R1. The domain of culture

Several commentators noted the need to clarify the notion of culture that underwrites the TTOM model. Whiten's insightful commentary suggests that accounts of culture should explain the cultural behaviors of non-human animals, which in recent decades have been documented extensively. In particular, Whiten notes features of animal culture that a unified theory of culture will have to account for, including: phylogenetic reach, that is, the ubiquity of transgenerational transmission of learning behaviors in many species; intraspecies pervasiveness, that is, the inclusion of many forms of behavior within culturally learned repertoires, which may exhibit diversity within a specific or community; stability and fidelity over time; and finally, adaptive preferences for certain modes or types of social learning.

Whiten suggests that, from the perspective of animal cultures, some of the features of culture we discuss are optional, for example, inter-group differentiation. However, it is precisely cultural diversity, borrowing, exchange, hybridization, and competition (i.e., differences that arise primarily through cultural learning) that allow the processes of cultural change and elaboration distinct from the rudimentary replication of limited behavioral repertoires. The 4300-year-long “fidelity” in nut-cracking technologies mentioned by Whiten, along with the minor (and similarly stable) regional variations in ape-foraging strategies, reflects local affordances rediscovered in each generation, apparently without intergenerational cultural transmission (Moore Reference Moore2013). On this view, such “fidelity” may actually reflect a rigid dependence on what we have called natural affordances (Ramstead et al. Reference Ramstead, Veissière and Kirmayer2016), without the cumulative intergenerational cultural elaboration that Tomasello has called the “ratchet effect,” which appears to be unique to humans (Tennie et al. Reference Tennie, Call and Tomasello2009).

The potential phylogenetic reach of basic forms of culture raised by Whiten may yield important insights into forms of learning we share with other animals. Our account, however, focuses on the phylogeny of the enhanced theory of mind modalities and cognition–culture iterative loops that appear unique to humans. Central to our account is the cognitive package that we call Thinking Through Other Minds (TTOM). Under the FEP, we construe this as package as comprising a set of abilities and a domain of statistical regularities that are coupled: abilities enabling us to tune to the minds of others and to navigate environmental uncertainty. The TTOM model explains this coupling by appealing to evolved and learned priors about conspecifics and their mental states.

Among hominids, current models for the evolution of TTOM emphasize cooperative breeding; a strategy that likely evolved in the Homo Erectus lineage circa 2 million years ago (MYA) (Hrdy Reference Hrdy2011) that selects for individuals who are skilled at understanding others’ needs, giving care, and eliciting care. Across species, the evolution of cooperative breeding likely follows different timescales and different pathways to similar traits. New World monkeys such as marmosets, for example, who share a last common ancestor (LCA) with humans 35 MYA, are cooperative breeders with better mind-reading and prosocial abilities than the non-cooperative breeding great apes (LCA with humans 5 MYA), and have as such been recognized as better models than chimpanzees for understanding human evolution (Miller et al. Reference Miller, Freiwald, Leopold, Mitchell, Silva and Wang2016). Humans, in turn, have refined skills for detecting the epistemic cues and their precision allowing for more complex forms of cultural transmission.

We use the computational construct of precision of epistemic cues to account for the relative stability, fidelity, and scaffolded elaboration of cultural forms of life over time. To account for preference guiding behavior based on the formal construct of culture acquisition proposed by TTOM, one might appeal to gene-culture coevolutionary explanations. This is what we suggested with the example of prestige biases and regimes of attention in humans. Prestige biases are an example of external component of the regime of attention – passed on as high-precision epistemic cues (scaffolded on, then divorced from physical dominance hierarchies) – whose effect is enabled by genetically inherited predispositions to social learning functioning as an internal component of regimes of attention (see Figure 4 of the target article).

In contrast to Whiten's appeal to animal culture, Vickhoff starts from the example of the arts, forms of human culture that involve elaborate systems of shared customs, codes, and scripts. Vickhoff considers music as an instance of culture, because it is governed by a system of shared expectations, and conforms to a “style” – a type of convention with esthetic value. To this, we would add a third salient characteristic of music, the arts and, indeed, of culture more generally: improvisation and creative invention (Torrance & Schumann Reference Torrance and Schumann2019). All of these aspects are related to TTOM. Vickhoff suggests that, as an enculturated agent, the listener can access the composer's mind as expressed through culturally scripted features of music. Hence, in Vickhoff's words, we not only access culture through other minds but also engage with other minds through culture. Although the TTOM model accounts for the idea of “set of expectations,” it may be less intuitive to apply the framework to thinking about cultural genres or styles. This is so because such uses speak to notions of esthetic value and of creativity and novelty-seeking, which do not have obvious interpretations under TTOM and the FEP.

Creativity is a kind of exploration of new possibilities for perception and action. Similar to other exploratory activities, it is driven by the human propensity for novelty-seeking or curiosity. Several authors have provided accounts of culture and curiosity in terms of the epistemic value of exploring a niche or a larger adaptive landscape to identify current or future possibilities (Friston et al. Reference Friston, FitzGerald, Rigoli, Schwartenbeck and Pezzulo2017a; Reference Friston, Lin, Frith, Pezzulo, Hobson and Ondobaka2017b; Moulin & Souchay Reference Moulin and Souchay2015; Schmidhuber Reference Schmidhuber2006; Schwartenbeck et al. Reference Schwartenbeck, Fitzgerald, Dolan and Friston2013). On this view, the FEP applies to a second-order process of optimization on longer time scales and, in some instances, across many alternate niches (Bengio Reference Bengio, Kowaliw, Bredeche and Doursat2014). A similar argument can be used for the value of creativity to generate an expanded adaptive repertoire of actions and responses. Improvisation, invention, and innovation are basic to human cultures, not only in the domains of esthetic experimentation that have come to be designated as “the arts” in contemporary societies, but equally in the most quotidian activities.

The relationship between culture and creativity in the framework of TTOM needs further development, to be sure, and Vickhoff gives some clues about how to address this. He mentions recent work on synchrony in human action, which is beginning to reveal the mechanisms of micro-coordination of action in real time and its consequences for feelings of emotional attunement and social affiliation (Kiverstein et al. Reference Kiverstein, Miller and Rietveld2019; Parkinson Reference Parkinson2019; Tschacher & Haken Reference Tschacher and Haken2007; Van de Cruys & Wagemans Reference Van de Cruys and Wagemans2011; Van de Cruys et al. Reference Van de Cruys, Chamberlain and Wagemans2017). Vickhoff notes, as well, that the reduction of uncertainty can be pleasurable (Vuust et al. Reference Vuust, Dietz, Witek and Kringelbach2018). The esthetic pleasure of music may arise in part from a balance between the setting up expectations (through rhythm and repetition of melodic and harmonic structures) and violating expectations with novelty, and then re-establishing order by further repetition or thematic resolution (Huron Reference Huron2006; Vuust et al. Reference Vuust, Dietz, Witek and Kringelbach2018). This play with the tensions of confirmation of expectation and surprise then constitutes a microcosm of the larger adaptive tasks of dealing with an often-unpredictable world (Friston & Friston Reference Friston, Friston and Bader2013).

A similar challenge to the view that all human action interactions are motivated by the minimization of uncertainty is spelled out in Mirski, Bickhard, Eck, & Gut's (Mirski et al.) commentary. For these authors, a free-energy account of culture fails to explain the emergence of new or competing normativities, and the varied ways in which many human actions appear to possess no fitness-enhancing or uncertainty-reducing function. They mention “people deciding to die or suffer for some highly abstract cause” as one such candidate exception to our model. First, we should note that the pursuit of “suffering” is typically patterned and prescribed socially, and seen in many cultural domains from rites of initiation, to religion and athletics (Atkinson Reference Atkinson2008; Coakley & Shelemay Reference Coakley and Shelemay2007; Gaines & Farmer Reference Gaines and Farmer1986). We see this as paradigmatic of, rather than an exception to the search for high-precision cultural modeling under the FEP. The most widely accepted account of “righteous violence” of dying for an abstract cause, in turn, uses Terror Management Theory to explain how perceived threats to one's group and belief systems coming from external agents can motivate altruistic death as a group fitness-enhancing strategy (Pyszczynski et al. Reference Pyszczynski, Motyl and Abdollahi2009). In the framework of active inference, altruistic self-sacrifice is motivated by an effort to reduce a perceived increase in the uncertainty of the social world. Cognitive anthropologist Scott Atran's work on suicide terrorism has yielded further clues on how some people come to die for “abstract ideas.” For Atran (Reference Atran2003), strong motives to defend an imagined community are not simply abstract: they are extrapolated from small-scale group bonding, where strong ties of solidarity – installed by military training – produce the motivation or willingness to die for one's friends, seen as an extension of a broader community. These examples can help us understand how new and competing normativities are also competitions for generative models of the world. Conflict over meaning and how the world ought to be invariably occurs within and between groups. We understand social change as occurring under these dynamics of optimizing world models through intra-group and inter-group competitions. We think that the FEP offers a way to account for both social stabilization and social change, and indeed for human historicity itself.

R2. Culture as embodied affective engagement with affordances

In describing human behavior as culturally patterned via regimes of attention that modulate salience and guide action, we aimed to ground our account in so-called 4EA approaches that emphasize the embodied, embedded, extended, enactive, and affective nature of cognition. As pointed out by several commentators, our attention-and-expectation-centered, information-heavy account left out important details on the affective, sensorimotor, physiological, and phenomenological dimensions of this puzzle.

Baggs & Chemero express the concern that our choice of terminology leads us to pursue an approach that is disembodied and inferential rather than embodied and enactive. They note that Vygotsky long ago proposed an approach to cultural learning as developmental engagements between social actors (initially child and parent) that is enabled by language (Vygotsky Reference Vygotsky1980). This developmental trajectory has been well documented (Tomasello et al. Reference Tomasello, Kruger and Ratner1993). The mother's discursive framing of the child's activity becomes the child's self-talk, which in turn becomes inner speech. The narrative self is born in this move from internalized dialog to monolog. Beyond this developmental picture, Baggs & Chemero insist that “language-involving cognition operates according to a different set of norms, and is not merely a more elaborated form of adaptive fitness.”

We agree that the ability to use language is a game changer for cultural learning. It allows humans to explore imaginary fitness landscapes, install higher-order priors without the need for individual lived experience, expand action repertoires by following recipes or instructions, and explicitly name, frame, debate, and deliberate about social norms. Most powerfully, language allows recursion and self-reflexivity. In theorizing how most human priors become saturated with other minds, and in this sense divorced from direct interaction with the world, we agree with Baggs & Chemero that this process can be cashed out under a Vygotskyian model of scaffolding via social learning and the education of attention. We have made this argument elsewhere (Ramstead et al. Reference Ramstead, Veissière and Kirmayer2016).

We strongly disagree, however, with Baggs & Chemero's claim that the affordances construct cannot be used to explain socially and culturally appropriate behavior. We see no reason to think the engagement of humans with their social and cultural worlds cannot be cast in terms of affordances. Indeed, Gibson himself used the concept of affordance to describe human social and cultural behavior (Reference Gibson1986, p. 119ff). The difficulty in accepting that dynamics of enculturation can be affordance-like, as we have argued elsewhere (Ramstead et al. Reference Ramstead, Veissière and Kirmayer2016), stems from the conservative definition of “affordance” adopted by many cognitive scientists, which applies the concept only to possibilities for action that can be “picked up” without instruction and do not require social learning to recognize and use (e.g., Moore Reference Moore2013).

Drawing in part from Chemero's (Reference Chemero2011) work, we have defined a cultural affordance as a relationship between the embodied skills of an encultured agent and relevant aspects of the cultural environment (Ramstead et al. Reference Ramstead, Veissière and Kirmayer2016). In our model, the relevance of an affordance is cashed out in terms of culturally shared priors about salience; and the ability to engage with cultural affordances corresponds to policy selection in active inference. In the target article, we provide an embodied dynamical account of how such engagement with a cultural niche is possible. That account explains what goes on “under the hood” (and indeed, “around the hood” in the environment) in scaffolded learning processes that stretch from imitative learning to explicit rule following and deliberation. It is with this view in mind that we have described culture as involving a set of deontic affordances. From this perspective, the cultural world not only solicits certain modes of attention and action, but also entails the obligation to respond to the world and other people in constrained ways. These obligations are felt in the body, expressed in habitual stances and actions, as well as (sometimes) in elaborated moral language and deliberation.

Clement & Dukes offer the important specification that culturally patterned practices do not simply direct attention in a value-free fashion, but entail the structuring of valence, as well. For these authors, “it is the valence and the intensity of others’ emotional expressions in particular that can be used to detect what is expected from each member” (para. 4). We could not agree more with this description of social learning as fundamentally affective, and culture as deeply axiological – less about “what is,” as Clement and Dukes put it, and more about “what matters and what is meaningful.” We understand this process as operating beyond explicitly normative situations, and extending to the affective qualities of the cultural world. For example, inferring whether someone's style of dress signals high or low social status, or assessing the “intrinsic” desirability of a prospective purchase, involves picking up on what relevant others expect the world to afford – simply learning fixed associations to social cues without this component of TTOM might lead to frequent errors as we move across contexts or niches. In other words, we are able to pick up the multidimensional valences assigned to specific features of the world by local regimes of attention. We can learn these valences as embodied dispositions to respond to cultural affordances, without explicit awareness of the relevant norms.

In their commentary, Smith & Lane show how the TTOM model could be expanded readily to make an explicit place for emotional processing and interoceptive embodied inference. In their approach, emotional processes are cast as emotional policy selection and emotional state inference (Smith et al. Reference Smith, Lane, Parr and Friston2019a). This extension of the formalism rests on an additional layer of parametric depth that biases action toward inference by the agent of its own affective states. We would add that interoceptive inferences are also modeled socially against the affective states of others – indeed, they are sometimes taught more or less explicitly in the way caregivers react to, and thereby teach us to react to and assign predictive meaning to our own internal states and social positions (Gendron & Barrett Reference Gendron and Barrett2018; Hoemann et al. Reference Hoemann, Xu and Barrett2019). This hierarchical and interpersonal structure of an active inference agent fits well with TTOM, which extends all the way from explicit high-level inferences involving emotional concepts to low-level automatic emotional response generation (Seth & Friston Reference Seth and Friston2016).

This model of emotion, when coupled with the concept of deontic policy selection (Constant et al. Reference Constant, Ramstead, Veissière and Friston2019b) discussed in the concluding section of the target article, can get us closer to what Buskell expects of TTOM: a view of characteristically human phenomena (such as radical organizational change), which requires an understanding of affect and normative behavior. We have begun some work in this direction (Constant et al. Reference Constant, Ramstead, Veissière and Friston2019b) showing how concept of deontic policy selection can help account for phenomena such as social conformity (Asch Reference Asch1955; Toelch & Dolan Reference Toelch and Dolan2015). In this model, we provide a formal description of how the acquisition of regimes of attention allows enculturated agents to zero in on what appears to be the most socially relevant response to a situation. This policy selection may have its own affective consequences which can influence subsequent action. This looping process follows directly from the TTOM model augmented with the concept of deontic action selection and the implementation of affect under the FEP by Smith et al. (Reference Smith, Lane, Parr and Friston2019a).

Van de Cruys & Heylighen provide a compelling example of affective social phenomena that is readily implementable with a model of affect coupled to TTOM. They propose that TTOM can help account for social phenomena that evoke, maintain, and are amplified by negative affect (e.g., dogmatism or radicalism). For instance, “guilt [can] also become a hidden cause efficiently explaining away someone's behavior in the eyes of others belonging to the culture.” Learning the mapping between one's actions, others’ suffering and feelings of guilt attaches an affective valence to policies that can guide behavior. Moral emotions such as guilt, then, contribute to social coordination in particular cultural contexts and indeed, as others have suggested, to the unique forms of pro-sociality found in human forms of life (Tomasello Reference Tomasello2009).

The commentary by Allen, Legrand, Correa, & Fardo suggests that our account is overly focused on cognitive and brain-bound prior beliefs, and does not sufficiently address embodied forms of inference that are ongoing at other scales. Indeed, work by Allen's group demonstrates that visceromotor processes and interoception can be usefully cast as forms of active inference (Allen et al. Reference Allen, Levy, Parr and Friston2019). We agree that social and cultural processes such as those described in the TTOM model may be built on a basis of embodied, interoceptive priors some of which have a long phylogenetic history. We would simply add that cultural processes also allow priors to be installed from the top down, through linguistic coding and mimetic processes, so that in any actual instance, culturally learned behavior is likely to be an outcome of both interoceptive and exteroceptive influences and both embodied and linguistic practices (Di Paolo et al. Reference Di Paolo, Cuffari and De Jaegher2018). This puts cultural history and learning on an equal footing with evolutionary and co-evolutionary influences.

The ongoing social dynamics of affective influence are foregrounded in several commentaries. As Mouras reminds us in his commentary, shared affective, motivational, and sensorimotor processes can be observed and experimentally manipulated directly in social interactions. Mouras's discussion of experiments on postural correlates of in-group bias in empathy for pain offers many useful pointers on how to operationalize models of shared agency in TTOM. Mouras and colleagues have made elegant use of the Minimal Group Paradigm, in which in-groups and out-groups are arbitrarily constituted in a lab environment (Tajfel et al. Reference Tajfel, Billig, Bundy and Flament1971), to demonstrate the quick, flexible, yet precise ways in which embodied and affective joint models can emerge as long as an assumption of shared goals is present. The experiments described by Mouras show that empathy for others’ pain can occur quickly once in-group conditions obtain, and that merely imagining fellow group members in pain affects sensorimotor processes with effects such as automatic (nonconscious) stiffening of posture. These findings echo recent studies on physiological correlates of group bonding that evince shared responses in both participants and spectators during high-arousal conditions of group synchrony, such as marching in a stadium (Jackson et al. Reference Jackson, Jong, Bilkey, Whitehouse, Zollmann, McNaughton and Halberstadt2018) and ritual firewalking (Konvalinka et al. Reference Konvalinka, Xygalatas, Bulbulia, Schjødt, Jegindø, Wallot, Van Orden and Roepstorff2011). Taken together, these studies point to an emerging socio-physiology that reveals some of the embodied interactions that undergird TTOM.

Further points raised by commentators help us describe social interactions as shaped by relational dynamics that include developmental histories and processes of attachment and affiliation. Strand's commentary, on the importance of attachment patterns with supportive others as potential modulators of cultural learning, raises important issues, including the claims that: (1) reward learning mechanisms are central to the patterning of culture via Bayesian inference; (2) neural pathways subserving these mechanisms are likely evolutionarily old, and (3) attachment styles are an important locus of cultural differences. In support of this view, Strand cites recent studies showing that collectivistic cultures exhibit a higher prevalence of insecure-anxious individuals, compared to higher rates of insecure-avoidants in individualistic cultures (Yamagishi & Hashimoto Reference Yamagishi and Hashimoto2016). We agree that attachment and reward mechanisms are central in the phylogeny and ontogeny of culture. On this view, cultural evolution has favored modes of interaction that fulfill and leverage our evolved need for rewarding attachments, enabling the construction of adaptive social ties in each generation (Boyer & Liénard Reference Boyer and Liénard2006; Reference Boyer and Liénard2008). We can think of the patterning of culture in terms of kinds of affiliative relationships, communal activities, rituals, and ceremonies that serve a general function of social uncertainty minimization, and contribute to specific, adaptive cooperative action, whereas offering a protective buffer against the risks of loneliness and rejection.

Strand's comments are also helpful to understand another strange loop in cultural evolution: attachments are central to the patterning of culture, yet are themselves culturally patterned. In Yamagishi and Hashimoto's (Reference Yamagishi and Hashimoto2016) niche construction model of cross-cultural differences in attachment, an avoidant style is understood as adaptive in individualistic cultures (in which help-seeking is socially costly), whereas anxious attachments confer social adaptiveness in cultures that require high levels of cooperativity and social deference (where individualism is socially costly). Of note here, both “individualism” and “collectivism” are collectively patterned via social norms and deontic cultural affordances that demand different levels of autonomy in specific contexts.

Pezzulo, Barca, Maisto, & Donnarumma's (Pezzulo et al.) insightful commentary on social epistemic action further assists us in specifying that, in addition to being rewarding for its own sake, human joint action often entails conveying cues for the sake of others. Pezzulo et al. explain that unlike epistemic actions (aimed at updating one's contextual beliefs about relevant features of the world) and pragmatic actions (acting on the world after resolving contextual uncertainty), social epistemic actions serve the function of directing others’ attention to patterns and regularities that we want them to infer. This is precisely the phenomenon we have referred to as regimes of attention (Ramstead et al. Reference Ramstead, Veissière and Kirmayer2016). The authors point out that although such actions can be verbal, humans routinely employ a wealth of kinetic and sensorimotor cues (what they cleverly call “motionese”) to convey the internal states and intentions they want others to infer, or to “signal more directly what and where salient information is.” Pezzulo et al. rightly emphasize the importance of social epistemic actions on “faster timescales” (e.g., during teaching, learning, and real-time communication) but we also find the notion useful for an account of the evolution of human sociality on longer co-evolutionary and developmental timescales. The existence of species-wide ostensive or indexical cues, such as pointing (to direct others’ attention), or holding out a hand (to signal helping behavior) offers strong evidence for evolutionarily old strategies for communicating universally legible intentions and emotions to others. The presence of laughter and vocal crying in apes, humans, and pre-verbal human infants offers further evidence for phylogenetically old and involuntary (or automatic) drives for social epistemic actions. Indeed, both actions occur under weak or absent voluntary control, and are best described as forms of honest signaling that reveal true internal states (Provine Reference Provine2017). For example, an experiment using short audio clips found that hearing joint laughter was sufficient for participants from 24 diverse language groups to differentiate pairs of friends from strangers (Bryant et al. Reference Bryant, Fessler, Fusaroli, Clint, Aarøe, Apicella, Petersen, Bickham, Bolyanatz, Chavez, De Smet, Díaz, Fančovičová, Fux, Giraldo-Perez, Hu, Kamble, Kameda, Li, Luberti, Prokop, Quintelier, Scelza, Shin, Soler, Stieger, Toyokawa, van den Hende, Viciana-Asensio, Yildizhan, Yong, Yuditha and Zhou2016).

This view of cultural transmission as entailing multiple forms of learning falling on a spectrum from automatic “picking up” of information to explicit teaching and deliberation assists us in taking up Gweon's helpful pointers for “a more complete picture of social learning.” Gweon points out, building on Cisbra and Gergely's (Reference Csibra and Gergely2009) Natural Pedagogy paradigm, that children are intrinsically motivated to identify high-utility, high-quality information from reliable sources, and are subsequently motivated to teach that information to their peers. We agree with Gweon that this view of children as active agents of both Bayesian inference and attention-directing communication is key to understanding cultural transmission. In this line, we welcome Michael & de Bruin's invitation to consider the importance of mind-shaping mechanisms and to clarify how the TTOM model of enculturation accounts for a full implicit-to-explicit inference spectrum. We find the postulated existence of evolutionarily old, developmentally early behavior-influencing mind shaping mechanisms (De Bruin and Strijbos Reference De Bruin and Strijbos2020; Zawidzki Reference Zawidzki2013) fully compatible with the account we have outlined above. Our account of social cognition recognizes multiple levels and instances of inference from the “quick and embodied” to the effortful and deliberative. In doing so, we join previous “multi-system” efforts by Christensen and Michael (Reference Christensen and Michael2016) intent on resolving tensions between theory–theory, simulation, and embodied accounts of mind-shaping and mindreading.

R3. Formal clarifications: What the FEP adds to our understanding of cultural learning

The TTOM model gains precision and explanatory power by leveraging the active inference framework. Unfortunately, some of the key features of active inference have been misunderstood in the recent literature, as well as in some of the commentaries. Clarifying these misunderstandings is crucial to appreciate how TTOM can be applied to model cultural learning.

R4. Applications and empirical tests of the model

We welcome comments from Brown, Brusse, Huebner, & Pain (Brown et al.), who speak from the vantage point of cognitive archeology. They agree that unifying models can help “dissolve disputes by bringing rival positions under a single theoretical framework,” but remain unconvinced that the FEP can offer such a framework. In particular, Brown et al. express their doubt that our model can generate testable predictions. This concern is shared by Dołęga, Schlicht, & Dennett (Dołęga et al.), who argue that our account does not sufficiently distinguish between the levels of explanation. They suggest that our model is best understood as applying to Marr's computational level of explanation (i.e., characterizing the information processing task). Of course, this specification has consequences for theory building. For example, by showing that one can cast both mindreading and niche construction as forms of active inference, we show that there is no need to choose between competing alternatives. Although our account shows how internalist and externalist accounts can both be justified in some contexts, Dolega et al. claim that it does not provide us with a way of deciding between the two options for any given phenomenon. We are thus, in their view, over-inclusive and gain explanatory scope at the cost of explanatory power and precision. These charges are well put and pose an important challenge to the usefulness of the TTOM model. However, we think the apparent empirical limitations of the model reflect its current formulation in generic terms. Detailed examples in particular domains are needed to produce specific hypotheses and test the fit of models with experimental data. Moreover, plenty of evidence (both neural and behavioral) is already available to demonstrate the viability of the principles that ground our model as applied to individual cognition (e.g., Friston Reference Friston2010; Keller & Mrsic-Flogel Reference Keller and Mrsic-Flogel2018).

The methods we employ to formulate TTOM are borrowed from theoretical neurobiology. These basic principles have been applied a wide range of problems and tested against empirical data, and we know they can explain and reliably predict many features of individual human behavior (Cullen et al. Reference Cullen, Davey, Friston and Moran2018; Mirza et al. Reference Mirza, Adams, Mathys and Friston2016). The application of these methods to cognitive systems beyond individual humans and their brains, however, is more recent. The challenge is to figure out what should count as empirical evidence for an ecologically extended, and spatiotemporally scalable model like TTOM. We think the paleoanthropological record holds some clues that may help us answer these questions, whereas addressing a second point raised by Brown et al. about what they perceive as our model's failure to address debates between the internalist and externalist approaches to cognition.

The TTOM model, to clarify, proposes an “interactionist” view of cognition as at once internal to individual brains and bodies, and extended, embedded, and enhanced through external features of the social world. Work in paleo-anthropology – that emphasizes the co-evolution of human cognition with their anthropogenic environment modifications – provides explicit examples of this process. For example, Dietrich Stout's analysis of cumulative cultural evolution in the lower paleolithic suggests that capacity for Theory of Mind was significantly enhanced through selective pressure to teach and learn the making of stone tools (Stout Reference Stout2011; Stout et al. Reference Stout, Passingham, Frith, Apel and Chaminade2011). In other words, tool production played a crucial role in selecting for better perspective-taking abilities. Using ethnographic evidence from the Ju/’hoansi Bushmen, Wiessner (Reference Wiessner2014) argued that the invention of fire, and the ritual practice of firelight talks at nighttime played an important role (in both phylogeny and ontogeny) in “evoking higher orders of theory of mind via the imagination, conveying attributes of people in broad networks (virtual communities), and transmitting the ‘big picture’ of cultural institutions that generate regularity of behavior, cooperation, and trust at the regional level.” On this interactionist view, then, fire also played a role in the evolution of cognition. On more recent accounts of the “Broad Spectrum Revolution,” a further leap in human cognitive, technological, and cooperative foraging skills occurred around 40,000 years ago after anthropogenic depletion of megafauna provided new selective pressures for diversification of hunting and gathering strategies (Sterelny Reference Sterelny2007; Reference Sterelny2011; Zeder Reference Zeder2012).

These diverse examples point to multiple ways in which humans have learned to resolve uncertainty through various TTOM-enhanced, environmentally mediated pathways. We have outlined a unifying account of human cultural co-evolution, which can be operationalized under the FEP. However, detailed predictions and testable hypothesis will have to take on board historical contingency and boundary conditions.

R5. Modeling diversity and difference in TTOM

TTOM models the acquisition of culture by a generic agent in a stable, relatively homogeneous social niche. However, the reality in most contemporary societies is one of the high levels of diversity. In our responses to Mirski et al., we have discussed how social change, emerging normativities, and social conflict can be described as processes of group selection and optimization. Many levels of difference and variation remain to be examined in our account.

Intra-cultural and inter-individual diversity, as well as non-conformity and intra-individual change (when individuals change their mind over time or in response to changes in context) raise thorny questions for TTOM. Whether our ongoing transition to a digital world with unpresented access to billions of minds is a game-changer for TTOM is another pertinent question. The commentaries by Christopoulos & Hong, Bouizegarene, and Clark are especially helpful in addressing these essential refinements to the model.

R6. Applications of TTOM to psychiatry

In addition to a better understanding of social problems, we think our model can help advance the study of psychopathology by highlighting interactions between the social and neural dimensions of various disorders.

In their wide-ranging commentary, Dumas, Gozé, & Micoulaud-Franchi (Dumas et al.) provide a very useful extension of our discussion of the shared affective and phenomenological processes that underwrite human experience. To advance an “interactive turn in psychiatric semiology,” Dumas et al. call for a unified computational framework that could account both for sociocultural variations in psychiatric symptomatology (including how patients interpret and enact different explanatory models of illness, and how clinicians leverage different explanatory models in their interactions with patients) and current neuroscientific findings on physiological disturbances in the brain. Indeed, they go beyond our original argument, to suggest that the TTOM model can contribute to the development of methods that examine how the whole extended cultural phenotype of humans interacts with the whole genotype in what they call social physiology. We welcome this ambitious project, and applaud Dumas et al.'s broader discussion of multi-scale approaches to psychiatric disorders grounded in active inference – in particular, their comments on neurodevelopmental conditions such as autism and schizophrenia that have been described as minimal self-impairments, and that may be operating at more basic levels than the cognitive-behavioral loops that are the focus of the current TTOM model.

The potential relevance of TTOM to psychiatry is also brought out by Lifshitz & Luhrmann in their commentary on the ways in which culture can structure the affective valence and content of hallucinations. In addition to pointing out, as Gold and Gold (Reference Gold and Gold2015) have done elsewhere, that all types of hallucinations appear to involve social scenarios and “the relevant others that humans think through” to guide their existence, Lifshitz and Luhrmann provide compelling examples of sensorially rich “hallucinations” (interacting with imagined agents who are not actually present) in contexts such as religious practice that do not involve any pathology or underlying brain dysfunction. This points to the far-reaching ways in which culture can affect perception. Their review of recent findings on the cross-cultural patterning of experience in schizophrenia also shows how basic TTOM impairments, which likely do entail dysfunctions of an organic nature (and, as such, might lead to behavior recognized as dysfunctional in any cultural context), predict widely different degrees of distress depending on the specific cultural information conveyed in hallucinations, and on local cultural assumptions about the nature of affliction. Lifshitz and Luhrmann thus raise important questions on the reach and limits of cultural influences on phenomenological plasticity of mental disorders in general.

Finally, we welcome the suggestions for extending TTOM offered by Bolis & Schilbach. We find their dialectical account of the interactions and relationships between the agents compelling and appreciate their call to extend our model to take into account the diversity of human experience, including neurodiversity. The idea that difficulties in interaction between the neurotypical and neurodiverse individuals result from the misalignment between individuals, rather than simply from some underlying deficit in the neurodiverse population, has far-reaching implications for our understanding of and attitudes toward individual differences and difficulties in adaptation. Their suggestion to focus more closely on the dynamics of real-time interactions among humans points to fruitful avenues for empirically testing the TTOM model.