The provocative thesis of Cognitive Gadgets (Heyes Reference Heyes2018) is that human abilities such as imitation, mindreading, and language – the traits that allow our species’ extensive cultural transmission – are not adaptations produced by biological evolution or, as repeated throughout the book, “in our genes.” Instead, these abilities are themselves “gadgets” that have been created and refined by cultural group selection. Although they give the illusion of innateness, they are taught to children through social practices, and learned with the support of enhanced domain-general mechanisms such as attention, social motivation, working memory, and – most importantly – associative learning. Except for potentiating these general-purpose cognitive tools, genetic evolution has had virtually no role in shaping the distinctive traits that define human nature.

Why evolution should have followed this route in our species is a mystery, and Heyes does not offer any rationale or theoretical model to make sense of it. In fact, she stresses that genetic evolution could have played a role – the evidence from cognitive science just happens to say otherwise. The first question, then, is whether the book makes a compelling empirical case for its almost-blank-slate argument. Cognitive Gadgets presents a wealth of interesting findings and useful criticism of previous research; but as a refutation of innatism I found it surprisingly weak. Consider Heyes’ treatment of genetic assimilation. In a nutshell, genetic assimilation occurs when traits that initially develop through learning (or other types of plasticity) get increasingly under genetic control, as selection favors variants that make the learning process faster and more reliable. In principle, assimilation can proceed so far that the trait develops entirely under genetic guidance, with no environmental input.

Heyes claims that she found no evidence of genetic assimilation for abilities like imitation and mindreading. Granting the premise for now, Heyes assumes that the heritability of a trait estimated from twin studies is an indicator of whether the trait develops with minimal environmental input (“poverty of the stimulus,” high heritability) or with considerable input from the social environment (“wealth of the stimulus,” low heritability). For imitation, the book cites one study of 2-year-olds by McEwen et al. (Reference McEwen, Happé, Bolton, Rijsdijk, Ronald, Dworzynski and Plomin2007) as showing that “identical twins are no more alike in their imitative ability than fraternal twins” (p. 208). But this is not what the study found. The correlation was significantly higher in identical twins, and the authors estimated the heritability of imitation at 30%. This figure is well within expectations: The heritability of cognitive traits is small in infancy, but increases to about 30%–40% in childhood and reaches 50%–60% by late adolescence (Briley & Tucker-Drob Reference Briley and Tucker-Drob2017). Heyes fails to cite another study of imitation in 2-year-olds (Fenstermacher & Saudino Reference Fenstermacher and Saudino2007), which also found a higher correlation in identical twins and estimated heritability at 45%. For mindreading, Heyes cites one study by Hughes et al. (Reference Hughes, Jaffee, Happé, Taylor, Caspi and Moffitt2005), which found the same correlation between identical and fraternal twins, indicating negligible genetic influences on individual differences in children's theory of mind. She omits to mention that, although the authors found no specific genetic contributions to theory of mind, there was a significant influence of nonspecific genetic factors shared with verbal ability (accounting for about 15% of variance). Other twin studies of theory of mind in children and adults have found heritabilities in the 15%–35% range (McEwen et al. Reference McEwen, Happé, Bolton, Rijsdijk, Ronald, Dworzynski and Plomin2007; Ronald et al. Reference Ronald, Viding, Happé and Plomin2006; Warrier et al. Reference Warrier, Grasby, Uzefovsky, Toro, Smith, Chakrabarti, Khadake, Mawbey-Adamson, Litterman, Hottenga, Lubke, Boomsma, Martin, Hatemi, Medland, Hinds, Bourgeron and Baron-Cohen2018). Thus, contrary to Heyes’ claim, both imitation and mindreading skills show a nontrivial proportion of genetic variance. Moreover, the apparent heritability of mindreading is most likely deflated by the rather noisy measures employed in these studies.

A deeper question is whether twin correlations and heritabilities are germane to the book's argument. In contrast with Heyes’ assumptions, the proportion of genetic versus environmental variance says very little about the nature of environmental inputs and the trait's history of genetic assimilation. Consider a genetically assimilated trait that has become fixed in a population, and shows little or no genetic variation among individuals. By necessity, most of the variance of such a trait would be environmental. Or consider a hypothetical developmental process in which an environmental variable triggers the expression of alternative, genetically specified behaviors that are the same for all the individuals in a species. The resulting trait would show low heritability and high environmental variance; but the role of the “stimulus” would be limited to selecting from a menu of pre-specified alternatives. To further complicate things, nonshared environmental variance in a trait may reflect random events and insults (e.g., infections) rather than learning or organized plasticity; and genetic variance may capture the effects of deleterious mutations besides those of functional alleles. In general, the factors that drive the development of a trait may not be the same factors that produce individual differences in that trait. Moreover, a particular skill can be both evolutionarily novel and socially learned, but depend for its acquisition on traits that show substantial genetic variation. To illustrate: playing chess is a cultural “gadget” if there ever was one, and yet interest and aptitude for chess are about 40–50% heritable (de Moor et al. Reference de Moor, Roeling, Boomsma, Vartanian, Bristol and Kaufman2013; Olson et al. Reference Olson, Vernon, Harris and Jang2001; Vinkhuyzen et al. Reference Vinkhuyzen, Van der Sluis, Posthuma and Boomsma2009). By Heyes’ criteria, one should conclude that playing chess is more likely to be a “cognitive instinct” than imitation or mindreading. In sum, the book's argument for rejecting genetic assimilation is conceptually flawed and supported with cherry-picked data.

To remain on the topic of mindreading, Heyes cites interesting cross-cultural evidence that the stages of theory-of-mind acquisition differ between individualistic countries like the United States and Australia and collectivistic countries like China and Iran. But these findings are damning only if one holds an inflexible model in which the various components of mindreading (Schaafsma et al. Reference Schaafsma, Pfaff, Spunt and Adolphs2015) can interact only in one pre-specified way, with no meaningful input from the social environment. Of note, the observed sequence changes typically involve two particular tasks out of five (“diverse beliefs” and “knowledge access”; see Duh et al. Reference Duh, Paik, Miller, Gluck, Li and Himelfarb2016; Kuntoro et al. Reference Kuntoro, Peterson and Slaughter2017; Shahaeian et al. Reference Shahaeian, Peterson, Slaughter and Wellman2011; Reference Shahaeian, Nielsen, Peterson and Slaughter2014; for a puzzling exception, see Dixson et al. Reference Dixson, Komugabe-Dixson, Dixson and Low2018). The overall picture, then, is one of patterned variation on a background of stability. Heyes also cites evidence that theory of mind development is markedly delayed in Samoan children (Mayer & Träuble Reference Mayer and Träuble2013). However, this literature contains several inconsistent findings that cannot be explained by cultural differences (see Liu et al. Reference Liu, Wellman, Tardif and Sabbagh2008; Mayer & Träuble Reference Mayer and Träuble2015). Some apparent delays may reflect culture-specific issues with task demands, as Mayer and Träuble (Reference Mayer and Träuble2015) noted in their follow-up to the original Samoan study. At the same time, theory-of-mind skills are not independent from other cognitive traits, and are significantly associated with IQ (e.g., Baker et al. Reference Baker, Peterson, Pulos and Kirkland2014; Rajkumar et al. Reference Rajkumar, Yovan, Raveendran and Russell2008). It may be impossible to fully make sense of the cross-cultural data on developmental trajectories without addressing the thorny issue of national differences in cognitive ability (e.g., Rindermann Reference Rindermann2018).

These examples serve to illustrate a double standard that is applied throughout the book: whenever the data do not support a rigidly “preformist” view of development, they are implicitly or explicitly counted as positive evidence for an associative account. But in several of the examples discussed in Cognitive Gadgets, associative learning is little more than a hypothetical mechanism (or a plausible contributing factor), and it is unclear if the models proposed by Heyes are able to explain the totality of the evidence. Moreover, the apparent simplicity of associative accounts often hides a lot of complexity (and inefficiency), which is revealed only by careful unpacking (e.g., Dickinson Reference Dickinson2012; Hanus Reference Hanus2016). For all these reasons, Heyes’ rejection of innateness in favor of almost-blank-slate associationism seems highly premature.

I will not discuss the book's case for cultural group selection in any detail, except to note that the argument is fully – and admittedly – speculative. To be clear, I see nothing wrong with bold speculation; but there is some irony in the sudden shift away from the hard-nosed empiricism of the rest of the book, precisely at the point where Heyes needs to explain how all the distinctive content of human nature can be outsourced to culture-mediated learning. For example, it is unclear if the selection process envisioned in the book could provide enough robustness and reliability to enable adaptive evolution; if it could work on a realistic timescale, given the long “life cycle” of groups compared with that of individuals; how it would respond to conflicts of interests between different social actors, and between group and individual fitness; and how it would prevent genetic adaptation from catching up with cultural transmission.

Even though my review of Cognitive Gadgets is critical, I strongly recommend the book to other evolutionary-developmental psychologists. It will stimulate them, challenge them to think more deeply about their assumptions, and prompt the field to open the developmental “black box” and become more explicit about computational processes. I see a clear parallel with much recent work in artificial intelligence (including neural networks), which shares the book's empiricist attitude and faith in the power of domain-general learning (Marcus Reference Marcus2018; see also Lake et al. Reference Lake, Ullman, Tenenbaum and Gershman2017). This new wave of research is a fantastic opportunity for evolutionary-developmental psychology. Understanding how learning is instantiated in the mind/brain, guided by evolved developmental programs, and integrated with innate information is a daunting task, which has been made even harder by a scarcity of explicit models (Frankenhuis & Tiokhin Reference Frankenhuis and Tiokhin2018). Computational tools like reinforcement learning can help understand what (and how much) pre-existing information is needed to perform efficiently and reliably in the real world (Frankenhuis et al. Reference Frankenhuis, Panchanathan and Barto2018), and how evolved developmental programs may respond to novelties in the environment, from optical mirrors to online interactions.

These questions can be approached in a spirit of synergy and integration (e.g., Frankenhuis et al. Reference Frankenhuis, Panchanathan and Barto2018; Lake et al. Reference Lake, Ullman, Tenenbaum and Gershman2017; Versace et al. Reference Versace, Martinho-Truswell, Kacelnik and Vallortigara2018), or – less productively – as a zero-sum competition between genetic inheritance and learning. Back to Cognitive Gadgets, it is unfortunate that Heyes sets up her main argument as a dichotomy between two extremes. Psychological mechanisms are either genetically encoded, domain-specific “instincts” that develop with minimal environmental input; or culturally transmitted “gadgets” that are learned through domain-general processes, with minimal or no contribution from genetic factors. The only middle-ground option entertained in the book – and quickly dismissed – is genetic assimilation (see above). This black-and-white contrast leaves out a world of more plausible possibilities. For example, psychological mechanisms may reliably develop a basic level of functionality with minimal input, but depend on learning (often directed and canalized) in order to reach full competence. Although basic preferences for sweet versus bitter flavors are present at birth, food preferences are expanded and fine-tuned through years of intensive but nonrandom learning, which yields cultural similarities as well as differences (Rozin Reference Rozin1990a; Reference Rozin1990b). Furthermore, even established preferences for or against certain foods can be adaptively overturned by conditions such as pregnancy and nutrient deficiency (Berthoud Reference Berthoud2011; Flaxman & Sherman Reference Flaxman and Sherman2000; Rozin Reference Rozin1990a).

By tuning their operating parameters, general processes such as associative sensory-motor learning can be canalized to reliably yield specific, adaptive outcomes.

My colleagues and I have proposed such a canalization hypothesis for the development of mirror neurons (Del Giudice et al. Reference Del Giudice, Manera and Keysers2009). Also, distinct mechanisms specialized for different tasks may reuse some basic information-processing algorithms – for example, reinforcement learning – while adapting them to the particular nature of each task. Modularity, functional specialization, and the difficulty of distinguishing between domain-general and domain-specific processes have been addressed in considerable depth in the work by Clark Barrett et al. (e.g., Barrett Reference Barrett2012; Reference Barrett2015; Reference Barrett, Tibayrenc and Ayala2017; Barrett et al. Reference Barrett, Peterson and Frankenhuis2016), which reconciles the notion of specialized adaptations with a sophisticated view of learning and plasticity. A powerful idea stemming from this approach is that cognitive mechanisms may develop hierarchically, through “module spawning” and progressive specialization induced by different categories of inputs (Barrett Reference Barrett2012; Reference Barrett2015). Heyes never considers these possibilities, which have been discussed for years in mainstream evolutionary psychology (e.g., Buss Reference Buss2015). It remains to be seen whether Cognitive Gadgets will herald a genuine paradigm change, or succeed mainly as a timely provocation.