The critical importance of the social environment in the emergence of language was put forward by Vygotsky (1978) and Bruner (1983). In their studies, which laid the foundations for a social constructivist view on language, the social world is discussed as a special kind of environment, distinct from the physical world. It is an indispensable yet external context for learning and development, not an internal matter of mind and cognition (Pléh, 2024). Additionally, their basic assumption, like many others’, has been that communication requires a code-based signaling system, that is, language. The idea that human communication is built on social cognition, as it is not mere information transmission, reverses the above order: communication may need to precede language acquisition. There are two dominant views regarding the developmental origins of human-specific communication and the role social cognition plays in it. Tomasello (2008), by and large, follows the (Neo-) Gricean tradition in emphasizing cooperation as the basis of communication and language in his joint attention framework; Gergely and Csibra’s Natural Pedagogy theory fits well with RT and the Post-Gricean perspective when proposing that ostensive cues play a central role in communicative interactions even in preverbal infants (Csibra and Gergely, 2009; Gergely and Csibra, 2013).

Although the Neo-Gricean and the Post-Gricean models are not mutually exclusive, they emphasize different aspects of communicative interactions and assume different sufficiency and necessity conditions for establishing common ground. Both models are primarily interested in explaining how to identify communicative intentions, with less emphasis on how information is actually transmitted (i.e., the informative intention). While communicative intentions may be recognized either through a code-like signaling system (ostension) or through the motivation for cooperation (joint attention), neither model argues for the necessity of mentalization, even though both involve attending to a social partner and identifying the partner’s focus of attention as a referent.

Another person’s attention or goal (i.e., the referent of a communicative interaction), however, cannot be but an attribution of attention (Elekes and Király, 2021) or of a goal (Csibra and Gergely, 2007). It may be argued that recognizing the attention or the goals of social partners relies on some sort of non-mentalistic mechanism, for example, on teleology (Gergely and Csibra, 2003). However, teleology does not seem to suffice to explain communication non-mentalistically because in communicative interactions, referent objects are not the goals of agents but the goals of the communication itself. Moreover, information transmission takes place only after minds establish common ground through joint attention and/or ostensive signals. Yet, minds never literally “join”; thus, common ground can emerge only in the minds of the two interlocutors separately. The two parties may mutually assume that the other has an identical belief about the state of affairs as themselves, but this can only be an ascription of a mental state to the other party. The point is that the machineries proposed for setting up communicative interactions by identifying communicative intentions are fundamentally attention-directing and attention-checking systems. These systems are aimed at identifying what the partner is calling the infant’s attention to, which is the partner’s own focus of attention, to establish that both parties have the same referent in mind. Perhaps due to the ambiguity of the term “communicative intentions,” a considerable group of researchers seems to believe it is obvious that language, like communication, involves mentalization, while another large group appears to hold it is rather obvious that neither communication nor language does.

There is another intriguing contradiction regarding pragmatic inferences, as highlighted by Bohn and Frank (2019). There is a long line of research arguing that infants reason skillfully about intentions already during language acquisition (Nelson, 1973; Bates, 1976; Bloom, 2000; Tomasello, 2003). Another line of research reports the difficulties children experience in deriving pragmatic inferences to interpret intended meanings (Papafragou and Musolino, 2003; Huang and Snedeker, 2009). Bohn and Frank propose to resolve the contradiction by defining communication as social cognition and reasoning about the goals of communicative partners. The Rational Speech Act framework suggests that pragmatic reasoning integrates all the elements of inferential communication, which are in place early on and foster the gradual development of language comprehension (Bohn and Frank, 2019).

The apparent contradiction may stem from the ambiguity of the term “communicative intentions.” It may refer to the social-cognitive inferential mechanisms employed to identify the intent to communicate, which may be present from early infancy onward on the one hand. It may also refer to the inferential mechanisms applied to the transmitted information content to recover meaning as intended, which may be challenging even for kids, on the other. While the former is the utilization of pragmatic-inferential mechanisms to set up a communicative infrastructure for the upcoming information (cf. Tomasello), the latter involves enriching the already available information with pragmatic inferences (cf. Grice).

The above two kinds of ambiguity surrounding the term “communicative intention”—(1) whether it is mentalistic or not and (2) whether interpreting it appropriately is easy (already in infancy) or difficult (even in late childhood) —seem to be orthogonal (Table 1). Some argue that inferring communicative intentions is essential for word learning, is available already in infancy, and involves mentalization (Bloom, 2000; Papafragou, 2002; Tomasello, 2003; Thompson, 2014). Others assume that word learning is based on inferential mechanisms, but it does not require mentalization (Sperber and Wilson, 1986; Csibra and Gergely, 2009; Bohn and Frank, 2019). Some show how non-mentalistic pragmatic inferences are challenging for kids when interpreting, for example, scalar inferences (Papafragou and Musolino, 2003; Huang and Snedeker, 2009; Noveck, 2018), contrastive inferences (Kronmüller et al., 2014), or logical terms (Noveck and Chevaux, 2002; Pouscoulous and Noveck, 2009). Finally, the position that pragmatic inferences develop slowly but co-develop with and involve mentalization has also been put forward (Rubio-Fernandez, 2021). The debate boils down to two key questions: (1) can communicative intentions be truly non-mentalistic (i.e., inferred based on code-like signals or regulative rules) and (2) what level of mentalization may be available in infancy (i.e., to what extent language acquisition may or should be tied to it)? Both of these questions are going to be addressed later on.

Taken together, there seems to be an agreement that social cognition, in the form of pragmatic inferences, plays a key role both in setting up communicative interactions and in deriving the implied meaning of utterances behind words and sentences. The first mechanism appears to precede the transmission of the linguistic code (i.e., setting up a communicative interaction, either via cooperation and joint attention or via ostensive communicative signals), while the second one follows it (i.e., enriching it inferentially to interpret it, either via checking for violations of maxims in a cooperative framework or via carrying out a relevance-based calculus). The first one seems to pertain to “communicative intentions” in the narrow sense: a scaffold of informative intentions. The second one uses the term in the broad sense: inferring meaning as intended. It may be argued that the latter actually concerns informative intentions, but this is not entirely clear from the literature. “Intended meaning” is typically referred to as what was intended to be conveyed (i.e., communicated), not what one was intended to be informed of. The term “information” may be the culprit here: it could mean either the content (utterance) or the form (sentence)—perhaps due to the remarkable influence of the code model. Notably, the question of how social cognition modulates the information content, specifically the link between linguistic form and semantic content, seems to have gathered limited attention. Yet, the linguistic signal is the most variable and rich source of input entering pragmatic inferential mechanisms.

From an information transmission point of view, it may not be surprising that language, as a stimulus class, enjoys a special status in human ontogeny. Well before birth, prenatal humans begin to pick up the prosodic properties of their native tongue (Abboub et al., 2016), and right at birth, they are sensitive to a range of physical-acoustic and phonological features of language (Werker and Tees, 1999). Newborns prefer speech to matched non-speech, forward-going speech to backward speech, their mother’s voice to other female voices, and their native language to unfamiliar languages; they can also differentiate between languages based on rhythmic properties even if they have never heard them before and can detect word boundaries, discriminate lexical stress, and even distinguish function words from content words based on acoustic characteristics (reviewed by Gervain and Mehler, 2010).

During language acquisition, infants use several of these acoustic properties, including rhythm (Goswami, 2022) and statistical distributional patterns (Kujala et al., 2023), to “break into the code” of language by approaching it from its information-transmitting potential. The acoustic-phonological-prosodic properties of language lend themselves readily to being deciphered as a code. First, the set of sounds the human vocal tract can produce is limited, making it computationally manageable. Second, phonology is not only governed by rules but also carries information about higher-level syntactic operations, both of which involve code-based computational structures. Even newborns rely on prosody and statistical learning of transitional probabilities to identify word boundaries (Fló et al., 2019). To segment the continuous speech stream into potentially meaningful units, they also employ various dedicated mechanisms to learn about the segments themselves (Fló et al., 2022). They also utilize innate pattern recognition mechanisms to pick up repetition structures (i.e., pseudowords with ABB structure, e.g., “mi-zu-zu,” as opposed to ABC random structures, e.g., “mi-zu-ka”) (Gervain et al., 2008). The repetitions may be engaging for them because they reveal a rule that may indicate syntactic structures, thereby providing a better opportunity to learn (about) language. Such sensitivity to rule-like structures extends to musical tones as well; however, only pseudowords, not tones, activate the left inferior frontal regions (Nallet et al., 2023). These regions include Broca’s area, which is responsible for processing the structural properties of language in adults (Musso et al., 2003; Liakakis et al., 2011; Friederici, 2012) and responds to language at birth (Peña et al., 2003; Perani et al., 2011). Newborns detect not only repetition structures but also their sequential position (ABB vs. AAB), and these two properties seem to be the two fundamental building blocks of any code-based system (Gervain et al., 2012). These findings strongly suggest that language is a unique signal for humans, engaging dedicated mechanisms, from statistical learning to pattern recognition, to identify word-like units and grammar-like rules right from birth. Newborns appear to be very well equipped to unpack the structural properties of the code system humans use to transmit information.

In the process of acquiring linguistic meaning, the only communicative cue that has been suggested to signal communicative intentions and is linguistic in nature is infant-directed speech (IDS) (Csibra and Gergely, 2009; Gergely, 2010); also called “motherese,” its characteristic prosodic pattern includes higher and broader pitch, greater amplitude variation, and slower speed than typical adult speech (Csibra, 2010). Although there seems to be some cultural variation (Cristia, 2023), sensitivity to IDS appears to be innate, present at birth (Cooper and Aslin, 1990), and universal (Fernald, 1992). The perceptual features of motherese appear to open the gateway toward the content of speech: IDS may simultaneously carry information about communicative and informative intentions. In fact, Sperber and Wilson (1986) consider language to be an ostensive cue in and of itself. Every communicative act carries its own relevance by definition, and a linguistic utterance is clearly communicative—at least for adults. Infants may exploit the prosodic layer of IDS for communicative intentions to gain access to its contents (i.e., informative intentions). IDS modulates electrophysiological responses to faces in 4-month-olds, perhaps because it generates communicative expectations (Sirri et al., 2020). It is interpreted as an ostensive cue by 5–6-month-olds, just as eye-gaze (Senju and Csibra, 2008; Parise and Csibra, 2013; Lloyd-Fox et al., 2015). It also facilitates 7-month-olds’ cortical tracking of speech (Kalashnikova et al., 2018). IDS appears to serve as a self-referential linguistic inroad toward linguistic meaning as it induces a communicative interpretation of the linguistic code. As an ostensive signal, it creates an expectation that incoming information refers to kinds and not individuals (Gergely and Csibra, 2013), which clearly aids word learning since, with the exception of proper names, words refer to categories.

Infants also appear to realize early on that language may carry information. In a series of remarkable experiments, Vouloumanos and colleagues (Martin et al., 2012; Vouloumanos et al., 2012; Vouloumanos, 2018) showed that infants expect speech—but not coughing or humming—to transmit information about intentions (i.e., objects preferences). Even 6-month-olds showed this expectation even when they had no chance to understand the transmission because it was non-sensical or foreign to them (Vouloumanos et al., 2014). Eleven-month-olds may prefer to interact with native speakers because they expect them to share information (Begus et al., 2016). Intriguingly, signal properties may modulate such expectations.

Using the so-called Flatfish paradigm, Tauzin and Gergely demonstrated that 10.5-month-old infants identify beeping entities as agents with preferences, but only if they exchange varying tone signals—not if they parrot each other using identical signals (Tauzin and Gergely, 2019). The authors argue that, from an information theoretical perspective, there are two elementary building blocks of communication: (1) there are two agents taking turns exchanging signals, and (2) the signals vary in an optimal way, with a high-but-imperfect level of contingency. The exchanged signals need to be similar enough to form a correspondence yet different enough to carry added information value (Tauzin and Gergely, 2021). At 13 months of age, infants assume a flatfish to update its falsely held belief about the location of an object only after an optimally variable signal exchange with another flatfish (Tauzin and Gergely, 2018). Communication, as information transmission, may directly modify mental state attribution through signal variation, irrespective of social-pragmatic inferences.

According to a recent study, humans may be sensitive to the information transmission value of language already at birth. When presented with grammatically structured ABB pseudowords, but now as an exchange between a female and a male voice, newborns showed increased activity near Broca’s area when the pseudowords were different tokens (female: kamumu; male: dekiki) compared to when they were identical (male: bulili; female: bulili) (Forgács et al., 2022a). These findings demonstrate, first, that neonates can identify the possibility of information transmission in communicative interactions, even when they may have no idea about its contents. Notably, they can do so even when they are not participants in the interaction and without relying on social cognition. If turn-taking were interesting in and of itself, there should have been no difference between the identical and variable signal exchanges. Second, the activation of Broca’s area suggests that it is the language-processing region of the brain that responds to the possibility of information transmission. Processing the potential for information transmission may be a core feature of human language.

The sensitivity to information value is independent of any particular semantic content or the structure of social interactions. In contrast, it is markedly missing from pragmatic models of human communication (Sperber and Wilson, 1986; Grice, 1989; Tomasello, 2008; Goodman and Frank, 2016). These models assume that information is transmitted as a code and then enriched and inferentially unpacked by social cognition. Yet, humans seem to be sensitive to information transmission even without knowing any code. Humans at birth appear to possess a structured representational template of an informative intention embedded within a communicative intention, allowing them to identify communicative intentions even when the embedded representational slot for the informative intention remains empty and even in the absence of social cues directed toward them. Of course, the above study did not provide direct evidence of a second-order representation or the recognition of a communicative intention or the relationship between the two; thus, these interpretations remain just as hypothetical as they are for infants (Csibra, 2010). Nevertheless, the underlying information estimation mechanism may be a third route for identifying communicative intentions alongside joint attention and ostension. Note that the ostensive cue of turn-taking is based on tracking proximal contingencies in interactions infants are part of, while the information estimation route capitalizes on distal contingencies.

Newborns’ sensitivity to information structure implies that humans may assume the existence of a code with content that can be sent and received, which may be just as important to language acquisition and processing as syntax and social cognition. This possibility is in sharp contrast with both Chomsky’s and Tomasello’s proposals. Chomsky suggests that syntax is the core feature of cognition in the form of recursion (Chomsky, 1965; Hauser et al., 2002) or merge (Berwick and Chomsky, 2011). However, its externalization, spoken language and communication are of no particular interest. Tomasello’s (2008) work implies that information can be transmitted communicatively only once the cognitive infrastructure for shared intentionality emerges. The notion that information transmission may be identified based on signal variability hints that humans may be able to enter the suspended space of human communication beyond the here and now, right from the very beginning of life—and language acquisition. Information may be identified without awareness of any form, content, or social context. But how may the actual meaning of language be figured out?

The idea that meaning emerges by establishing word-to-world mappings (Waxman and Lidz, 2007) via linking objects to sounds can be traced back at least to John Locke (Locke, 1975). In fact, it may be a unique feat of our species that a single system, rather than two separate ones, handles both conceptual representations and communication (Miller, 1990). The way these connections are established is still debated, however. The classical view of associations (Hume, 1978; Sloutsky et al., 2017), a form of statistical learning (Smith and Yu, 2008), has been seriously questioned on the grounds of social-pragmatic cognition (Tomasello, 2003; Bohn and Frank, 2019) and by placing intentions at the center stage (Macnamara, 1972; Bloom, 2000).

One outstanding challenge in explaining word learning is the question of referentiality. Referentiality is the idea that words single out and point to things in the world. However, they do so not at the level of individuals—and based on associations—but at the level of kinds (Waxman and Gelman, 2009). This definition is a minimalist one because proper names pick out individuals, but it suffices for most words. According to a series of well-crafted studies, when objects are labeled consistently, with pseudowords rather than tones, 3-month-olds form categories based on sets of objects and generalize membership to previously unseen novel members (Perszyk and Waxman, 2018). On the other side of the same coin, words refer also in the sense that they pick out objects. It has been shown that 4-month-olds follow the gaze direction of an actor faster to locate an object if the actor utters a pseudoword beforehand—backward speech, no vocalization, or looking at the infant instead of the side of the screen where the object is to appear do not do the trick (Marno et al., 2015). These findings show that very young infants can link linguistic signals to conceptual categories and expect these signals to indicate objects.

The first word infants seem to grasp is their own name, at least by 5 months of age (Grossmann et al., 2010; Parise et al., 2010). They do not take long to have at least some understanding of at least some—food-related and body-part—words by 6 months of age (Bergelson and Swingley, 2012, 2015; Tincoff and Jusczyk, 2012). Even these first words are organized in a semantically structured manner (Bergelson and Aslin, 2017), although word frequency and cross-linguistic differences may play a role here (Kartushina and Mayor, 2019; Steil et al., 2021). These findings refuted the long-held idea that during the first year of life, infants primarily learn the phonology of their native language(s) and that word learning proper begins only around their first birthday (Bloom, 2000; Kuhl, 2011).

Word comprehension undergoes qualitative changes during the first year, nevertheless. An electrophysiological indicator of semantic processing, the so-called N400 event-related potential (ERP) (Kutas and Hillyard, 1983; Kutas and Federmeier, 2011), can be elicited in infants by mislabeling objects (Friedrich and Friederici, 2004; Parise and Csibra, 2012). It appears as early as 6 months of age but only during the encoding phase of novel object-label pairings; a day later, infants show only a so-called N200-N500 phonological familiarity effect (Friedrich and Friederici, 2011). These results reveal that word forms are processed and semantic memory structures are in place but function at a limited capacity in 6-month-olds. Even at 9 months of age, the semantic system requires some support to produce an N400, such as words being produced by the infants’ caregiver instead of by an experimenter (Parise and Csibra, 2012) or infants being familiarized with word labels in the lab (Junge et al., 2012). Only the top third high word producers of 12-month-olds exhibit the N400, and it can be reliably evoked only in 14-month-olds (Friedrich and Friederici, 2005, 2008; Forgács et al., 2019). A turning point in word learning at 14 months of age is underscored by the dramatic increase in infants’ performance in Bergelson and Swingley’s (2012) data as well. Werker and colleagues also demonstrated that only 14-month-olds, but not 12-month-olds, can link objects with labels during habituation training (Werker et al., 1998). A boost in the acquisition of abstract words has also been reported in this age group (Bergelson and Swingley, 2013), as well as a more sophisticated understanding of common ground (Moll et al., 2008). These shifts occur right before the onset of the supposed vocabulary spurt, an intense, albeit debated, expansion of the mental lexicon (Bloom, 2000; McMurray, 2007).

To expand their vocabulary, kids are thought to employ a number of dedicated learning strategies—not simply general inductive mechanisms. They rely on word learning constraints (Markman, 1990), such as the whole-object assumption, the taxonomic assumption—from 18 months of age (Markman and Hutchinson, 1984)—and the mutual exclusivity assumption—from as early as 12 months of age (Pomiechowska et al., 2021). They also utilize semantic (Pinker, 1984) and syntactic bootstrapping mechanisms (Brown, 1958; Gleitman, 1990), whereby they infer the meaning of words based on the meaning of the surrounding words in the former and by their syntactic role in sentences in the latter case. Taken together, the semantic system, which is thought to store the meaning of words, seems to be operational from 6 months of age and fully functional by 14 months of age. Word learning is thought to be aided by social cognition, which is thought to be external to the semantic system and pertaining mostly to pragmatic interpretative mechanisms.

Just as with the diverse use of the term “communicative intentions,” there is a continuum among researchers who advocate for the role of mentalization in acquiring the meaning of words and those arguing against it. Those who believe that mentalization is crucial on the road toward linguistic meaning—beyond the recognition of communicative intentions—mostly aim to account for referent resolution (Bloom, 2000; Tomasello, 2008). Tomasello’s (2008) line of reasoning practically seeks to resolve referential ambiguity: recursive mind reading is necessary for appreciating shared goals, which creates joint attention, giving rise to common ground, which in turn allows for identifying the content of pointing, pantomiming, or words. Even those who do not explicitly argue for mentalization in referent resolution rely on some form of attention-guiding mechanism (Baldwin, 1991). Ostensive cues play a very similar role when they serve to establish a cultural learning interaction and, thereby, a unique interpretative context by guiding attention toward objects (for nouns), actions (for verbs), or functions/properties (for adjectives) (Csibra and Gergely, 2009). However, attention tracking may not be a good substitute for mentalization, as it still requires an attribution (Elekes and Király, 2021). One important motivation for leaving out mentalization from language acquisition has been uncertainty about whether ToM is available before 4 years of age. Pragmatics may have seemed a safe place to introduce mentalization in language acquisition because it fitted well with an unspoken, linear developmental order and the sequential conceptions of online language processing inherited from the serial comprehension models of Grice and Chomsky.

When the question of ToM was first raised in cognitive science (Premack and Woodruff, 1978), it soon became a tool to explain autism spectrum disorder (ASD) (Baron-Cohen et al., 1985; Frith and Happé, 1994). Autism had previously been treated mainly as a language deficit, but the new argument was that ASD children are unable to learn to use language in a socially appropriate manner because of a lack of a well-functioning ToM module and concomitant reduction in social motivations that curtail the necessary linguistic input. In an interesting twist, this idea was reversed while researchers scrambled to explain the classic explicit ToM tasks such as the Sally-Anne or Maxi task (Wimmer and Perner, 1983). The argument shifted to the idea that it was language development that enabled ToM (de Villiers and de Villiers, 2000), although the possibility of bidirectional influences was also offered (de Villiers, 2007). Some proposed the necessity of semantic development: as conceptual enrichment unfolds hand-in-hand with word learning (Gopnik and Meltzoff, 1998), ToM becomes available through learning mental words such as “think” or “believe” (Olson, 1988; Brooks and Meltzoff, 2015). Others emphasized that the emergence of ToM depends on grammatical structures (de Villiers, 1998; de Villiers and Pyers, 2002; Hale and Tager-Flusberg, 2003). Paralleling the finding that the acquisition of mental words is aided by complement clauses (“thinking or believing that”) (Papafragou et al., 2007), mental state attribution is made possible by learning the syntactic structure for embedding propositions into propositions, in a meta-representational format (“Maxi thinks that »the chocolate is in the cupboard«”). Again, others have argued for the role of pragmatics (Harris et al., 2005; Frank, 2018; Rubio-Fernandez, 2021). A meta-analysis found that language indeed exerts a considerable influence on ToM: syntax and semantics, alongside receptive vocabulary size, memory for complements, and general language ability, were all positively associated with it (Milligan et al., 2007).

This direction of thinking has taken for granted, however, that ToM becomes available only once kids are able to pass explicit ToM tasks around 4 years of age (Wimmer and Perner, 1983). In such paradigms—the Maxi, the Sally-Anne, or the Smarties task (Perner et al., 1987)—children are explicitly asked about the mental contents of social partners (e.g., “What does Sally think, where are her marbles?”). Perhaps it is no wonder that language competence and ToM abilities have consistently been found to be interrelated.

When it emerged that preverbal infants exhibited ToM abilities (Scott and Baillargeon, 2017) as early as 6–8 months of age (Kovács et al., 2010; Southgate and Vernetti, 2014; Kampis et al., 2015), the idea that various language abilities lay the foundations for ToM was seriously challenged. Explicit tasks may not be tapping into mentalization per se but could instead run into some communicational-pragmatic burden (Helming et al., 2014). The implicit ToM results have been swiftly questioned (Ruffman, 2014) on methodological grounds (Poulin-Dubois et al., 2018), or they were explained away, either entirely (Heyes, 2014), or by suggesting that infant mentalization is inferior to that of adults. It was suggested to be ape-like and not suitable for coordinating perspectives (Tomasello, 2018) or that it is perceptual and “low-level,” restricted to some sort of object tracking and physical perspective-taking system (Apperly and Butterfill, 2009; Low et al., 2016).

Nevertheless, a growing body of findings is proving to be increasingly difficult to explain without assuming adult-like meta-representational ToM in infancy. Observations that false beliefs can be ascribed to social partners without knowing their actual mental content (Kovács et al., 2021; Kampis and Kovács, 2022) suggest that infants attribute structured belief files (Kovács, 2016). Moreover, the ToM of 14-month-olds is capable of handling semantic representations that are in the appropriate “high-level” representational format for beliefs proper (Forgács et al., 2019, 2020). Based on these findings, it is well possible that ToM contributes to or enables language development rather than the other way around.

Recent findings on the so-called social N400 effect have profoundly challenged the received knowledge on the neurocognitive organization of language processing and its relation to social cognition (Rueschemeyer et al., 2015; Westley et al., 2017; Jouravlev et al., 2019; Hinchcliffe et al., 2020). When participants were required to track the comprehension of a confederate while reading semantically incongruous sentences together (“The boy had gills”), they exhibited an N400. Surprisingly, this occurred even when they heard context sentences beforehand (“In the boy’s dream, he could breathe under water”), which should have attenuated the N400 by providing interpretative context. The intriguing finding is that a social effect, which should have engaged pragmatic mechanisms, elicited a semantic response.

In a paradigm designed to directly manipulate the belief state of a communicative partner during language comprehension, even 14-month-olds produced a social N400 in response to the miscomprehension of a social partner (Forgács et al., 2019). In a puppet theater experiment, infants were presented with familiar objects that were always correctly labeled from their perspective but sometimes incorrectly labeled from the perspective of an observer. The observer, seated on the other side of the stage, had visual access to objects only when an occluder was lowered. First, an object was placed in front of infants (e.g., a cup), which was revealed to the observer as well; however, when the occluder moved back up, the observer turned away, and the first object was replaced by a second one (e.g., a car), unbeknownst to the observer. When the observer turned back, the second object was labeled (“car”), which was congruent for infants but incongruent with the false belief of the observer. Despite experiencing no semantic processing demands, infants produced an N400 (Forgács et al., 2019, 2020). Thus, the ERP indicator of language comprehension responded to a mentalistic manipulation, not simply a social one. These findings are relevant for ToM research because they demonstrate that false beliefs can be attributed as semantic content, which is compatible with a propositional meta-representational format. Conversely, ToM may be at full capacity already in infancy (Leslie, 1994). The findings are also remarkable from the perspective of experimental pragmatics because they show that not simply social cognition but specifically mentalization can impact language comprehension, not only at the level of pragmatic-inferential mechanisms but also at the level of semantic processing.

The social N400 appears to have two constituents: the false belief N400 and the social presence N400 (Forgács et al., 2022b). When presented with congruent and incongruent object-labeling events, adults showed an enhanced N400 response not only to incongruity but also to the mere presence of another person, in contrast to when they were alone. The typical N400 seems to be best explained as a semantic memory retrieval effort (Kutas and Federmeier, 2011; Brouwer et al., 2012; Urbach et al., 2020), which is evoked at all times but reduced when semantic predictions are met. Thus, the social presence effect can be understood as a lesser reduction of the N400 when someone is simply present. This may be due to a broader range of semantic elements remaining activated, which is likely to enhance potentially ensuing social interactions. The false belief N400 can be elicited in adults as well, over and above the social presence N400. In the false belief N400 paradigm, an observer is always present. However, an additional N400 effect is evoked only if participants are explicitly instructed to follow the comprehension of the other person (Forgács et al., 2022b)—just as in the information asymmetry social N400 experiments (Jouravlev et al., 2019). In sum, semantic processing seems to involve two mentalistic components: a spontaneous one, the social presence N400, and a strategic one apparent only following instructions, the false belief N400.

The social presence N400 is evident already at 14 months of age, right at the developmental onset of the N400. Nonetheless, the effect appears only in response to incongruent labels, not to congruent ones (Forgács et al., submitted). It seems that infants ration their limited cognitive capacities to engage in semantic mentalization only when incongruent labels potentially incur divergent perspectives and false beliefs rather than when congruent labels require the attribution of true beliefs. In such cases, it may be sufficient to assume a shared, normative belief (Király et al., 2018). It may be argued that no attribution of beliefs is necessary for the social presence effect and that attributing perception may suffice. This may be true, but it is based on the assumption that the N400 is an indicator of perceptual processing. While it has been argued that the language system is fundamentally a reflex-like perceptual system (Fodor, 1983), the more broadly accepted view is that semantic mechanisms pertain to the conceptual system in some way. Additionally, the attribution of perception could also be viewed as a form of mentalization, as it involves ascribing an experience as a mental state.

The mentalistic social N400 is a riddle for pragmatic theories. For Neo-Griceans, social cognition, let alone mentalization, should not influence semantic mechanisms—only pragmatic inferences. For Post-Griceans, since inferential mechanisms may already be involved in developing explicatures, the impact of social cognition on semantic processing may not be unexpected. However, mentalization should be an input to the pragmatic module (together with the logical frame) and should not influence the initial lexical retrieval. The thought-provoking aspect of the mentalistic N400 is that none of these experiments were supposed to elicit an N400, as they did not pose any semantic processing demands per se. Instead, they well could have evoked ERPs associated either with ToM, including parietal or frontal responses (Liu et al., 2009; McCleery et al., 2011), or with pragmatics and contextual processing, such as the P600 (e.g., Van Berkum, 2009). Mentalization apparently impacted language comprehension not on a pragmatic but on a semantic level, which was not predicted by any pragmatic theories.

At a minimum, these results suggest that the ToM network may coordinate very closely with the language network (Paunov et al., 2019). The ToM network is a bilateral system, perhaps slightly more right-lateralized, with centers at the temporoparietal junction (TPJ) and the middle prefrontal cortex (mPFC) (Frith and Frith, 2003; Saxe and Kanwisher, 2003). It is part of a broader network of social cognition (Schurz et al., 2020). The language network is a more left-lateralized system of temporal and frontal regions (Binder et al., 1997). It has been argued that these two networks work independently (Shain et al., 2023), despite some apparent overlaps, and have stronger connections within themselves than between each other (Paunov et al., 2019). However, there are concerns with explaining the social N400 based on an interaction between the two networks. It is a lexical input that should trigger the ToM network (instead of the language processor), which in turn should activate the semantic system soon enough to produce an N400, yet not for semantic retrieval but to represent the mental state of a social partner. Thus, not only was an N400 not expected in the social N400 experiments (in the absence of semantic processing demands), and no other ERPs were observed (to indicate the activation of the ToM network), but specifically an ERP associated with the semantic system responded to the language comprehension and miscomprehension of a social partner.

It is true that belief attribution was accompanied by frontal effects in infants’ false belief N400 experiments (Forgács et al., 2019, 2020). However, these effects were inconsistent between French and Hungarian infants, and frontal regions may be engaged during false belief processing for a variety of reasons beyond belief computations, from inhibitory control through response selection to resolving conflicting representations (Southgate, 2020). It is also true that the infant social presence study (Forgács et al., submitted) involved no false beliefs, only the tracking of another person’s experience of a semantic incongruity, which nevertheless still seems to qualify as at least some form of belief attribution. The overall pattern of results suggests that the semantic system is engaged in processing ToM in the mentalistic N400 experiments. Such an interpretation does not preclude the possibility that the ToM and language networks are separate systems that work closely together (Paunov et al., 2019; Shain et al., 2023; Fedorenko et al., 2024). The semantic system could work mentalistically without subserving other ToM functions.