Common misunderstandings of

memes (and genes)

The promise and the limits of the genetic analogy to cultural transmission processes

Francisco J. Gil-White

fjgil@psych.upenn.edu ; http://www.psych.upenn.edu/~fjgil/

Assistant Professor of Psychology
University of Pennsylvania
3815 Walnut Street, Suite 400
Philadelphia PA 19104-6196

Word Count: Abstract = 248 words; Main text = 12,313; References = 1,154; Entire Text = 13,903.

Short Abstract: ‘Memetics’ suffers from conceptual confusion and not enough empirical work. This paper attempts to attenuate the former problem by resolving the conceptual controversies. I criticize the overly literal insistence—by both critics and advocates—on the genetic analogy, which asks us to think about memes as bona-fide replicators in the manner of genes, and to see all cultural transmission processes as ultimately for the reproductive benefit of memes, rather than their human vehicles. A Darwinian approach to cultural transmission, I argue, requires neither. It is possible to have Darwinian processes without genes, or even close analogues of them. The cognitive mechanisms responsible for social-learning make clear why.

Long Abstract: ‘Memetics’ suffers from conceptual confusion and not enough empirical work. This paper attempts to attenuate the former problem by resolving the conceptual controversies, which requires that we not speculate about cultural transmission without being informed about the cognitive mechanisms responsible for social learning. I criticize the overly literal insistence—by both critics and advocates—on the genetic analogy, which asks us to think about memes as bona-fide replicators in the manner of genes, and to see all cultural transmission processes as ultimately for the reproductive benefit of memes, rather than their human vehicles. A Darwinian approach to cultural transmission, I argue, requires neither. It is possible to have Darwinian processes without genes, or even close analogues of them. The insistence on a close genetic analogy is in fact based on a poor understanding of genes and evolutionary genetics, and of the kinds of simplifications that are legitimate in evolutionary models. Some authors have insisted that the only admissible definition for a ‘meme’ is ‘selfish replicator.’ However, since the only agreement as to the definition of ‘meme’ is that it is what gets passed on through non-genetic means, only conceptual confusion can result from trying to make a hypothesis into a definition. This paper will argue that, although memes are not, in fact, ‘selfish replicators,’ they can and should be analyzed with Darwinian models. It will argue further that the ‘selfish meme’ theoretical calque imported from genetics does much more to distort than enlighten our understanding of cultural processes.

KEYWORDS: Cultural transmission, culture, evolutionary genetics, meme, memetics, replicator, social-learning.

Given an incredibly simplistic notion of genes, memes are not in the least like genes. . .One problem with interdisciplinary work is that any one worker is likely to know much more about one area than any of the others. Geneticists know much more about the complexities of genetics than of social groups. Conversely, anthropologists and sociologists tent to be well-versed in the details of social groups. To them genetics looks pretty simple.—Hull (2000:46)

Many of the claims made about memes could be false because the analogy to genes has not proven productive.—Aunger (2000:8)

Introduction

Should we demand that ‘memes’ be exactly like genes if we are to apply Darwinian tools of analysis to culture? No.

The formal similarities between genes and what (after Dawkins 1989[1975]) are now called ‘memes’—the units of cultural transmission and evolution—suggest cultural transmission processes are ripe for Darwinian analysis. A vigorous debate is emerging over how to think about ‘memes’ (for a recent compendium of views see Aunger 2000). This is an evolutionary but also cognitive issue because memes are stored in human brains.

New fields will always use analogies and borrowed yardsticks, and these can be a source of fresh insights, but also cause misunderstanding. The yardstick which requires ‘memes’ to be essentially identical to genes if Darwinian analyses are to apply is a source of much confusion. This regrettable error is advanced by both critics and defenders of ‘memetics’ and—to boot—the specific arguments are often based on a poor understanding of genes and evolutionary genetics. The standard chosen is therefore not only erroneous but would indict evolutionary genetics as well (genes, it turns out, are not sufficiently like ‘genes’ either!).

There are too many insistent definitions of ‘meme’—typical in a new research program given that careers (especially in social science) are often boosted by getting particular definitions adopted. The prize is large because the term ‘meme’ is on everybody’s lips. If definitions were advanced only with conceptual progress in mind, this would be fine. But here, more than in other fields, the various protagonists must be aware that the contest is memetic, yielding a tendency to produce ‘catchy’ definitions that ‘sell well’ at the expense of conceptual advance and scientific utility.

The definition of meme as a ‘replicator’ is very catchy. Introduced by Dawkins (1989[1975]), and developed by Dennett (1995) and Blackmore (1999, 2000), it has helped mobilize our intuitions for population-driven processes involving genes, which are bona-fide replicators producing perfect descendant copies of themselves. As a heuristic device there is nothing wrong with this. But as a statement of what Universal Darwinism is—i.e. find a replicator, then apply Darwinism—it is a garden path. And a tortuous one. Consider that Blackmore (2000:26) says “memes are replicators,” but on the preceding page claims that, “As long as we accept that people do, in fact, imitate each other, and that information of some kind is passed on when they do, then, by definition, memes exist.” By definition? By definition ‘replication’ takes place when perfect copies are produced, not when “information of some kind [my emphasis] is passed on…” Proponents of memetics who uphold the ‘gene standard’ must weaken and mutilate the meaning of ‘replication’—which they take to result from ‘imitation’—in order to claim that memes are ‘replicators’ and that, since they are, Darwinism applies. They insist, therefore, not on the concept ‘replication’ but on the word, the use of which is assumed magically indispensable to the possibility of Darwinian science. But this is absurd.

Critics of memetics who also uphold this same ‘gene standard,’ on the other hand, stick closely to the definition of ‘replication’ as ‘perfect copying,’ and this is good (why butcher the language?). However, they fetishize the concept, for they accuse that the poor copying fidelity of memes—i.e. memes are not, after all, replicators—supposedly makes Darwinian analyses to culture inapplicable in principle. In my view, these critics, as much as the proponents, are chasing a mirage. Replication is not necessary for cumulative adaptations through selective processes (Boyd & Richerson 2000:153-158), and is therefore not the standard both critics and proponents are looking for. Replication is a red herring.

The ‘selfish meme,’ like its ancestor the ‘selfish gene,’ is another catchy idea. It answers the question cui bono? by saying that the unit being transmitted—the meme—is the ‘entity’ which ‘benefits’ in the cultural selective process. Again, this began with Dawkins (1983:109), who stated that a meme is “a unit of cultural inheritance…naturally selected by virtue of its…consequences on its own survival and replication,” and again developed by Dennett and Blackmore. In this picture “We humans. . .have become just the physical ‘hosts’ needed for the memes to get around. This is how the world looks from a ‘meme’s eye view’” (Blackmore 1999:8). In a manner parallel to the ‘gene’s eye view,’ we are here supposed to interpret every meme that succeeds at proliferating as having done so by dint of being well designed for proliferation. Cultural selection is reduced to the continuous editing of meme content until memes end up optimally designed for colonizing human brains. I will argue that only some rather specialized kinds of memes satisfy this analytical calquing from genetics to culture. But, again, this does nothing to wreck the applicability of Darwinian analysis or the usefulness of thinking in terms of memes—it merely indicts the fetishizing of the genetic analogy. Reducing cultural transmission to ‘selfish memes’ requires that we ignore much of social-learning cognition and miss most of the picture.

It should be obvious this far that I feel no compunction to accept Dawkins’ (1989), Dennett’s (1995), and Blackmore’s (1999) definition of ‘meme’ as selfish replicator. A recent compendium of views (Aunger 2000) makes it clear that neither do many others. It is best not to insist on a research program that rises or falls on whether memes defined as selfish replicators exist. That is a careerist semantic game that tries to assume or impose as a definition something that must be investigated, and such a game does not advance the science of cultural transmission—a science that will be carried out anyways because we must.

Most of us seem to accept the Oxford English Dictionary’s definition, which says: ‘an element of culture that may be considered to be passed down by non-genetic means.’[1] So ‘selfish replicator’ I will treat as a hypothesis about what the stuff that gets transmitted through non-genetic means is like. The relevant questions, then, are: (1) does this stuff look like a selfish replicator?; (2) If not, does this really make Darwinian analyses of culture impossible? Related questions are: (3) if they don’t replicate, is it impossible to find the boundaries of memes?; and (4) can we import from biology, willy-nilly, the ‘selfish gene’ idea? I will answer “no” to each of these questions. But I will still call what is transmitted culturally a ‘meme,’ and so—I will bet my house—will everybody else. The term ‘meme’ has already been selected for, so rather than forcing its meaning to coincide with a particular hypothesis about cultural transmission, let us do some science.

I. What is required for genetic cumulative evolution?

Darwinian systems involve simple and blind algorithmic processes that nevertheless produce gradual accumulation of (sometimes very complex) adaptive design. They have three main requirements: information must be able to leave descendant copies (inheritance), new information should be routinely generated by some process (mutation), and there should be forces responsible for causing some items of information to leave more descendants than others (selection).

Genes satisfy all three. They are inherited through reproduction; new genes are routinely created because of occasional copying mistakes, or ‘mutations’, during DNA duplication; and a gene, through its effect on its carriers, affects the probability that it will increase in number. Thanks to selection and inheritance, when a particular gene causes increased reproductive success, more copies of it are passed on, and its relative frequency in the population increases (absent frequency dependent effects, eventually the whole population will have it). Thanks to mutation, new alternative genes get generated which occasionally amount to improvements, allowing the population to continue to evolve.

Cumulative genetic adaptations are possible because (1) genetic mutations typically introduce incremental rather than massive changes, and (2) the mutation rate for genes is low. It is these latter two requirements for cumulative evolution in genetic systems that inform some scholars’ intuitions that ‘replication’—that is, high-fidelity copying—is crucial to cumulative evolution through memes as well, which intuitions then damn Darwinian approaches to culture if memes are found not to replicate. For this reason these two requirements deserve further attention here.

Massive change is by definition the opposite of the accumulation of design, where each successive design change is a minor alteration on the margins of the previous template. But should we expect organic evolution to consist of small, incremental changes? Yes. The space of maladaptive designs is vast relative to the space of adaptive ones, so random changes to any current design (and mutations are random) are unlikely to cause adaptive improvements. Imagine that a monkey types a character at random as I am writing this essay. Will it improve? Without vanity, I can say that the chances are exceedingly low. A random typo is unlikely to yield English, let alone better English. But should the monkey press a key which launched a program to rearrange all of the letters in my essay, then he would be infinitely less likely to improve it—slim as his chances were anyway. In population-driven processes, for a novelty to last longer than an instant, it is typically constrained to cause a small modification.

Mutations must also be infrequent because, unless designs are relatively stable across time, we cannot get cumulative evolution. Suppose the offspring of A’s are mostly non-A’s. Even if A reproduces better than its competitors B and C, this cannot have an evolutionary consequence because the information responsible for A’s reproductive prowess is almost always lost after reproduction. On the contrary, if an A typically begets another A, then A’s higher reproductive success will soon make everybody in the population an A (absent frequency-dependent effects). Later, when a rare mutation results in a slight improvement to ‘A design’—let us call the new design A°—these A° mutants will outreproduce mere A’s and the population changes again (but only slightly).

This covers the intuitive basics of genes as replicators allowing for cumulative cultural evolution. But how similar to genes are memes? Well, memes certainly have the properties of inheritance, mutation, and selection. We constantly acquire and learn things from each other through social interaction, so in a broad sense at least it makes sense to say that the information I possess can create a ‘descendant copy’ in you (inheritance). People can make mistakes when acquiring information, and can also have stupid or bright novel ideas, which leads to new items of information (mutation). And some ideas are more popular than others, so they are copied more, stored longer, and rebroadcast more often, which in turn means they leave more descendants than competing ideas (selection). What makes some ideas more ‘popular’ than others are the properties of human social-learning psychology. This is not the only force acting to favor certain memes over others, but it is a very important one and I shall restrict myself to it here.

So much for intuitively stated formal similarities. The devil, as usual, lurks in the details. To many critics, the dangerous phrase above is “in a broad sense…information can create a ‘descendant copy.’” How broad? How similar must ancestor and descendant memes be?

Some assert that selectionist approaches cannot work because memes are not true replicators, making cumulative evolution impossible (e.g. Sperber 1996; Boyer 1994). Others, however, have not considered this a problem and proceeded to build Darwinian selectionist models that in their fundamental assumptions are quite similar to those used in evolutionary genetics, but adapted for cultural idiosyncrasies (e.g. Boyd & Richerson 1985; Lumsden & Wilson 1981; Cavalli Sforza & Feldman 1981; for a review, see Feldman & Laland 1996). As Laland & Odling Smee (2000:121) put it: “For us, the pertinent question is not whether memes exist. . .but whether they are a useful theoretical expedient.” Their critics, however, will counter that such models do not help us explain human cultural processes because the units employed are nothing like what exists in real-life cultural transmission. To find out who is right, we need first to examine closely whether it matters that memes are poor replicators.

II. Do memes mutate too much?

To Dan Sperber (1996), contagious pathogens such as viruses are a better analogy than genes for understanding the spread of cultural information. Populations of brains are infested in successive ‘epidemics’ of memes (which Sperber invariably calls ‘representations’—a favorite term in the cognitive literature). He cautions, however, that the analogy can be taken only so far.

. . .whereas pathogenic agents such as viruses and bacteria reproduce in the process of transmission and undergo a mutation only occasionally, representations are transformed almost every time they are transmitted. . .—Sperber (1996:25)

. . .recall is not storage in reverse, and comprehension is not expression in reverse. Memory and communication transform information.—Sperber (1996:31)

For example, does anybody ever retell a story exactly? No, and this is Sperber’s point.

In the case of genes, a typical rate of mutation might be one mutation per million replications. With such low rates of mutation, even a very small selection bias is enough to have, with time, major cumulative effects. If, on the other hand, in the case of culture there may be, as Dawkins [1976] acknowledges, ‘a certain “mutational” element in every copying event,’ then the very possibility of cumulative effects of selection is open to question.—Sperber (1996:102-103)

It is important to see exactly what the argument is. Genes are very stable across generations because they very rarely make copying errors during duplication—hence, for the most part, they replicate. As observed above, this allows cumulative genetic adaptations to emerge because small, cumulative changes can only be added if there is an overall template which remains—for the most part—stable. There is nothing absolute about the acceptable rate of mutation, of course. Rather, this is always relative to the strength of selection. For example, even if there is a moderate rate of mutation, cumulative evolution will still happen if the selective process culls suboptimal variants fast enough that the favored design is stable at the populational level, and from generation to generation. G.C. Williams (1966) made this principle famous in his definition of an ‘evolutionary gene,’ which is “any hereditary information for which there is a favorable or unfavorable selection bias equal to several or many times its rate of endogenous change.” This definition was taken willy-nilly by Dawkins and applied to his definition of the ‘meme,’ and recently stated very clearly by Wilkins (1998:8):

A meme is the least unit of sociocultural information relative to a selection process that has favorable or unfavorable selection bias that exceeds its endogenous tendency to change.

Sperber is accepting this move to assume (1) that ‘replicators’ are the things to look for; (2) that Dawkins’ reinterpretation of Williams gives the universal definition of a replicator, and (3) that Darwinian analyses will apply to memes only if they can satisfy this definition. In fact, Sperber eagerly forces the issue by ruling that any other conceptualization of ‘the meme’ is trivial (Sperber 2000:163). His stance is therefore that cumulative adaptations through cultural selection are possible only if we can find bona-fide cultural replicators. But memes in fact mutate in every single act of transmission, so he concludes that cultural selection cannot conceivably act fast enough because the meme’s dizzying rate of endogenous change creates a ceiling effect (Atran 2001 echoes this argument). Sperber therefore believes that we must understand how cognitive processes of information storage and retrieval cause mutations in particular and systematic directions. With this information, we can build (orthomemetic?) models of directed mutation rather than selectionist models (Sperber 1996:52-53, 82-83; 110-112) of cumulative change.

There is some irony in this. Hull (2000:47) quotes the above definition by Wilkins approvingly as a starting point for a science of memetics that he optimistically believes to be possible, although he fully expects “howls of derision” to come from unreasonable critics who will accuse this definition of not being sufficiently “operational.” Something very different has already happened, however! A prominent critic of selectionist approaches to culture—Sperber—has eagerly embraced that very definition in order to explain why selectionist approaches to culture are supposedly impossible.

It would seem as though either Hull or Sperber must be wrong, for they agree on how to define units of cultural processes that would be legitimately Darwinian, but they reach exactly opposite conclusions as to whether human culture passes or fails the test. However, I believe they are both mistaken because they are sparring on the wrong battlefield. The standard chosen, rather than enlighten, blinds us to the general requirements for a Darwinian system by insisting narrowly on the terms of one particular solution to them—the genetic one—as if this were the only possibility.

I shall accept Sperber’s point that the mutation rate for memes is 1: they mutate in every act of transmission. And I will agree, too, that often they are systematically biased. But this is neither here nor there. What matters is how big these mutations are, and how strongly biased in particular directions, as we shall see.

III. ‘Replication’ is a red herring

Sperber’s argument may seem intuitively appealing, but I think it is specious. Near-perfect copying fidelity is certainly important in genetic selection, but it is not a requirement for any Darwinian system. If the high rate of mutation is not the meme’s only distinction, then perhaps its other idiosyncrasies make it possible for regularly imperfect—or even invariably imperfect—meme-copying to support the emergence of cumulative adaptations.

I shall make the case with a toy example. But first, a few preliminaries. In genetics, a ‘locus’ is the physical location of a ‘gene’ on a chromosome. This is where the information ‘for something’ can be found. If we are talking about, say, the ‘eye-color’ locus, then the gene found there may be the ‘brown-eye’ gene, or the ‘blue-eye’ gene, and so forth. What is the analogue in memetic transmission? For example, imagine something like, say, a tennis-serve ‘locus’. Whatever is in your tennis-serve locus causes your behavior when beginning a new point in tennis. There are in principle a vast number of different behaviors that people could store at the tennis serve locus (just as there are many different sequences of nucleotides that may be stored at the chromosomal eye-color locus). Waving hello to your mom, or baking a bread, would be ruled illegal by the judges, but in principle this does not prevent you from storing such information at that locus (just as a random and useless sequence of nucleotides could, in principle, be stored at the eye-color locus).

It hardly matters that the tennis-serve locus may not be physically located in the same piece of brain for every individual. To insist on this is to push the genetic analogy to an absurd extreme where it begins to straight-jackets thought rather than inspire insights. The relevant and crucial similarity is functional, not physical: if individuals recognize that an item of information becomes relevant when, in a game of tennis, a new point is beginning, then the ‘cultural locus’ has all the requisite functional similarity to the genetic locus that we need. In cognitive terms, the cultural ‘locus’ is a tag plus retrieval function—it is a matter of categorization rather than physical location in the brain. The information retrieved at the start of a new tennis point is that which I tag as ‘tennis serve’. Waving to my mom or baking a cake have not been tagged this way (even though, in principle they could be), and, since they have not been, they do not compete to ‘occupy’ my tennis serve ‘locus.’ The true alleles of my current serve, therefore, are other behaviors which I also tag as ‘tennis serves’ because some individuals in the population perform them in the context of beginning a point in a tennis match. I may choose to acquire one of these later on, and in so doing will replace my current serve.

These obvious functional similarities readily dismiss the criticism that, because memes do not have the same kind of physical reality as genes, selectionist approaches to culture are a nonstarter. We are not talking here of the duplication of exact neuronal structures analogous to the duplication of exact nucleotide sequences in DNA, but we are speaking of the duplication of a certain behavior, understood to belong in a certain context, and in competition with other behaviors also understood to be candidates for the same context. The lack of similarity in the material basis of genes and memes is not a problem.

A. The right mix of stability and variation

To see whether a meme’s inability to properly replicate makes cumulative cultural adaptations are impossible, we must examine the full spectrum of theoretical possibilities.

Suppose that in our population, Bob’s serve is the most attractive, and seeing it performed gets people excited to make changes in their own tennis-serve loci. There is a continuum of different things that could happen, bounded by two extremes. At one extreme—replication—people acquire precisely the same content that is in Bob’s own locus. For example, you acquire the exact same top-spin service with a slight jump that Bob favors. At the other extreme—causation of random changes—people rewrite the information in their locus such that it typically bears no resemblance to Bob’s serve. Here, for example, you might ‘write’ into your tennis serve locus the idea that you should wave at mom when up to serve.

Please take note that I am not following the information in the brain here, although of course it is necessary for the process. What I am keeping track of here is the actual behaviors, and I am completely ignoring the question of what particular information content in the brain may be causing them. The latter is not always unimportant (Gil-White 2002a), but it does not concern me in the present analysis, and it is irrelevant to the points I will make. When I talk about ‘replication failure,’ what I mean here is the inability of the copier to perform a serve that is identical to Bob’s.

Let us look first at the causation of random changes. This will look silly, but we cannot gain the proper insights until we examine the full spectrum of possibilities. As silly as it sounds, suppose I put ‘wave at mom’ in my tennis serve locus after watching Bob’s top-spin serve. You will put randomly different, but typically equally dissimilar, information to Bob’s serve in your own tennis serve locus. What will happen? We are assuming that it is the content (i.e. the sequence of motions) involved in Bob’s serve that make it attractive, in turn precipitating changes in the tennis-serve loci of other people. Given this, I myself (who now wave at mom when I ‘serve’)—and all others who randomly changed the information at their tennis serve loci after watching Bob—are not similarly beacons of change; our new ‘tennis serves’ look nothing like Bob’s and they therefore get nobody excited (and mostly irritate the judge because they are not admissible). Bob’s serve has not become more common, nor has the mean serve of the population moved in the direction of Bob’s serve. Since evolution is about statistical changes in a population, the fact that this process does not produce reliable directional movement in the population’s mean serve implies that this process cannot lead to cumulative design changes. After all, the first requirement for cumulative adaptive design is the possibility of directional change.

Now consider the other extreme. This will look silly too. Here, watching Bob’s serve produces verbatim replicas in observers’ tennis-serve loci. People copy perfectly, so there is never any mutation—not ever. What happens? Because Bob’s is the most attractive serve, all of the people who now have Bob’s serve in turn become models for other people, who again copy the serve precisely and so forth. Bob’s serve spreads until everybody is serving identically. Here, too, selection cannot lead to cumulative design changes because the serves are all identical to Bob’s. The future will be spent forever more serving exactly like Bob, by everybody. No other serves will ever emerge because nobody ever makes a copying mistake.

We see that at either end—random changes, or perfect replication (100% copying fidelity)—there can be no accumulation of adaptive design. So this can occur only somewhere ‘in the middle’, where descendant changes are relatively similar to the ‘parent’ stimulus, but somewhat different. There are two ways in which this can happen: (1) descendant serves are always identical to the parent, except that every once in a long while there will be an accidental difference; or (2) the descendant serves are always accidentally different from the parent serve, but jump around relatively closely to the average of copying accuracy. In both cases we get more attractive future serves by making marginal changes to Bob’s, which in turn makes the marginally improved serve the new model (and this is what allows for cumulative adaptation). I examine each in turn.

(1) Copying involves mistakes only once in a long while. Here the information ‘written’ in a person’s tennis-serve locus is a pristine replica of the ‘parent’ serve. There is a very small probability of replication failure so, very rarely, a random modification results. Such modifications will typically make Bob’s serve less effective because a tennis serve is a complex behavior where many variables must be kept within narrow ranges to ensure success. I am assuming that only effective serves are attractive, and so most random changes will result in less attractive serves. But very, very occasionally, a random copying mistake begets a more effective—and therefore more attractive—serve, which then displaces Bob’s as people now begin making perfect replicas of the improved serve. Many iterations of this cycle will lead to ever better serves. I have just described a process of accumulation of adaptive design emerging from cultural transmission that is exactly parallel to cumulative genetic evolution by natural selection. Sperber (1996) claims that in order for selection to produce cumulative design in cultural transmission, the process should look like this. But let us take a look at a rather different process.

(2) Copying always involves mistakes, but around an average of perfect accuracy. This process is illustrated below in fig. 1. Every time somebody sees Bob’s top-spin serve, the goal is to copy it exactly, but there is always some error, and thus there is almost never a perfect copy. However, the errors are relatively small and not biased in any particular direction, so that Bob’s serve is obviously the template for all descendant serves. In this scenario, replication is the occasional exception. However, the population’s mean serve is still Bob’s, even if no individual serve is a true replica. The errors amount to a constant introduction of modest variations, from which a serve superior to Bob’s will emerge, and which then will become the new model serve—the new template to copy—for all of us. When that happens, this new serve becomes the new mean of the population, with a new cloud of error around it.

If we concentrate on the population mean, it is clear that cumulative design is taking place. This is not like genetic evolution by natural selection (where replication is very high fidelity), but it is certainly the accumulation of adaptive design due to selection (and it is faster than natural selection because variants are introduced in every copying attempt).

Fig. 1. Copying with modest errors. Think of the units in the X-axis as being very small, so that the distance between the left-most bar and the right-most bar is not too great—that is, we are assuming that all serves produced are minor deviations from the target serve (which is Bob’s).

In the second case just considered replication rarely if ever happens; the norm is replication failure. It is a good summary description of the assumptions that go into many of the selectionist models that Boyd & Richerson (1985) introduced in their approach. This condition of replication failure as the norm is what Sperber claims renders cumulative adaptations from cultural transmission impossible. But we have just seen that it is certainly conceivable, and this lays bare that replication itself is a red herring. It is neither here nor there. What cumulative adaptation requires is (1) sufficient inaccuracy in the production of descendants such that superior variants can occasionally emerge; and (2) sufficient accuracy that, at the populational level (the mean), we can speak of meaningful, directional change (cf. Boyd & Richerson 2000).

B. Mutations may have consistent biases

But what about directed mutation? This idea posits an attractor, created by a psychological bias, towards which serves will tend because the copying mistakes we make are on average in the direction of the attractor. That is, the mean of our copying errors will not be zero. Contra Sperber, this is still not a problem—at least not in principle.

The attractor could be anywhere at all, but we can get our bearings by again considering the two extremes, namely, (1) when the attractor is the optimally effective serve, and (2) when it is in a direction opposite to the optimally effective serve.

(1) The mutation attractor is the optimally effective serve. This case is illustrated below in figure 2. As before, suppose that every person tries to copy Bob’s serve exactly, but fails within a cloud of error with mean zero. A few people, however, can see forward to the kinds of modifications that would make Bob’s serve even better, and attempt these. This means that the actual mean ‘error’ for the whole population will be skewed by these innovators in the direction of the optimal serve. Does this prevent cumulative adaptive design? No. On the contrary, it speeds up the process that takes the population to the optimal serve because mutations in this direction are slightly more likely. The design is cumulative because foresight does not extend to the optimal serve itself, merely to slight modifications of observable serves that take them in that direction.

Fig. 2. Adaptive mutation bias. In this case the population mean is closer to the optimum, after copying, than is Bob’s.

(2) The mutation attractor is in a direction opposite to the optimal serve. This case is illustrated below in figure 3. This could mean, for example, that there is something about the way it feels natural to move our bodies that makes us more likely to make errors in a direction away from the optimally most effective serve. But the phrase here is more likely. It doesn’t mean that copying errors in the direction of a better serve never happen. Thus, what happens is that the mean copying effort results in a serve somewhat lower in quality than Bob’s, but if the cloud of copying error occasionally produces a serve better than his, this serve will become the new target for copiers. This results in a new population mean that is again less good than the new target serve, but it is not less good than the previous mean serve in the population. Thus, the population mean will have moved closer to the optimal serve despite the fact that the mutation bias always makes it lag behind its current target.

Fig. 3. Maladaptive mutation bias. In this case the population mean is further away from the optimum, after copying, than is Bob’s serve. However, some copiers will make mistakes to the right of Bob, and since this yields a better serve, it will become the model for the next generation.

Only when the attractor is so far away that it prevents the emergence of any variants better than Bob’s serve would the emergence of cumulative design be short-circuited, as shown below in figure 4.

Copying mistakes

that result in improvements

Fig. 4. Overly strong maladaptive bias. Due to a strong mutation attractor, the population mean is so far away from Bob’s serve in a maladaptive direction that better serves practically will never appear.

The last example above shows that, when directed mutation occurs, it should be modeled together with selection. The direction of the system will then result from the algebraic sum of all the forces considered. We don’t have to decide whether either mutation or selection is the force to consider in our modeling exercises. For problems having the structure just considered, Sperber will be right that constant, directed mutation, prevents cumulative adaptation only if and when such mutation is (1) not towards the optimum and, (2) of sufficient strength. This is an empirical question, and it may be true for some domains and not for others. But we will not find the answer under the armchair.

But do we have empirical examples of cumulative cultural adaptations through selection? Yes. Other than tennis serves, we could name tennis racquets. In fact, we could name anything in the large domain called ‘technology’. Here design has obviously accumulated gradually. And even here Sperber’s dictum that replication is a limiting case rather than the norm is correct (except in the case of our very modern manufacturing techniques).

One can also point to institutions. Certainly institutions have been ‘constituting’ themselves on paper for a long time, but institutional organization pre-dates paper. Moreover, though the rules of an institution may be written, institutional behavior is always in the (sometimes very) flexible neighborhood of what is written down, rather than a rigid instantiation of it. In this sense—as living, breathing organisms—institutions are always imperfectly copied (for an example, consider that the Mexican political constitution is—on paper—almost a replica of the American, on which it was modeled). And yet institutions accrete cumulative changes. The evidence that they do so adaptively is in the incontrovertible fact that complex societies have outcompeted simple ones, and in the fact that different institutional arrangements have been the key to success in the competition between different complex societies (McNeil 1963, Landes 1998, Diamond 1997, Wright 2000). Technological and institutional change are not the only examples, merely the most obvious ones. But they occupy much of what is important in cultural evolution, so they make the case that selectionist approaches will be quite significant to explaining culture.

Given that cumulative cultural adaptations don’t require memes to replicate, this was not the litmus test for Darwinian analyses to culture. And if my critique of gene-analogy fetishism among the critics of ‘memetics’ is acceptable (for a mathematical demonstration of my core arguments, see Henrich and Boyd 2002), it simultaneously refutes the arguments of proponents such as Dawkins, Dennett, and Blackmore, who fetishize the alleged importance of ‘replication’ for opposite reasons.

IV. ‘Imitation’ is another red herring

A related point can be made about ‘imitation’ (i.e. what we do when we copy Bob’s serve). Blackmore insists on imitation as the memetic p