In chapter 10, Spelke (Reference Spelke2022) articulates what is, in our view, the most thought-provoking contribution in her expanded review of the core knowledge (CK) theory. She argues that the concept of “social agent,” that is, an agent who interacts with others through object-directed actions (e.g., giving an object or demonstrating its function to someone), is not a part of CK. Instead, infants acquire this concept through exposure to their native language. By participating in a community of relevant and efficient speakers, they discover how to combine the representations of two distinct CKs: An agent system that interprets agents as acting on their environment to bring about instrumental goals, and a social cognition system that interprets agents as interacting with others through shared experiences. Spelke argues that prior to this discovery, occurring at around 10 months of age, infants fail to interpret social interactions involving object-directed actions because they lack the means to integrate the aforementioned systems. In this commentary, we express empirical and theoretical reservations regarding this proposal.

To begin with, the developmental literature reveals several exceptions to the proposed timeline, suggesting that, at least in some settings, young infants are capable of interpreting the instrumental goals of “social agents.” For instance, by 4 months of age, infants begin to form expectations of distributive fairness, which requires representing interactions based on resource transfer (Buyukozer Dawkins, Sloane, & Baillargeon, Reference Buyukozer Dawkins, Sloane and Baillargeon2019; Geraci & Surian, Reference Geraci and Surian2023). Around 6 months of age, infants infer social dominance on the basis of relative group size in scenarios involving agents with conflicting instrumental goals (Pun, Birch, & Baron, Reference Pun, Birch and Baron2016; a finding that Spelke explains away by suggesting that infants construed the agents' movements in the presence of group members as a “social gesture”: p. 412). By this age, infants also begin to show a manual preference for prosocial characters (Hamlin, Wynn, & Bloom, Reference Hamlin, Wynn and Bloom2007; Kanakogi et al., Reference Kanakogi, Inoue, Matsuda, Butler, Hiraki and Myowa-Yamakoshi2017), suggesting an incipient understanding of instrumental actions directed at increasing the utility of other agents. Even if the interpretation of these findings as providing evidence for the representation of second-order social goals remains contested (Powell & Spelke, Reference Powell and Spelke2018), the success of the manual-choice measure, which entails selecting one of two ostensiviely presented characters, shows that young infants can appropriately respond to acts of offering. Furthermore, by 9 months of age, infants can already represent the joint goals of agents involved in simple forms of collaboration (moving synchronously toward a common target; Begus, Curioni, Knoblich, & Gergely, Reference Begus, Curioni, Knoblich and Gergely2020), and start to leverage ostensive signals not only to guide their attention to objects, but also to learn about their featural information (Okumura, Kanakogi, Kobayashi, & Itakura, Reference Okumura, Kanakogi, Kobayashi and Itakura2020; Thiele, Hepach, Michel, & Haun, Reference Thiele, Hepach, Michel and Haun2021).

Even discounting such exceptions, the failure of younger infants in interpreting the goals of “social agents” does not necessarily indicate a lack of corresponding conceptual frame. As Spelke points out, CKs compete for attentional resources. The task of identifying a teleological relation between the participants of an object-mediated interaction is bound to be more challenging than in the case of instrumental actions (or acts of social engagement) simply because the former, being a more structurally complex event, generates more concurrent goal hypotheses. For instance, typical helping actions (e.g., A helps B to open a box) contain cues of social affiliation (A approaches B and mirrors some of their movements) and first-order instrumental goals (A brings about a change of state in the box), either of which may prime well-formed goal hypotheses that prevent observers from recognizing the teleological dependency between the helper's object-directed efforts and the facilitation of the helpee's goal fulfillment (Schlingloff-Nemecz, Tatone, & Csibra, Reference Schlingloff-Nemecz, Tatone and Csibra2023). Similarly, instances of taking can be reduced to nonsocial acts of object acquisition, disregarding the effects that the action has on the original resource possessor (the takee). Supporting this possibility, infants and adults spontaneously interpret taking actions involving unreactive takees as nonsocial instances of resource acquisition (Tatone, Geraci, & Csibra, Reference Tatone, Geraci and Csibra2015; Yin, Csibra, & Tatone, Reference Yin, Csibra and Tatone2022; Yin, Tatone, & Csibra, Reference Yin, Tatone and Csibra2020), leading to the omission of the takee from the event frame.

The aforementioned evidence cautions against the claim that the composite concept of “social agent” does not compete for attention with concepts from other CKs (p. 422). To date, we do not know of any experimental evidence attesting to this claim (for a similar argument, see Revencu & Csibra, Reference Revencu and Csibra2023). In fact, the above findings suggest that even adults, who should presumably leverage such a “social agent” concept (and thus bypass the hurdle of competing available frames), similarly experience the conflict between interactive and instrumental construals. Under our account, the existence of such conflict well beyond infancy is evidence that arbitrating among competing frames is a challenge inherent to the process of action interpretation (Tatone, Reference Tatone2022), and not specific to developmentally early instances of attentional-resource competition among CKs. Importantly, unlike Spelke's, this account does not entail any strong kind of dispositional ascription: it is not the agent, but their episodic behavior, that is being construed as instrumental or social, based on the available goal hypotheses. Because of this, infants (and adults) should be able to entertain agents as having goals pertaining to different CKs over time (e.g., engaging with a partner now, and pursuing their own instrumental goals later). This does not seem to be the case for Spelke, who characterizes CK frames as an instance of kind categorization, which, upon being deployed, constrains the type of potential goals that an agent may in principle pursue (e.g., acting on the environment vs. establishing social engagement).

The notion that the concept of “social agent” may derive from the infants' pragmatic interpretation of object-directed communicative acts raises further concerns. To begin with, it is not quite clear how the linguistic combination of two distinct relations (one toward an object, the other toward a partner) could guide infants to discover specific utility functions underpinning a given social behavior (Powell, Reference Powell2022). Even if infants could linguistically expand the interpretation of an agent's object-directed actions to include social goals based on “engagement, shareability, and experience” (p. 349), this affiliative motive does not seem sufficiently precise to distinguish between different types of interdependence (Aktipis et al., Reference Aktipis, Cronk, Alcock, Ayers, Baciu, Balliet and Winfrey2018), such as prosocial versus mutualistic interactions. Arguing that the relevant distinctions may be acquired by participating in the linguistic community is rather problematic, for two reasons. First, it would require linguistic inputs to exhibit a tight mapping between thematic and beneficiary roles (i.e., who acts vs. who stands to gain from the action), which it is not always the case (Newman, Reference Newman1996). Relatedly, linguistic descriptions of social interactions do not reveal the relational concepts (e.g., dominance, reciprocity) underlying these episodes (Fiske, Reference Fiske1992), despite these concepts being nonetheless inferred from the occurrence of specific social behaviors already by prelinguistic infants (Mascaro & Csibra, Reference Mascaro and Csibra2012; Tatone & Csibra, Reference Tatone, Csibra, Denison, Mack, Xu and Armstrong2020).

More generally, Spelke's reliance on natural language (as opposed to a language of thought) for combining core representations seems to run into an evolutionary conundrum: if the acquisition of the concept of “social agents” indeed depends on interactions with efficient and relevant speakers geared with the appropriate conceptual frame, how is this process originally bootstrapped? That is, what kind of combinatorial resources could the first language users tap into? Furthermore, if, as Spelke seems to recognize, concepts of “social agents” capture universal and species-wide experiences (e.g., giving, communicating), unlike a myriad of other concepts that are the product of socioculturally contingent traditions (e.g., integers), what accounts for this critical distinction, considering their surface similarity (i.e., being all discovered by participating in a linguistic community)? The reliable acquisition (across ontogenesis and different sociocultural niches) of the “social agent” concept, one may concede, attests to its adaptive relevance and centrality to human livelihood. But if that is the case (as we believe), we should expect concept learning to have been canalized through the emergence of dedicated mechanisms over evolutionary time (Baldwin effect; Barrett, Reference Barrett, Eliis and Bjorklund2005). It is hard to assess whether the developmental timeline that Spelke sketches – with “social agents” being systematically the first noncore concept that infants acquire – bears evidence of such preparedness. Spelke does entertain the possibility of adaptive canalization when discussing our predisposition to associate snake avoidance with fear displays (p. 416), but considers specialized mechanisms to be mostly the province of the cognition of nonhuman animals.

With respect to nonhuman species, the proposal that the understanding of object-mediated interaction is heralded by the acquisition of natural language leaves us with two possible scenarios. One is that nonhuman animals, equipped with the same CKs as humans (as per Spelke's criteria), lack a concept of “social agent,” and would thus be forced to navigate interactions based on food provisioning by construing them as either instances of nonsocial resource collection or goal-demoted forms of social bonding. This possibility seems unlikely, given that these interactions are among the first instances of caregiving in species with dependent offspring (from birds to social carnivores; Jaeggi & Van Schaik, Reference Jaeggi and Van Schaik2011), and are thus pivotal to growth and survival. A more plausible alternative is that, in light of the fitness-relevant role that these interactions have, nonhuman animals evolved specialized mechanisms for recognizing and appropriately partaking in them (e.g., food offering calls or gestures; Jaeggi & Gurven, Reference Jaeggi and Gurven2013; Scheid, Schmidt, & Noë, Reference Scheid, Schmidt and Noë2008). But if these arguments can be appealed to for explaining the proficiency of nonhuman animals in identifying certain interactions, what prevents a similar logic from being applied to our species?

Proposals in this direction have already been put forth (Frankenhuis & Barrett, Reference Frankenhuis, Barrett, Rutherford and Kuhlmeier2013), also with regard to object-mediated interactions. In the domain of giving, for instance, it has been argued that infants are prepared to treat giving-like outcomes as potential goal states (Tatone et al., Reference Tatone, Geraci and Csibra2015) and to readily structure event representations in the service of capturing such goals. Buttressing this claim is the evidence that infants ascribe giving goals based on minimal cues of possession transfer, even when other perceptual indicators of engagement (e.g., social approach or receipt acknowledgment) are amiss (Tatone, Hernik, & Csibra, Reference Tatone, Hernik and Csibra2019). Such preparedness, we maintain, is to be expected because a generative model of action understanding governed solely by assumptions of instrumental rationality would be unable to compute the utility of actions, such as resource transfer, which inherently entail the voluntary self-imposition of costs. Since Spelke's model relies on similar assumptions, it also requires the repertoire of instrumental goals that infants are originally geared with to be expanded to include the suite of social goals directed at increasing or decreasing another agent's utility. How linguistic scaffolding enables the discovery of such goal states and their integration however, is presently not discussed.

Spelke's theory of CK represents a trailblazing milestone in the study of human cognitive development. Drawing on an impressive body of experimental evidence, her latest book cements the tenets of this theory and expands its scope in daring new ways. Our commentary provided critical counterpoints to the idea that learning a natural language would allow infants to combine instrumental and social construals in the service of a new concept of “social agents.” In our view, the present proposal falls short of explaining with sufficient precision how this concept is discovered, how its utility function is derived and distinguished from the core concept of social engagement, and what accounts for its systematic acquisition. We hope our critical analysis may help spur further elaborations.