In recent years, Bayesian approaches to the mind/brain have become very influential. Two lines of research deserve to be highlighted: one explores how cognitive development can be accounted for by rational constructivist models (Gopnik & Wellman Reference Gopnik and Wellman2012; Griffiths et al. Reference Griffiths, Chater, Kemp, Perfors and Tenenbaum2010; Tenenbaum et al. Reference Tenenbaum, Kemp, Griffiths and Goodman2011; Xu Reference Xu, Carruthers, Laurence and Stich2007), whereas the other investigates how neural processing can be understood as a form of predictive coding (Clark Reference Clark2013b; Shipp Reference Shipp2016), and more generally, how entropy reduction is made possible through active inference (Friston Reference Friston2010). As yet, very few attempts have been made to apply these emerging theories to the study of culture. In this respect, Veissière et al.'s endeavor can only be applauded. Although I am sympathetic with the general spirit of the authors’ endeavor, I will contend that their theory is not viable because of at least two important flaws.

Drawing upon two concepts – “thinking through other minds” (TTOM) and the “free-energy principle” (FEP) – Veissière et al. intend to explain the origins of implicit cultural norms, beliefs, and habits (sect. 1.1, para. 2). As a brief reminder, TTOM has it that “information from and about other people's expectations constitutes the primary domain of statistical regularities that humans leverage to predict and organize behavior”; moreover, FEP stipulates that “living systems act to limit the repertoire of physiological (interoceptive) and perceptual (exteroceptive) states in which they can find themselves.” Can these two concepts help us understand the mechanisms of enculturation? I doubt it. Indeed, contra TTOM, it is known that some priors are shaped through strictly intrapersonal – and not interpersonal – processes; and, contra FEP, there is ample evidence of human behavior not complying with entropy minimization.

Veissière et al. mention “optical illusions” as one of the explananda of their theory (sect. 1.1, para. 2). Therefore, let me first discuss this specific example. The best theory we have of visual illusions – the natural scene statistics theory – shows that visual priors responsible for illusions are shaped by regularities in the surrounding environment (Howe & Purves Reference Howe and Purves2002; Howe et al. Reference Howe, Yang and Purves2005). This theory accords well with cross-cultural work demonstrating that variation in the perception of visual illusions directly results from the exposure to every day's environment (Miyamoto et al. Reference Miyamoto, Nisbett and Masuda2006; Segall et al. Reference Segall, Campbell and Herskovits1966). For instance, individuals growing up around complex and ambiguous scenes will be more likely to develop a “holistic perceptual style” and to be tricked by illusions requiring context-independent scrutiny. The process through which the enculturation of visual priors takes place does not involve one's expectations about other people's expectations; in other words, it is TTOM-free. Therefore, it is difficult to understand what the authors might want to mean when they claim that TTOM can shed light on the enculturation of implicit priors responsible for optical illusions.

Importantly, this objection is not restricted to the domain of vision; it applies to numerous other domains. For example, categorization and reasoning have been shown to vary across cultures because of “ecocultural factors” (e.g., being a farmer rather than a fisherman) (Uskul et al. Reference Uskul, Kitayama and Nisbett2008). Now, these factors are all about individual exposure to specific environmental patterns and have not much to do with TTOM. In sum, it seems that Veissière et al. have overlooked the wealth of evidence showing that a good deal of enculturation takes place completely outside of TTOM.

Another central claim of the article is that humans tend to minimize entropy. Interestingly, Veissière et al. point out that entropy reduction is consistent with temporary entropy increase (sect. 3.1 and 3.3). When humans happen to be seeking uncertainty, the authors note, it is only because they anticipate that a dramatic drop in entropy will take place soon after (the peekaboo game is mentioned to illustrate this point). Unfortunately, here again, apart from anecdotal evidence, no experimental data are offered by the authors to corroborate their claim. Crucially, against FEP, numerous studies have shown that in esthetics (Delplanque et al. Reference Delplanque, De Loof, Janssens and Verguts2019), music perception (Chmiel & Schubert Reference Chmiel and Schubert2017), visual perception (Chetverikov & Kristjánsson Reference Chetverikov and Kristjánsson2016), consumer behavior (Kao & Wang Reference Kao, Wang and Yamamoto2013), etc., humans have a preference for medium entropy patterns rather than low-entropy patterns. Entropy and liking follow an inverted U curve: expected (low entropy) patterns are judged to be boring, completely unexpected (high entropy) patterns are deemed too difficult/demanding, and medium entropy patterns are liked and looked for (cf. Berlyne Reference Berlyne1966).

Things happen to be even more intricate than just suggested, for if, on the one hand, plenty of studies have shown a preference of humans for medium entropy, on the other hand, an increasing number of studies demonstrate that the preferred level of entropy is highly variable across individuals (e.g., Güçlütürk et al. Reference Güçlütürk, Jacobs and van Lier2016; Güçlütürk & van Lier Reference Güçlütürk and van Lier2019). This line of research suggests that the relationship between liking and entropy may be largely shaped by cultural factors, and as a consequence, that any normative claim – for example, “humans seek to minimize entropy” – is pointless.

Optimality claims are particularly knotty (Frank Reference Frank2013); this is why, in response to critics, some Bayesians have recently proposed to distinguish between normative and descriptive Bayesian models, and have further argued that descriptive Bayesianism fares better against criticisms (Tauber et al. Reference Tauber, Navarro, Perfors and Steyvers2017). It is unfortunate that the authors do not address this important issue and fail to adumbrate an optimality-free version of their framework. Last but not least, Friston has elsewhere acknowledged that FEP is not empirically falsifiable (Friston et al. Reference Friston, Fortier and Friedman2018, p. 21); therefore, it is not clear to me whether Veissière et al. intend to make an experimentally testable claim when they state that humans tend to minimize entropy. For the same reason, it is not clear either whether FEP can be of any avail to social scientists.

In conclusion, I wish to emphasize that none of the above criticisms undermines the prospect of a Bayesian theory of culture. What I have argued, rather, is that if such a theory is to be achieved, it will build upon descriptive constructivist Bayesian models of cognition (e.g., Fortier & Kim Reference Fortier, Kim, Zedelius, Müller and Schooler2017) rather than FEP.