Experimental studies of cultural transmission
Even though transformation is most commonly associated with the cultural attraction approach, transformation is actually a common subject of study within the standard cultural evolution literature. For example, Mesoudi and Whiten (2004) examined the transformation of event knowledge (descriptions of everyday events such as going to a restaurant) as it is passed along chains of participants each of whom receives the previous participants’ recall as their input. In line with schema theories from cognitive psychology, it was found that low-level actions (e.g. ‘he opened the door, took a seat at the table) were spontaneously subsumed into medium-level (e.g. ‘he entered the restaurant) and high-level (e.g. ‘he went to a restaurant) goals, despite the latter being absent from the original description. The driver of change here is clearly individual modification, which occurred similarly and independently in all participants based on previously acquired and common knowledge structures. While not described as (narrow) cultural attraction in the paper, it is clearly an example of it.
In the same way, although they do not use the term ‘cultural attraction’, experiments and models in the Bayesian inductive reasoning or iterated learning tradition (Kirby et al. 2007; Griffiths et al. 2008) also appear to capture the kind of change envisioned by cultural attraction proponents. For example, Xu et al. (2013) experimentally simulated changes in colour labels, noted by Claidière et al. (2014) to be potentially constrained by psychophysical aspects of cognition that may form a cultural attractor. Xu et al. had participants learn to pair novel words with specific colour shades. While the first participant learned random word-colour pairings, subsequent participants were trained in the word-colour pairs produced by a previous participant, along chains of 13 cultural ‘generations’. Each participant made non-random changes such that the fictional terms gradually converged on real-life colour term clusters. Similar iterated learning experiments have shown that properties of languages, such as compositionality, spontaneously emerge as participants individually modify artificial languages to make them more learnable (Kirby et al. 2008). The source of change in all these cases lies in individual participants’ cognition and perception, which act in similar ways across people to drive cultural representations towards similar end-points, in line with the notion of cultural attraction.
Other studies have combined guided variation and selection-like biased transmission. Bettinger and Eerkens (1999) found that different patterns of prehistoric arrowhead variation in North America showed signatures of different mechanisms of cultural transmission. Arrowheads from Nevada showed little variation, coming in a small number of uniform types. Contemporary arrowheads from California showed extensive variation, with no uniform types. Bettinger and Eerkens (1999) argued that the former pattern was generated via model-based biases, with prehistoric hunters preferentially copying the arrowhead designs of successful or prestigious hunters, thus creating a small number of popular types. The more diverse Californian arrowheads, on the other hand, were influenced by guided variation, as each arrowhead maker modified their design according to trial-and-error. Mesoudi and O’Brien (2008) subsequently experimentally simulated these hypothesised transmission processes, confirming that model-based bias and guided variation can, under certain circumstances, generate the observed patterns of low and high variation respectively. This general hypothesis is noteworthy because the individual modification component is not a result of common content-based cognitive biases, but instead due to contentless trial-and-error (associative) learning.
In the only experimental study that we know of that has explicitly compared cultural attraction and selection, Eriksson and Coultas (2014) examined the transmission of stories that invoke to varying degrees emotional reactions of disgust. In one experiment they passed stories along chains of participants in the standard manner, finding that elements rated highly disgusting were preserved over elements rated low in disgust. This can be seen as a form of cultural attraction, with the stories mutating at each step in a non-random direction to contain relatively more disgusting content. In a further experiment, Eriksson and Coultas (2014) allowed participants to choose whether to read, and then whether to pass on, a story to a subsequent participant, without altering the story. Hence this resembles selection (specifically, content/direct bias), because the stories change in frequency without being altered. Both methods revealed a bias towards disgusting content, indicating that disgust bias operates both through the non-random transformation of content as it is remembered and reconstructed, and also the non-random selection of content as it is chosen and replicated. Although Erikkson and Coultas did not discuss this, it appears that selection had a larger effect, in that low-disgust material was entirely absent at the end of the chains in the experiment in which only selection was possible, whereas in the experiment in which only transformation was possible, low-disgust material was still present at the end. In explaining the real-life preponderance of disgusting urban legends (see Heath et al. 2001), both cultural attraction and cultural selection can potentially be seen to be working together.
Here are some take-home messages from these studies. First, cultural attraction, transformation and cognition are not ignored in standard cultural evolution research. Many studies, in particular transmission chain studies, have explicitly examined transformative processes. If anything, it is rarer to find transmission chain studies that examine cultural selection. Second, few studies have explicitly studied both selection and attraction. Eriksson and Coultas (2014) is a rare exception. Third, studies such as Mesoudi and O’Brien (2008) highlight that cultural attraction does not have to be due to cognitive universals. The individual modification that occurs in cultural attraction can occur via individual trial-and-error. If a task has multiple solutions, then perhaps trial-and-error will lead different people to different solutions (as it did in Mesoudi and O’Brien’s study), such that cultural attraction can generate and maintain cultural diversity.
Preservative versus reconstructive processes depend on the granularity of the analysis
In some cultural evolution studies, the unit of analysis is the cultural trait, that is, what is transmitted in the cultural transmission process. Examples of cultural traits include names, fairy tales, ways to tie a knot, recipes for lasagne, hammers, and the like. In others, the unit of analysis is the individual person (see also El Mouden et al. 2014). If each individual has exactly one cultural variant of a particular type, then these units will coincide. However, where individuals can possess multiple cultural traits, then classifying cultural change as attraction or selection-like becomes complicated. Image that person A has ideas X, Y, and Z, and person B learns from A only ideas X and Y, with no modification of those traits. From the trait-as-unit-of-analysis perspective, transmission is preservative: traits X and Y are being selected and transmitted with high fidelity, while trait Z has been selected against. From the individual-as-unit-of-analysis perspective, however, transmission may be considered reconstructive, as person B has a different set of traits (XY) compared to person A (XYZ), from whom she copied.
Take, for example, a transmission chain experiment by Mesoudi et al. (2006), in which multiple stories varying in their social complexity were passed along chains of participants. Over successive transmission episodes, the social stories remained largely intact, while the non-social stories virtually disappeared. If one takes the individual as the unit of analysis, then this appears to be a case of cultural attraction. The first people in the chains had a mixture of social and non-social stories, the final people had mostly social, such that there is a non-random transformation due to (according to Mesoudi et al.) biologically evolved and universal aspects of cognition (humans’ ‘social brains’). If instead one considers each separate story as a ‘trait’, then the process seems more selection-like. The social stories were more likely to be preserved, and the non-social stories less likely to be preserved, with no modification to the traits (there was little distortion or confabulation in this particular study). This change in trait frequency therefore resembles selection. Note that there is no explicit, conscious ‘selection’ of stories by the participants here, just unconscious selection as a result of (probably implicit) memory biases; the population level consequences of both explicit/intended and implicit/unintended selection will be the same, however. In sum, there doesn’t seem to be a ‘correct’ answer to whether people or traits are the unit of analysis, but which decision we take determines whether the process is transformative (attraction-like) or preservative (selection-like).
The issue regarding preservation and transformation in transmission, however, is generally considered assuming the trait-as-unit-of-analysis perspective. Consider again Cinderella. We used it above as an evident case of reconstructive cultural transmission since, each time one retells the story, it will be extremely unlikely that she will repeat exactly the version heard. However, what are we considering here as the cultural trait? A coarse-grained description of the cultural trait is “a story involving a young lady, first oppressed by her stepmother and stepsisters, and then succeeding in marrying a prince”. Because this basic plot structure is likely to be maintained through successive iterations, the transmission is, at this level, preservative. At an intermediate level we can consider, for example, Cinderella as a combination of sentences. In this case, assuming that one repeats all the sentences, one might change some words, saying: “Once upon a time there lived a sad young girl” instead of “Once upon a time there lived an unhappy young girl”. This would count as reconstructive. Finally, a fine-grained description could focus on the single words of the story. Imagine one summarises Cinderella in few sentences, using words picked from the perhaps longer version she heard. One could interpret this as a preservative process, in which some cultural traits (the words used) have been selected and reproduced without mutation.
Moreover, cultural selection and cultural attraction are likely, in the majority of cases, to act together within the same traits, at different levels of generality. As we mentioned above, supernatural concepts may be favoured because they are minimally counter-intuitive entities. As an optimal combination of intuitive and counter-intuitive features, a generic undead being (like a ghost, or a zombie, or a vampire) is an effective cultural trait. However, an explanation of the cultural success of a specific undead entity, say, Dracula, needs to include selective processes. The spreading of Dracula is most likely due both to attraction-related factors, that explain why, in general, undead beings are favoured in respect to other entities, and to selective factors, that explain why, among all other undead beings, the Transylvanian vampire enjoys such popularity.
One could hope that, when we find the correct unit of analysis for cultural evolution, we would be able to settle the debate. Unfortunately, this might be unlikely, as there is continuing disagreement over how to define a cultural trait. Sperber and Claidière (2008) criticize Richerson and Boyd (2005) for seeming to oscillate between an “internalist” view of cultural traits as “(mostly) information in brains” (ibidem, p. 61) and an “externalist” one where “some cultural information is stored in artifacts” (ibidem, p. 61). However, the cultural attraction approach adopts a similar strategy assuming that both mental representations and behaviours/artifacts should be considered cultural traits (Sperber and Claidière 2008). Claidière et al. (2014), for example, explicitly discuss how mental representations and public narrations of a folktale should be both treated as cultural traits with “equally potent causal roles” (ibidem).
One proposed solution to this puzzle is to consider the information, wherever stored, as the equivalent of the biological genotype, and the expression of the information in behaviours or artifacts as the equivalent of the biological phenotype (Dawkins 1976). The problem here is that it assumes that, when copying, we have access to a “cultural core” (Sperber and Claidière 2008), which represents the information/genotype, which we then use to build variable phenotypic expressions. This might be loosely the case: the classic example is the transmission of a recipe to cook, say, lasagne, where the recipe represents the transmitted, stable, genotype, and what you serve to your guests at dinner is the variable phenotype. However, in many cases, we do not have access to a “recipe”, but we extract the information from the result/phenotype (such as when we try to reproduce lasagne after tasting it at a friend’s home). Richerson and Boyd (2005) make a similar point when noting how the mental representations of different individuals who have tied the same bowline knot might in principle be very different. What is the genotype here? The individual, variable, mental representations of the bowline knot cannot be the genotype, as they are not, in general, transmitted, because they are different. For the same reason, the information stored in the artifact itself does not transfer directly in the (variable) mental representations.
Furthermore, even solving the internalist/externalist debate would not settle the reconstructive/preservative question. Imagine that everybody agreed on an internalist view, so that the real cultural traits in the transmission of a folktale are the mental representations, which we could access with some advanced neuroimaging technique. As we suggested in the Cinderella example, would they be the mental representations of “a story involving a young lady, first oppressed by her stepmother, etc.” or more detailed mental representations of the plot, or something else?
While this may appear pessimistic, we believe that pluralism in the conceptual definitions of the unit of analysis in cultural evolution is not a problem (see also Lyman and O’Brien 2003; O’Brien et al. 2010). Biologists, too, work simultaneously with multiple concepts of the ‘gene’, varying with context and use (Stotz and Griffiths 2004). Depending on various domains, and on the questions one is interested in, an opportunistic strategy can be the best choice. Moreover, moving from coarse to fine grained units can indeed clarify how the interplay between attraction and selection can be important for the success of specific cultural traditions, as the Dracula example illustrates.
Preservative versus reconstructive processes depend on the empirical domain
Besides the decision of what to consider a cultural trait, the fidelity of cultural transmission likely also varies in different empirical domains. We initially compared two cases. In the oral transmission of stories, we can infer from high variability of the successive reproduction of the “same” story that reconstructive processes strongly influence cultural evolution. In first names diffusion, instead, the innovation rate is extremely low, and cultural transmission is highly preservative, such that selective processes are more important than attraction-based processes. Many other examples are possible across the domains of technology, language, art and social customs.
The general scepticism of proponents of cultural attraction towards the idea that high fidelity imitation is the unique, or even the main, support for cultural evolution is a useful counterbalance to a naïve view of humans as perfect and indiscriminate copy-machines, and that this is enough to explain cultural stability. Not only are copying mechanisms often characterised by low fidelity (as in the Cinderella example), but also long-term, stable, traditions are not necessarily supported by high fidelity copying (as in the religion example, where supernatural concepts may be reconstructed each time). However, it does not follow from here that copying mechanisms are always scarcely faithful, or that stable traditions are never supported by high fidelity copy.
Many technologies that we use are, for example, causally opaque (Csibra and Gergely 2011), meaning that we do not know or understand the mechanism by which they produce the result we use them for. Experimental studies have demonstrated how common high fidelity copying is for technology-related actions. Flynn and Smith (2012) had adult participants observe a model perform some operations with a box (using a tool to drag some bolts, tapping with the tool, lifting a door, inserting a tool into a hole) in order to retrieve a reward from inside. Only the last two of these actions actually retrieved the reward, the others had no causal effect in relation to the goal. One group of participants observed the model interacting with a transparent box and were thus able to see which actions were unnecessary. For another group the box was opaque, obscuring which actions were causally relevant. Flynn and Smith (2012) found that adults, like children (Lyons et al. 2011), showed a high likelihood of copying all actions—both relevant and irrelevant—under both conditions, even the transparent condition where the irrelevant actions are revealed to be irrelevant. This phenomenon, dubbed ‘over-imitation’, indicates that high fidelity copying is often the default approach to solving unfamiliar problems, even out-weighing causal reasoning. Interestingly, however, when the model was another participant (rather than the experimenter) Flynn and Smith (2012) found that participants did not reproduce the unnecessary actions when the box was transparent (they still did when the box was opaque). In other words, when potential sources of prestige are removed (thus removing the possibility of prestige biased cultural selection), a causally transparent technology elicited reconstructive transmission, while a causally opaque technology elicited preservative transmission. More generally, we suspect that the more a technology is opaque, the more cultural transmission will be preservative; the more a technology is transparent and model-based biases are absent, the more cultural transmission will be reconstructive.
Other domains that might be characterised by generally preservative transmission are domains in which a final result is reached through a sequence of actions, and sequences of actions that are even slightly different to the correct one produce an unusable result (Acerbi et al. 2011). Tying a Windsor knot is a serious affair that involves a sequence of precise actions. Performing correctly, say, nine of the ten actions required does not produce 90 % of a Windsor knot, but will likely produce a shapeless configuration of fabric. The task of tying a Windsor knot can be visualised as a search space with a single slender peak (the correct knot) surrounded by a vast flat territory (all the action combinations that produce unusable results). For these tasks, individual learning—or reconstruction—is in general an unsuccessful strategy, because the final result does not provide any feedback about “how close” one is to the correct solution, nor has genetic evolution provided us with precise intuitions about knot-tying. Individual learners need to explore each time the full space of possible actions. The great majority of modern technological tasks probably fits this description. While constraints that can help individual search and reconstruction do exist—an airplane has to fly and a kayak has to float—their guidance is so loose that only preservative cultural transmission can sustain those traditions (Acerbi et al. 2012). Notice that examples of opaque or slender-peaked tasks are not necessarily restricted to the technological domain. Other activities that require performing arbitrary but well-defined sequences of actions, like dancing or rituals, could in the same way require preservative cultural transmission to persist (Tennie et al. 2009).
Saying that high fidelity copying is the best strategy in certain situations, or that some traditions need to be supported by high fidelity copying, does not guarantee, of course, that this is what happens in reality. We may indeed use suboptimal strategies, and persistence of traditions can be explained by something else. However, we have good reasons to believe that, for some domains, this is indeed the case. Csibra and Gergely (2011) suggest that a suite of species-specific cognitive adaptations for cultural learning, which they label ‘natural pedagogy’, may be responsible for the capacity of preservative cultural transmission of opaque technologies. Natural pedagogy indicates that social learning is accompanied by ostensive communication, that is, a form of deliberate communication (“Look at what I am doing with this stick!”) that guides the learner through the critical aspects of the process. Similarly, Herrmann et al. (2013) showed that verbally framing a demonstration stressing the conventionality of the actions involved (as opposed to their instrumentality) is sufficient to increase imitative fidelity in preschool children. Others (Tennie et al. 2009) have emphasised how high fidelity in human cultural transmission can be achieved through a combination of process-oriented imitative social learning (humans tend to pay attention not only to the final result of a demonstration—a Windsor knot—but also to the actions performed to reach the result) and a form of cooperation that favours active teaching and social motivations to copy. The afore-mentioned over-imitation studies, where people copy both relevant and irrelevant actions demonstrated by others, provides evidence for this.
Another important factor that increases the fidelity of cultural transmission is the use of epistemic tools (Sterelny 2006). Epistemic tools are modifications of the environment—in a broad sense—that improve the cognitive capacities of individuals. Tasks that are hard for children to learn, such as tying their shoes, can be encoded in vivid images and rhymes such as “Bunny Ears”. Tehrani and Collard (2009) argue that they are able to trace robust phylogenies (a sign of preservative transmission) of Iranian tribal textiles because craft learning is scaffolded in such a way that different designs are embodied as a set of motor routines that are difficult to rewire. Modern culinary recipes are another good example of epistemic tools. They convey detailed information through numbered lists of ingredients, with universal measures, explicit sequences of actions, and possibly images of the various phases of the preparation. As with all technical idioms, cookery language has developed a series of specific terms (to sauté, to simmer, to reduce, etc.) that decrease ambiguity and, again, favour preservative transmission.
Of course, language itself is a preeminent epistemic tool, and written language has been explicitly considered as a technology that favours preservativity of cultural transmission, compared to oral communication (Ong 1982; Rubin 1995). One innovative line of studies examined the hand-copying by scribes of stories before the invention of the printing press, stories such as The Canterbury Tales (Barbrook et al. 1998; Howe et al. 2001). Phylogenetic analyses accurately reconstructed the evolutionary relationships between the different manuscripts due to the high fidelity copying. There were also copying errors intriguingly similar to those found in genetic inheritance, such as the insertion or deletion of words or letters, or the random swapping (or ‘crossing over’) of sentences from one manuscript to another. In these cases, where the express goal is to replicate a text, there was seemingly very little directional transformation.
Today, we can observe a new shift that involves digitally mediated cultural interactions. The transmission of Internet content (think of social media “sharing”) is a form of highly preservative cultural transmission, where the information is practically replicated with no mutation. Intriguingly, there are several examples of short texts that have become “viral” in social networks such as Facebook or Twitter which users are explicitly asked to not automatically share or re-tweet, but to copy and paste manually (Adamic et al. 2014). This re-introduces the possibility of transmission errors or conscious modifications, or, in other words, makes transmission more reconstructive in a preservative media like the Internet. Adamic et al. (2014) found indeed a decrease in transmission fidelity (a mutation rate of 11 %), with some non-random modifications. For example, the phrase “No one should die because they can’t afford health insurance…” was transformed by conservatives into “no one should die because the government is involved with health care…”, reminiscent of Bartlett’s (1932) early studies where information is distorted to fit pre-existing opinions. Overall, it is a fair question to ask whether the ubiquitous presence of digital communication is making cultural transmission more preservative than reconstructive and what the consequences are of this transformation.
All these examples show that it is important to not automatically assume that human culture is sustained by perfect transmission, but how, in some domains, the fidelity of cultural transmission is higher than in others. Rather than deciding whether attraction or selection is in general more important, it is more interesting to ask the extent to which transmission is preservative or reconstructive in different domains, and how attraction and selection consequently interact to shape cultural variation.