s nominalism about property-types in this paper, see Quine 1953, If we think of selection as a causal process, then thinking of tokens qua instantiations of types will be the more plausible of these two versions (unless we want to take property-types to be the relata of causation).

4 Thus, the suggestion is that property-types are a necessary feature of the theory of selection: talking about property-types makes evolutionary explanations possible to begin with (see Sober 1981, esp. pp , for an explicit defense of this claim, see also Sober 1984a, pp ). The type-selection view seems to be the majority view of selection. But this does not mean that this view is uncontroversial (a couple of examples: Mayr 1959/1996, Smith-Varzi 2001, 2002, Godfrey-Smith 2009, Rosenberg 1983, Nanay forthcoming a, forthcoming b, forthcoming d). The alternative is the claim that the theory of natural selection is about trait tokens, not about trait types. As this view is less dominant in the literature, I will spend some time explaining why it is a plausible view and cashing out what it entails. The claim that selection theory is not about trait types may seem provocative, especially given the repeated mention of genotypes and phenotypes in the literature. We have already seen that the view that selection theory is about types is the majority view. However, Elliott Sober did not just accept this view as the default position, but gave some forceful arguments for the claim that we cannot even formulate evolutionary theory without the help of property-types. 6 In order to make the idea of token-selection plausible, I need to address the most important argument Sober gives in favor of the claim that evolutionary theory is about property-types (I will address another, related, argument he gives for the same claim in Section IV below). He writes: Objects, whether they be genes, organisms, or groups, may change in frequency owing to natural selection. But that these changes in composition occur, owing to selection, leaves open the question of what traits were selected for. In virtue of what did the change occur? Natural selection is selection for traits; the upshot is the selection of objects. The issue of artifact versus cause cannot even the expressed without the idea of selection for properties [that is, property-types]. (Sober 1981, p. 166) This distinction between selection for a trait and the selection of an object (see also Sober 1984a, Sober 1984b) has been extremely influential and widely discussed (see, for example, Goode-Griffiths 1995). So if selection for is always selection for a trait-type, then is token-selection even viable? Is there such a thing as selection for a trait token? Sober explicitly claims that selection for is always selection for a trait-type and not a traittoken. As he says, if there is selection for X, every object which has X has its reproductive chances augmented by its possessing X (Sober-Lewontin 1982, p. 171). My claim is that selection for, just like selection per se, has two interpretations, a token and a type one. Sober clearly presupposing the latter, but the former is also viable. In fact, one way of cashing out the difference between typeselection and token-selection is to ask whether selection for is selection for trait-types or traittokens. And we have no good reasons to exclude the latter option. If we maintain that selection for is selection for trait-tokens, we can still talk about the distinction between selection for a trait and the selection of an entity. If we think of selection as selection for trait tokens, then the same distinction can be formulated without helping ourselves to the concept of trait-types. An entity has a number of trait tokens. Any one of these trait tokens can be selected for: its color, its shape, its size, etc. If any of these trait tokens is selected for, the entity is 4 6 On the opposite end, however, see Nanay forthcoming d s argument on why talking about property-types in some domains of biology could be thought to be problematic.

5 more likely to survive (and/or, its reproductive success will increase). The entity is more likely to survive (and/or its reproductive success increases) in virtue of one of its token traits. The other token traits of the entity are irrelevant in this process. To sum up, we have two different ways of thinking about selection: as type-selection and as token-selection. In the next two sections, I will examine the consequences of this distinction in the context of some of the most important debates of philosophy of biology. I will argue that depending on which way of thinking about selection we endorse, we find ourselves on very different sides in these debates. III. The first consequence: What can selection explain? In this section, I examine how the difference between type-selection and token-selection relates to the epic debate over what selection can explain. One of the most important recent debates in philosophy of biology is about whether natural selection can explain why organisms have the traits they have. The view that selection can play a role in explaining why organisms have the traits they have, has been defended by Karen Neander (1995a, 1995b, see also Millikan 1990, Nanay 2005, Nanay 2010, Matthen 1999). On the other side of the trench the central figure is Elliott Sober (1984a, 1995, see also Walsh 1998, Dretske 1988, 1990, Pust 2001, Lewens 2001, Cummins 1975 and Stegmann 2010). Sober claims that selection is a negative force: it does not create; it only destroys (Sober 1984a, Chapter 5). Random mutations create a variety of traits (or genetic plans) and selection eliminates some of these, but the explanation of the traits of one of these individuals is provided by random mutation and inheritance (and some developmental factors), not by the elimination process. Selection can explain why certain individuals were eliminated, but it cannot explain the traits of the ones that were not eliminated. Karen Neander argues against the validity of this argument, which she calls the argument for the Negative View of selection, at least as far as cumulative selection is concerned (Neander 1995a). After a couple of rounds of exchanges without any sign of rapprochement, one gets the sense that there is some sort of miscommunication between Neander and Sober. One gets the sense that the opponents and the advocates of this argument may not mean the same by the term selection. My claim is that this disagreement is due to the fact that while Sober takes selection to be type-selection, Neander takes selection to be token-selection (see also Nanay forthcoming a, Nanay forthcoming b). This disagreement is the most explicit in the way they treat cumulative selection of continuous traits. Cumulative selection is a selection process, whereby changes accumulate through generations. The distinction between continuous and discrete traits is a standard one in evolutionary biology (I borrow the labels from Okasha 2007, p. 24, see also Mayr 1982, p. 512). The length of an organism s fingers is a continuous trait, the number of these fingers is a discrete one. Continuous traits come on a continuum; discrete traits do not. There is no possible value between 5 and 6 when it comes to the number of fingers, but there are many possible values between 5 and 6 centimeters when it comes to the length of these fingers. Selection for (or against) polydactyly is selection of discrete traits. Selection for (or against) long fingers is of continuous traits. I will limit the scope of my argument to continuous traits in what follows. For Sober, cumulative selection means that the change in trait frequencies, that is, the frequencies of certain property-types in a population, accumulates: in the first generation, we have 50% trait F and 50% trait G in a population. If there is selection for F, in a couple of generations, 5

6 the ratio can change to, say, 60%/40%. Importantly, F and G are property-types. As we have seen in Section II, selection for F, one of Sober s key notions, is always selection for a property-type (see Sober 1981 and Sober 1984a, pp for a detailed defense of this claim). Hence, in Sober s model of selection, property-types are competing with one another. But while the relative frequency of these property-types can and does change if there is selection, these property-types themselves do not change (or, at least, they do not need to change for there being a selection process). 7 The entire process can be described in terms of property-types (see also Sober 1980, p. 370). What is crucial for our purposes is that if we think of selection this way, then we can have selection without any change in the selected property-types. Selection increases the frequency of a specific trait type, but it does not modify this trait type. Hence, when explaining why this trait type is the way it is, selection is of no help. Sober is right when he suggests that what explains the characteristics of this trait type is mutation and what explains why an organism has an instantiation of this trait type is explained by inheritance. Selection does not play any role in this explanation. In short, Sober is right when he says that selection does not explain anything about why the selected property-type is the way it is. If we take selection to be type-selection, then Sober is justified in saying that selection fails to explain why specific individuals have the traits they have. 8 Take the alternative model, where selection is interpreted as token-selection. Here is a very simplified example (where I ignore sexual reproduction and limit the traits relevant for selection to only one). The neck size of giraffe x is 12 feet. She has three offspring, a, b and c. Giraffe a s neck 6 7 This point highlights that Price s equation, which is often used by population biologists as well as philosophers of biology as a way of describing natural selection (or the change of trait frequencies from one generation to another in general), presupposes type-selection. According to Price s equation, the change of trait frequencies from one generation to the other (the left hand side equation) is the sum of two independent values, Cov (ω, w) and E w (Δz), where Cov (ω, w) is supposed to be the result of natural selection and E w (Δz) is supposed to be the result of the transmission bias (Price 1972, Frank 1997, note that Samir Okasha attacks this causal interpretation of the two components (see esp. Okasha 2007, p. 26)). In other words, selection can happen even if the transition bias is zero. What this means is that while the relative frequency of these propertytypes can and does change if there is selection, these property-types themselves do not need to change in order for selection to happen. Price s equation takes selection to be type-selection (see (deleted) for a longer version of this argument). 8 One is tempted to take Sober s stance to be an instance of what Mayr calls typological thinking in spite of Sober 1980 s endorsement and analysis of population thinking. Here is what Mayr says about the way natural selection is explained by the typological thinker: Natural selection is meaningless to [the typologist], for it can [ ] only eliminate deviations from the type. For him natural selection is by and large only a negative process, able to eliminate the unfit but unable to play a constructive role (Mayr 1982, p. 517, see also Mayr 1982, p. 591 and Mayr 1959/1994, p. 328). I do not want to suggest that Sober was a typological thinker or that the type-selection/tokenselection distinction is the same as the one between typological and population thinking. A careful analysis of the relation between these two distinctions can be found in Nanay forthcoming a.

7 size is 10 feet, b s is 12 and c s is 14 feet. If the branches are very high up, then c is more likely to survive, than a and b. Thus, c makes it to the next generation and she has three offspring, d, e and f. As c s neck size was 14 feet, this will be the trait that gets transmitted to her offspring, who will have the neck size of 12, 14 and 16 feet respectively. Again, f, who has the longest neck is the most likely to survive. And so on. What we have here is a cumulative selection process: changes accumulate. But as this selection process is token-selection, it is not only the frequency of the selected traits that change: the selected traits in one generation will also be different from the ones in the previous generation: c s neck is longer than x s and f s is longer than c s. In the first generation, x s neck size was 12 feet, in the nth generation, the neck size of the individuals in the population will be close to the height of the lowest branches of the trees in the environment; it will adapt to the environment. 9 The explanatory scheme I sketched in the previous paragraph only talks about trait tokens. Importantly, it describes the selection for long neck in this population as token-selection. It is organism f s property-token of being 16 feet tall that is selected over organism e s property-token of being 14 feet tall. This selection process is not type-selection. Remember that type-selection explains why a certain type of organism is selected with the help of the property-types these organisms all have instantiations of. Token-selection, in contrast, explains why a certain specific organism is selected with the help of the property-instance of this specific organism. The explanatory scheme about f s long neck is clearly of the latter kind. But one may worry about whether my account of the selection for long neck is really an instance of token-selection. One way of describing the selection process in the giraffe population is that there is really only one trait type that is involved in this selection process: the trait type of neck size. But this trait type admits degrees. Shouldn t we conclude that this selection process is typeselection after all? The answer is that in the selection process I described specific instances of the property-type of neck size are selected over other specific instances of the same property-type. Again, it is organism c s property-token of being 14 feet tall that is selected over organism b s property-token of being 12 feet tall. Thus, when we consider the property because of which c survives and b dies (or because of which c has more offspring than b), this property is to be referred to as a property-instance and not as the general property-type of neck size. In short, if we take selection to be a process whereby property-types compete, as Sober does, then we would be justified to say that selection cannot explain why specific individuals have the traits they have. But if we take selection to be token-selection, then Sober s argument fails to apply and Neander s point that cumulative selection explains why specific individuals have the traits they 7 9 We need to be clear about what it is that adapts to the environment: trait tokens cannot be said to adapt to anything as they only exist for one generation. And as I aim to show that we can explain why a specific organism has the traits it has in terms of cumulative selection without appealing to property-types, we should hope that it is not property-types that adapt to the environment. The short answer is that the token traits of particular lineages adapt to the environment. A lineage is traditionally defined as "an entity that changes indefinitely through time as a result of replication and interaction" (Hull p. 327): in the giraffe example, the entity consisting of individuals x, c and f, etc. would constitute a lineage. If we take the sequence of the neck size of these individuals, they converge to the height that is optimal in the given environment: we can say that the token traits of this particular lineage adapt to the environment. See (deleted) on lineages as units of selection/adaptation.

8 have stays untouched (maybe positing some additional conditions, such as the limitation of environmental resources, see Nanay 2005, Nanay 2010). An important clarification is in order about the scope of this claim. The claim I made is about selection only. I am not claiming that the theory of evolution, or biology in general, is all about token traits or that we should (or can) re-write our biology textbooks without any mention of trait-types. Trait-types play a very important role in evolutionary theory (denying this would amount to denying the truth and relevance of important findings of population biology, such as the Lotka- Volterra equations). My claim has a much more limited scope: it is about selection only. 10 To sum up, the question about the explanatory power of selection is not about whether selection can explain why specific individuals have the traits they have, but about which account of selection can explain this. The one that takes selection to be type-selection (Sober s) can t. The one that takes selection to be token-selection (Neander s) can. Again, I do not want to jump to conclusions about which way of thinking about selection is the correct one (some jumping to conclusions will come in the last section). The aim at this point is to trace the respective consequences of thinking about selection as type-selection and as tokenselection. IV. The second consequence: The concept of fitness Distinctions about various ways of thinking about selection are very much in fashion among contemporary philosophers of biology. But the distinction between type-selection and tokenselection is not what the contemporary discussion is about. The contemporary discussion focuses on different ways of thinking about fitness. As evolution by natural selection is taken to be the heritable variation of fitness, depending on what conception of fitness we endorse, we arrive at different ways of thinking about selection. The two most influential contemporary debates about selection are about whether it should be taken to be population-level or individual-level process and whether it is a a causal or a statistical phenomenon (Matthen-Ariew 2002, Walsh et al. 2002, Millstein 2006, Brandon 2006, Bouchard- Rosenberg 2004, Rosenberg-Bouchard 2005, Ariew-Lewontin 2004, Stephens 2004). Depending on how one takes sides in these debates, one s concept of fitness will also be different. It has been pointed out that the concept of fitness is used in two different ways: as an ecological descriptor and as a mathematical predictor (Sober 2001, p. 319; this distinction may be traced back to Kitcher 1984, p. 50). Building on Sober s distinction Mohan Matthen and André Ariew made a distinction between vernacular and predictive fitness. 11 Vernacular fitness is a measure of the overall competitive advantage traceable to heritable traits (Matthen-Ariew 2002, p. 56). Predictive fitness, in contrast, is the expected rate of increase (normalized relative to others) of a gene, a trait, or an organism's representation in future generations" (Matthen-Ariew 2002, p. 56). Vernacular fitness plays a role in the informal 8 10 My claim is not even about the evolution of all traits. The evolution of radically new traits may be difficult to explain with the help of natural selection only (see Maynard Smith-Szathmáry 1995 for a summary on the major transitions of evolution ). My account is silent about how radically new traits evolve. 11 Ariew and Lewontin (2004) refer to these two concepts of fitness as 'Darwinian' and 'reproductive' fitness.

9 presentations of natural selection, whereas predictive fitness is used in mathematical formulations of population genetics. Vernacular fitness is a comparative measure, whereas predictive fitness is a quantitative one. Vernacular fitness is usually taken to be a cause of selection, whereas predictive fitness is taken to be a measure of selection, not its cause. Matthen and Ariew argue that we should dispose of the concept of vernacular fitness altogether. Others disagree. Some defend the concept of vernacular fitness by denying that natural selection is population-level process (Bouchard-Rosenberg 2004, Rosenberg-Bouchard 2005). And some others concede that it is a population level process, but maintain that it is a causal one (Stephens 2004, Millstein 2006). There are some further questions about fitness. Is it fixed throughout the organism s lifetime (Ramsey 2006)? In what way does it depend on the environment and how can we characterize the environment it depends on (Abrams 2007)? What kinds of entities are the bearers of fitness in the case of clonal organisms (Ariew-Lewontin 2004, esp. pp , Nanay forthcoming c)? It would be tempting to examine which conceptions of fitness would give rise to an account of selection that would count as type-selection and which ones would give rise to token-selection. But I will argue that the type-selection versus token-selection distinction has a more radical consequence with regards to these contemporary debates about the nature of fitness and of selection. In short, if we accept token-selection, we need to radically reinterpret (or even dispose of) the concept of fitness. I will use Elliott Sober s argument in order to show why this is so. Sober argued at length that fitness should not be thought of as a property of organisms, but rather of trait types (Sober 1981). As he says: the idea of trait fitness is absolutely essential to any formulation of evolutionary theory (Sober 1981, p. 162). Sober differentiates between trait fitness and organismic fitness. Trait fitness is the fitness of a trait type, whereas organismic fitness is the fitness of a specific organism. Sober argues that the fitness of a trait type cannot be reduced to organismic fitness. Hence, trait fitness, and with it, trait types, provide an essential component of the theory of natural selection. Here is the masterargument: If a single gene combination underlies two phenotypic traits, it is possible for the two phenotypic traits to differ in fitness, and yet for both to increase because of natural selection. This can happen when one of the phenotypic traits is selected for, while the other is neutral, or even slightly deliterious. The second trait will then be a 'free rider', increasing its representation in the population because there is selection for another trait with which it is pleiotropically linked. Notice that according to this understanding of pleiotropy, two phenotypic traits may be present in exactly the same organisms, and yet differ in fitness. If this is so, then trait fitness is not to be identified with the average fitness of organisms having the trait. (Sober 1981, p. 164). Sober s reasoning continues: the fitness of trait-types cannot be reduced to organismic fitness. But then fitness is an irreducible property of trait-types. Hence, any theory of natural selection must talk about trait-types. Without committing ourselves to Sober s conclusion (which would imply the endorsement of type-selection), it is possible to use his argument in a neutral manner. If fitness is to be understood as the fitness of property-types and if a theory of natural selection does not talk about property-types, we have to conclude that a theory of natural selection should not talk about trait- 9

10 fitness. Thus, if we hold that selection is token-selection and we accept Sober s argument from pleiotropy, we have to conclude that we need to give an account of selection without talking about the fitness of trait types. Is this a reductio argument? Some would say it is: evolution by natural selection is, by definition, the heritable variation of fitness (Lewontin 1970, 1980/1985, Maynard Smith 1987, Godfrey-Smith 2007), by which Sober means trait-fitness: therefore, every account of selection necessarily relies on talking about trait-fitness. Again, as Sober puts it, the idea of trait fitness is absolutely essential to any formulation of evolutionary theory (Sober 1981, p. 162). But there are two ways of resisting this reductio argument. First, we can deny that fitness is to be interpreted as trait fitness and, second, we can deny that evolution by natural selection is to be understood as the heritable variation of fitness. Sober s argument for the claim that fitness is to be understood as trait fitness is aiming at what he took to be the only plausible alternative: organismic fitness. And his argument from pleiotropy is a nice way of showing that the concept of organismic fitness has its problems. But Sober does not consider the possibility of talking about the fitness of trait tokens, which would be different both from trait fitness and from organismic fitness. Sober s argument from pleiotropy does not rule out the possibility of the fitness of trait tokens. As the fitness of the selectively advantageous trait token would be higher than the fitness of the neutral trait token, there seems to be no contradiction in reducing the fitness of a trait type to the fitness of trait tokens. In fact, as we have seen, a very general way of differentiating type-selection and tokenselection is the following: when a theory of selection talks about the difference between traits of the units of selection (whatever they may be) that are responsible for (or, that explain) which units of selection will get to be selected, then we need to be able to tell whether these traits are to be taken as trait types or as trait tokens. If we take the concept of fitness to be explanatory at all, then proponents of type-selection should interpret fitness as the fitness of trait types (as Sober in fact does), while the proponents of token-selection should interpret it as the fitness of trait tokens. I am fully aware of how revisionary this concept of fitness as the fitness of token traits is and it would take more than a couple of paragraphs to give a convincing account of this revisionary concept. Further, even if one accepts that the concept of fitness is to be understood as the fitness of trait tokens, it does not follow that this concept of fitness can be plugged into the heritable variation of fitness account of evolution by natural selection in an unproblematic manner. 12 In other words, even if the concept of the fitness of trait tokens is an unproblematic concept in and of itself, it may not satisfy the various desiderata a concept of fitness is generally taken to satisfy (for three such important desiderata, see Ariew-Ernst 2009, who, incidentally, claim that no account of fitness can satisfy all of these). Although I myself find the idea of the fitness of trait tokens a very promising one (see (deleted)), I think that there is a much safer bet for those who are leaning towards the idea of token-selection. This safer bet is the following. Evolution by natural selection does not have to be conceived of as the heritable variation of fitness. Darwin s original formulation of the theory of natural For simplicity, I will say the heritable variation of fitness account of selection (rather than the heritable variation of fitness account of evolution by natural selection) in what follows.

11 selection, for example, does not talk about the heritable variation of fitness. In fact, he does not talk about fitness at all (although he does talk individuals as being fitter than other individuals). 13 Importantly, there is a widespread way of formulating a theory of natural selection without using the concept of fitness. According to this account, selection consists in repeated cycles of two separate processes. As Ernst Mayr says, natural selection is actually a two-step process, the first one consisting of the production of genetically different individuals (variation), while the survival and reproductive success of these individuals is determined in the second step, the actual selection process (Mayr 1991, p. 68, see also Mayr 1982, pp , Mayr 2001, p. 117 and Mayr 1978): David Hull calls these two steps replication and interaction (Hull 1981, Hull 1988, Hull et al. 2001). He defines selection as: The repeated cycles of replication and environmental interaction so structured that environmental interaction causes replication to be differential. (Hull et al. 2001, p. 53) This replication-interaction account of selection was introduced as an improvement on the heritable variation of fitness account, and it is supposed to clarify a number of details left implicit in the heritable variation of fitness account. More precisely, the replication-interaction account has been thought to help us to understand what is at stake in the units of selection debate: if selection is replication plus interaction, then we should not talk about the units of selection, but rather the units of replication and the units of interaction, which may not be (and in fact most often are not) the same. The replication-interaction account has different versions and the most widespread of these is highly problematic. I will distinguish two versions of this account, the replicator-interactor account and the property-replication account and argue that the latter is a good bet for those who are looking for an alternative to the heritable variation of fitness account. According to the first version of the replication-interaction account, replication is the copying of an entity, the replicator. 14 Hull defines the replicator as an entity that passes on its structure largely intact in successive replications (Hull 1988: 408, see also Godfrey-Smith 2000, Brandon 1990, Nanay 2002 on the concept of replicator). The unit of interaction, interactor, on the other hand, is defined as the entity that interacts as a cohesive whole with its environment in such a way that this interaction causes replication to be differential (Hull 1988: 408). I will call this version of the property-replication account the replicator-interactor account as it identifies replication with the copying of an entity, the replicator. In the last decade or so, some philosophers have argued against this replicator-interactor account of selection (Okasha 2007, pp , Godfrey-Smith 2007, p. 515, Godfrey-Smith 2009, Avital & Jablonka 2000, p. 359, Jablonka & Lamb 1995, Richerson & Boyd 2005, chapter 3, Griesemer 2000, pp , Griesemer 2002, p. 105). Their main claim is that the copying of In addition, what Darwin meant by fit and fitter is very different from the way the term is used by contemporary biologists. See Paul Note that the replicator does not have to be the gene. In fact, an important motivation for introducing this concept was to extend the circle of potential replicators to non-biological entities, such as memes.

12 replicators is not necessary for selection; hence, selection cannot consist of repeated cycles of replication (conceived of as the copying of replicators) and interaction. There are ways of transmitting information (extragenetic inheritance, cultural transmission) that do not count as replication but that are (given other conditions) sufficient for selection (Okasha 2007, p. 15, Avital & Jablonka 2000, p. 359, Jablonka & Lamb 1995, Richerson & Boyd 2005, chapter 3). Samir Okasha summarizes this line of objection: evolutionary changes mediated by cultural and behavioural inheritance cannot be described as the differential transmission of replicators (Okasha 2007, p. 15). To put this objection in more general terms, selection can happen if there is sufficient phenotypic parent-offspring resemblance. Replication is not needed (Okasha 2007, p. 15). An example: 'maternal effects': cases in which large mothers have large offspring as a result of laying eggs with larger food reserves (Uller 2008). It is important that these problems are problems for the replicator-interactor account and not for the replication-interaction account in general. Remember that the original alternative to the heritable variation of fitness account was the view that selection consists of repeated cycles of replication and interaction. It is an additional requirement that replication should be thought of as the copying of an entity, namely, the replicator. We may be able to salvage the general gist of the replication-interaction account if we deny that replication is the copying of an entity. We could conceive of replication as the copying of property-instances (Nanay forthcoming a, see also Nanay 2002, p. 113). The hope is that this version is not vulnerable to the objections raised against the replicator-interactor account. I will use the term property-replication account for this version of the original replication-interaction account to contrast it with the replicator-interactor account. It is important to clarify the difference between these two versions: about what is meant by entities and properties here. The cup in front of me is an entity. It has lots of properties, some interesting, some others less so. Its colour is one property, its shape is another one, etc. Thus, the copying of an entity and the copying of one of the properties of this entity are very different processes. Properties are always properties of entities of course. But it is possible to copy a property of an entity without thereby copying the entity itself. The claim is that replication is the copying of properties: we can have a replication process without there being a replicator that is being copied. The definition of replication is then the following (Nanay forthcoming c, Section 4): property P of object a is a replica of property Q of object b if and only if: (i) P is similar to Q and (ii) Q is causally involved in the production of P in a way responsible for the similarity of P to Q. An important feature of this definition is that a and b are not necessarily objects of the same kind. Object b may be an apple and object a a color photograph of this apple. The color of the photograph can be a replica of the color of the apple under my definition, but this does not mean that the objects themselves are replicas or copies or replicators in the old sense of the word. This notion of replication is very weak: many non-biological copying processes, like photocopying, will also qualify as replication. Is this a problem? No. The same is true of the traditional concept of replication as the copying of replicators (Godfrey-Smith 2000, Nanay 2002). Importantly, any account that conceives of selection as the repeated cycles of replication and interaction needs to acknowledge that not every replication process will be particularly interesting from an evolutionary point of view. But this is what we should expect: the notion of replication is only the starting point for an account of selection. Further additional criteria need to be met in order for replication to lead to selection: replication needs to give rise to an interaction process that makes the next round of replication differential. 12

13 Note that P and Q in this definition of replication are property-instances. Thus, this account of replication does not presuppose talk of property-types. This account of replication could be, and has been in Nanay s original account, formulated in terms of tropes: properties that are logically incapable of being instantiated by two different particulars at a time. The same goes for the concept of similarity: we can talk about the similarity between two property-instances without evoking any property-types. 15 In short, selection conceived as property-replication is token-selection, not type-selection. How can this property-replication account handle the objections raised against the replicator-interactor account? First, according to the property-replication account, both extragenetic inheritance and cultural transmission can count as replication. Nothing in the definition of replication suggests that the replicated property needs to be a property of the DNA. Thus, extragenetic properties can replicate just as the properties of the DNA can. If property P of the offspring is similar to property Q of the parent and the latter is causally responsible for this similarity, then we do have replication, regardless of whether these properties can be called genotypic or not. More generally, if we accept the property-replication account, then phenotypic traits can replicate. Take the maternal effects example I mentioned above. According to the propertyreplication account, there is a property that replicates in this case: the property of being large. The offspring s instantiation of this property is similar to her mother s (in as much as the degree of similarity between the size of the two individuals is higher than it is between the size of two randomly chosen individuals in the population) and her size is causally responsible for this similarity. Thus, we do have selection in this population, but we also have replication. We do not have replicators though. It is not the aim of this paper to work out the details of a plausible alternative to the heritable variation of fitness account of selection. But it is important to point out that rejecting the heritable variation of fitness account of selection does not mean that we are left without any plausible alternatives. A final worry: The replication-interaction account was introduced as being consistent with (in fact, as an addition to) the heritable variation of fitness account. Can we make sense of the concept of fitness in the framework of the property-replication account? We can, but it will be the fitness of the replicated property-instances. In the original replicator-interactor account, fitness was understood as the fitness of concrete particulars: typically of interactors. In short, in the original replicator-interactor account we could talk about organismic fitness. As the property-replication account does not talk about entities at all, but only about property instances, if we want to talk about fitness at all in this account, the most plausible candidate would be the fitness of (genotypic and phenotypic) trait tokens. But, importantly, one can accept the property-replication account without using the concept of fitness at all. VI. Conclusion: How to decide? The main aim of this paper was to identify a major ambiguity in talking about selection and trace its consequences in a number of contemporary debates in philosophy of biology. In conclusion, I also That is exactly what trope nominalists do when they talk about the similarity between tropes, see, e.g., Campbell 1990, Bacon 1995, and Schaffer 2001 for an overview.

14 want to suggest which way we should resolve this ambiguity. My suggestion will be that we should think of selection as token-selection. Let us summarize what the distinction between type-selection and token-selection has entailed: 14 Type-selection Selection cannot explain the traits of specific organisms Heritable variation of fitness account Token-selection Selection can explain the traits of specific organisms Replication-interaction account Thus, we seem to have two package deals when thinking about selection. 16 We can choose to think of selection as type-selection. This way of thinking about selection will be consistent with the heritable variation of fitness account of selection. However, if we think of selection this way, its explanatory power will be limited as it will not be able to (partially) explain why the traits of specific organisms are the way they are. The other package deal is to think of selection as token-selection. This way of thinking about selection may or may not be consistent with the heritable variation of fitness account of selection. On the other hand, if we think of selection this way, selection will (partially) explain why specific organisms have the traits they have. We have a trade-off here. We can say that the advantage of type-selection is that we get to keep the concept of fitness, and the disadvantage is that the explanatory power of the concept will become limited (as it will be explanatorily irrelevant in explaining why specific organisms have the traits they have). If, in contrast, we choose token-selection, then we may have to do without the concept of fitness (or we may have to reinterpret it radically), but, and that is the advantage, the concept of selection will be explanatorily very relevant as it will explain why specific organisms have the traits they have. I want to suggest that this trade-off should force us to take selection to be token-selection. Endorsing token-selection does not strictly entail giving up the concept of fitness we may be able to salvage the concept by reinterpreting it as the fitness of trait tokens. But even if we do give up the concept of fitness, it is not really a disadvantage: we have a perfectly coherent way of describing selection without talking about fitness. In fact, it has been argued that the replication-interaction account may have independent explanatory advantages over the heritable variation of fitness account (see Nanay forthcoming c). In contrast, the fact that taking selection to be type-selection will lead to selection being explanatorily superfluous in explaining token traits should worry those who take it to be the central tenet of evolutionary theory to explain the apparent teleology of nature in terms of a dumb causal process of selection. 16 I argued that type-selection cannot explain why a specific organism has the traits it has. I also argued that the best bet for token-selectionists is to endorse the replication-interaction account and that token-selection can explain why a specific organism has the traits it has. I have not argued that type-selection entails the heritable variation of fitness account of selection. In principle, the replication-interaction account could be made consistent with the idea of type-selection. As this view would combine the respective perceived disadvantages of the two package deals, I will not discuss it here.