3.1 The problem of emergence and the necessity for foundational concepts

Natural cognition is the ability to acquire information from the environment, retain it, and use it for the purpose of adaptive behavior. Mental representation is argued to be the central process making this possible. We should ask, then, what it means for mental representation to be in service of information acquisition, retention and behavior guidance. At the most general level, it means that the organism understands something about the represented reality: representation should give a clue to the organism about what it can expect to happen, what it can do, and what it should do considering its goals and current states.

Traditionally, representation is argued to achieve the above in virtue of its correspondence to the represented reality and systematic relationship to other mental representations (Cain 2013; Fodor 1975, 1983; Pylyshyn 1984; Smith and Medin 1981). This has classically been viewed as a causal, informational, and ruleful relationship between the representation, the represented, and other representations of the representational system. To capture that, mental representations have been assumed to have semantic properties much like logical terms, propositions and propositional attitudes do; they refer to reality via the meaning encoded in them and to each other via their syntactic properties. Consequently, much of cognitive psychology today views cognition in terms of manipulation of semantic symbols in this sense. Environmental information is said to enter the system through the senses, after which it is transduced into a symbolic or representational format that is independent from the specific sense modality through which the information was acquired. That is when sensory representations become concepts. Once in the amodal format, information is processed in a way similar to logical inferences. Finally, results of this computation are transduced back into the embodied format of motor directions.

There is a problem, however, in accounting for the emergence of the amodal semantic content—concepts must be already in place for the sensory information to be transduced into them and back into motor information (Bickhard and Terveen 1995). In fact, emergence of semantic content is impossible; there is no way for the organism with such a code to actually know what it is about (for the criticism see Bickhard 2009b; Bickhard and Terveen 1995), and for that reason any model built around semantic representation is forced to be foundational—to necessarily assume a set of conceptual primitives from which development can start. Below, we point out how the traditional ToM approaches share this problem.

3.2 Nativism

Researchers with nativist views generally follow the standard modularist model of cognition (Carruthers 2013, 2015, 2016; Helming et al. 2016, 2014; Leslie 1994; Scholl and Leslie 1999). As with similar accounts of other cognitive abilities (e.g. Lightfoot 1989; Pinker 2014/1994; Wynn 1992), the story here goes as follows: There is an innately specified cognitive mechanism, or module, dedicated to a specific domain, in our case—to mindreading. This mechanism is independent from and insensitive to virtually any extra-organismic factors and develops according to a biologically predetermined schedule. The innate, encoded information that it contains consists of basic inborn concepts (e.g. BELIEF, DESIRE, SEE, and PRETENSE) and heuristics (e.g. “seeing leads to believing”) that enable the child to pick out relevant stimuli, cognize it, and draw basic conclusions (in an unconscious modular way, that is). These are then fed in some form to the central system, adding an aspect of another’s agency to the child’s perceived reality.

Nativists take this to be the cognitive process that underwrites 15-month-olds’ performance on the spontaneous–response FBT; infants pass it because they have this basic inborn theory which makes them expect the false-belief congruent scenario. As Westra and Carruthers (2017) offer, this innate theory is open for learning; it is claimed that it gets enriched with more complex concepts throughout development, or that its harmonization with the rest of the cognitive system can improve, domain-general processes putting the mindreading module to work for their purposes (Carruthers 2015).

Accordingly, any differences in performance on the elicited-response FBT and related tests across populations—e.g. different cultures, parenting practices, or linguistic inputs—are explained away by factors other than an actual lack of mentalistic understanding of the mind (this is most clearly argued in Westra and Carruthers 2017, but see also Helming et al. 2014, 2016). Nativists take two directions here. One is the claim that the tasks in question make demands for more complex concepts and heuristics than the basic ones (culturally embedded ones not yet acquired). The other is the recourse to the competence-performance chasm. The latter path is naturally necessary for FBTs. Children are said to understand false beliefs innately and the varying performance is due to (a) misconstruals of implicatures of the test questions, (b) lack of adequate vocabulary in the language they are growing up with, or (c) undeveloped executive functions or general-processing resources.

Nativism is thus openly foundationalist, and seeks a solution to the impossibility of emergence in the claim that the representational primitives are innate. This does not help, however, as the idea of inborn semantic content cannot be defended either. We discuss this in Sect. 4.

3.3 Rational constructivism

The alternative, rationalist constructivist idea has been that children construct a theory of mind much like a scientist would (Gopnik et al. 1999; Gopnik and Wellman 1992). This faced a challenge in light of the spontaneous–response FBT results with infants. In order to explain the gap between infant and preschooler mindreading, and not to commit to innate belief understanding, constructivists were forced to show that infant mindreading could be explained by simpler concepts than those of a belief-desire theory of mind.³

Wellman (2014) argues that infant data can be explained with only a desire-awareness conceptual framework, which with time is built on and becomes a proper theory of mind with belief understanding. The way the original concepts are modified and enriched is modeled with the use of Bayesian hierarchical networks, and the child’s conceptual development is viewed as a theory revision process, the child being a “little scientist.” Thus, in contrast to the nativist view, constructivism of this kind views ToM development as utilizing domain-general resources, and conceives of ToM as a somewhat real theory in the mind of the child, not a modular mechanism operating according to the same principle as a theory of mind. It does, however, start with a set of foundational mental representations (Wellman 2014, p. 197), and necessarily so. For a theory-based development, the initial states must be concepts understood as encodings; theory by its nature just has to start with initial concepts that enable hypothesis formation and further theoretical change. It is fairly unquestionable that any account has to start with something, to take something for granted. Rational constructivism, however, forces us to assume that these initial states are conceptual in nature, and this is clearly a case of foundationalism.

3.4 Two systems

The two-systems view finds a middle ground in between nativism and constructivism. Although there have been a number of different proposals that advance two systems (e.g. De Bruin and Newen 2012), Apperly and Butterfill’s account is most usually associated with the term (Apperly 2012a, b; Apperly and Butterfill 2009; Low et al. 2016).

The basic assumption of the two-system view is that children pass the spontaneous–response FBT because they are in possession of mindreading system 1, which is a limited foundational theory of mind: a belief-tracking mechanism (Apperly 2012a; Apperly and Butterfill 2009; Butterfill and Apperly 2013).

Imputing spontaneous–response FBT results to the workings of system 1, two-system proponents claim that passing the elicited-response FBT requires a much more effortful and explicit way of thinking about other minds—what they call system 2—that children before preschool cannot really use. Although Apperly and Butterfill do not really address the development of system 2, it would make sense that it is constructed or somehow acquired just as other explicit ways of thinking about reality, and so its development might be influenced by environmental factors, which would explain the cross-cultural data we have presented before.

Further, it does not change matters much that they view their minimal theory ascribed to infants as only a construct at the computational level of explanation (cf. Marr 1982/2010). The computational level explanation still imposes significant constraints on possible implementations. Apperly and Butterfill’s framing of the computational problem in terms of a minimal theory of mind still poses a foundationalist problem: The computational-level theory has to be implemented in a way that merits being called “a theory”—that is, there have to be implementational equivalents of computational-level processes that relate to each other in the prescribed, theory-like way (compare to the discussion on tacit knowledge in Davies and Stone 2001; Fenici’s 2013, application of Davies and Stone’s ideas).

The computational problem they describe still consists in ascribing states to observed agents. Thus, system 1, though minimalistic, still presupposes concepts of object and agent and registration, whose contents are similarly left unexplained developmentally. As far as system 2 is concerned, here they run into all the problems that rational constructivism does—they openly draw an analogy to Fodor’s central system (Apperly and Butterfill 2009, p. 956), where explicit theories reside, when explaining why system-2 theory of mind should be effortful but flexible.

Notably, it can be argued that the two systems account does not aspire to explain the development of the representational states in question, and thus the charge of foundationalism does not apply to it. Still, the problem generally has to be solved to give an exhaustive account of socio-cognitive development, and there does not seem to be much potential in the two-systems account to do so, as it frames the problem in cognitivist terms.

The general insight of the two-systems theory, however, is consistent with the interpretation offered by our action-based account. Following the action-based principle, we too arrive at the closely similar conclusion that there is an important chasm between processes underlying competent social interaction, and explicit theorizing about other’s mental life. How we reach that interpretation is however importantly different, as later parts of the paper will show.

5.1 The action-based principle

Here is what follows from the criticism of the ToM approaches we offered above, which we contrast with what the action-based principle claims: (1) For the criticized theories, every action or cognitive skill is underwritten by disembodied representational competence of sorts; for action-based models, action can precede representational mental content.⁷ (2) For the criticized theories, the development of social competence must start with an inborn base of amodal representation; for action-based models, it does not have to—representation can emerge from non-representational phenomena.

It is instrumental to stress at this point that the action-based representation (which we take from interactivism) and the standard idea of representation as amodal encoding differ fundamentally.⁸ First, what we are interested in when modeling representation is not solving the metaphysical problem of reference (see, e.g. Quine 1960/2013, pp. 23–72), but only proposing such an idea of representation which is a viable way in which real organisms can represent reality. An action-based representation does not represent on the basis of reference or correspondence; it is not a disembodied symbol with a semantic stand-in for what is being represented. It does, however, have the necessary properties of representation—intentionality and truth value. And most importantly, it has them in virtue of processes which are consistent with developmental reality and which allow for representation to emerge in ontogeny (and phylogeny) from non-representational phenomena. How it does so, and how it achieves intentionality and truth value should become clear in our exposition of the action-based principle below. This is drawn from interactivism (Bickhard 2007, 2009a, b, 2010; Bickhard and Terveen 1995).

An internal state S is a detection of an external state S*. The organism does not know that, but merely experiences the internal state as “this state.” Being in internal state S, the organism undertakes action A (from among others; let us assume that for newborns actions can be random at first for the sake of the illustration), which results in the external state S* changing to Y*. The organism’s physical organization is such that Y* evokes another internal state, state Y. This way, on the basis of non-representational detective properties, the organism can create an action-internal state contingency pattern that while in state S, action A leads to state Y. This provides the germ for normativity—the organism will now (implicitly) know that state S is not only just “this state” but also such a state that can lead to state Y via action A. Thus, the organism functionally predicates something about the current situation; and it does so in virtue of the action-internal state contingency it has the knowledge of, without the need to refer to the outside world at all. Moreover, the predication can be false or true, and the truth value can be potentially known to the organism—all it takes is to engage in action A to find out whether it is possible to go to state Y from state S. If it was not, then the previous situation was not the situation that should have produced state S (no external state S*) or state S needs some other states co-occurring in order to afford going to Y, which were only accidentally present in the previous interactions that went from S to Y via action A. The organism can accordingly update its functional predications after failing to achieve the expected result of its action.

Now, there may be processes in more complex organisms whose main function is to probe the external reality in the way presented above. They can serve to keep track of what interactions are possible in the given situation so that the organism can be a competent agent that can choose from an array of affordable actions in light of its current goals. Bickhard refers to these kinds of representational phenomena as apperception. They are much more like the classic idea of representation as their function is mainly not to directly indicate the possibility of a given action, but only to predicate something about the current situation. This predication can then be the basis for many other interactions and constitutes what Bickhard calls the situation knowledge. Thus, apperceptive processes have the function of representation for other interactive processes that rely on them in order to guide adaptive behavior. In fact, any representation of a possible interaction can serve two functions—to cue the organism to engage in the interaction, or to provide information about the situation for another interaction representation (something like knowing that you can order a taxi anytime at a party makes you entertain staying after the last bus home leaves). Moreover, there are most likely many levels of recursive interactive systems within the architecture of the human mind, in which one system interacts with interaction potentials (representations) of a lower one, enabling explicit thought about its properties (Bickhard 1998; Campbell and Bickhard 1986). However, it is the function of representation for the organism’s processes that makes it representation, not some amodal, symbolic format it is coded in.

Importantly, the action-based representation offered by interactivism is inherently embodied and situated. The content is constructed through interaction and is grounded in the modalities of the experiences. (Ap)Perceptual processes constituted by sensorimotor contingencies (SMCs) may be grounded in one particular modality (O’Regan and Noë 2001), but the situation knowledge they create, and on which higher-order interactive representations rely, will span all the modalities. Representing the car, for instance, will base off SMCs grounded in past explorations of how cars look, smell, feel, what sound they produce, and for some maybe, what they taste like. The interactive representation of car will consist in connecting the expectations of all these modalities under a common contingency—if I hear a car, visual contingencies associated with it activate too and I expect to see it when I turn my head. The central point is that the content of action-based representation is modal by definition as it is past experience that constitutes it.

Consider a simple example (this is just to make a point, not an empirical claim). The infant experiences a state of hunger and starts crying (crying action could occur randomly, but it is plausible that such a simple state-action coupling could form prenatally; note, however, that the infant does not know anything about why she acts this way). Crying has its impact on the environment such that the mother comes and starts feeding the baby. As a result, the state of hunger changes to the state of satisfaction. Thus, the infant functionally and implicitly comes to know that when hungry, crying leads to satisfaction, and she has some (again, functional and implicit) idea what the internal state of hunger “means”. This knowledge—of hunger-crying-satisfaction contingency—does not have to be innate as the information about it is reliably present in the environment of growth; hence, evolutionary selection was more likely to predispose the child to quickly learn it in the above way rather than prewire it whole.

If the mother is not in the room, however, then crying will not have its effect.⁹ Making sure that the mother is in the room will therefore be an apperceptual process in service of hunger satisfaction via crying. Perception is an apperceptual process par excellence, and we have an action-based account of perception nicely worked out by O’Regan and Noë (2001).¹⁰ This way, through apperceptual processes, the child keeps track of whether the given situation really is a situation in which crying will lead to feeding and satisfaction of hunger.¹¹

The way that language changes the interactive context needs to be noted as we refer to it in our account of the three groups of empirical findings. Linguistic interaction will build on non-linguistic interaction patterns and words will come to represent interaction possibilities they have been used in. Consider a game in which a linguistic utterance is a part of—a simple naming game that mothers play with children many times, where when presented with a toy, one has to say its name. Grounded in such an interactive pattern, the word Zebra, for instance, can come to represent the toy as it was used in the game upon the presentation of the toy Zebra (feedback for failures in naming it so could have been provided in the form of undesirable interaction on the mother’s part—negative facial expressions or continuation to hold the toy instead of progressing to the next one). Linguistic units grounded in such a way can be naturally uttered in any situation, and the child will learn this when her mother uses the same expression in a different context, which evokes the interaction potentials in the child’s mind grounded in the past use of that word. Something similar will take place when the child herself uses it in a different context and observes its impact on the external, social world. Importantly, further contexts of use can be themselves linguistic, which is possibly how abstract meanings emerge.¹²

There is much more to be said about language in an action-based framework (Bickhard 2007). However, for the present purposes, we want to point out that some socio-cognitive abilities could require language to develop, while some would not. There are some social competences which will be greatly improved by associating them with linguistic interaction, some that are embedded in linguistic interaction entirely, and there are such that do not gain much from it. For instance, mere physical sequential social interaction of changing the diaper would not gain anything by adding linguistic components to it, whereas going for a walk would (linguistic structuring of activities outdoors makes them both safer and more interesting for the child). And it goes without saying that kinds of interactive competences that are entirely embedded in language—such as being able to hold a conversation—would need previous linguistic experience in order to exist at all.

It should be clear after this exposition that an action-based perspective solves the problem of foundationalism and is consistent with the multiply contingent nature of developmental phenomena. In fact, many of the aspects of the model mentioned above—such as emergence of representation from non-representational phenomena, naturalized normativity, system-detectable error, or the possibility of multiple, interrelated but qualitatively different representational processes within the organism—are simply absent in the framework upon which the traditional ToM accounts are based. As such, an action-based perspective offers a much more comprehensive alternative to modeling socio-cognitive development that can replace or potentially complement the standard models.

5.2 Constraints on experience in an action-based framework

With the above model of cognitive development, the kind of interactive experience the child engages in will determine the content of its cognitive structure. It is possible that some internal state-action contingencies are already present in newborns (think of all kinds of reflexes), but it does not seem likely that they come with prewired complex interaction representations of theory of mind. A more viable thesis is that the macroarchitecture of the nervous system predisposes the child to certain kinds of experience, and it is these experiences that form its microstructure, furnishing the mind with representational contents (understood pragmatically, not semantically). Natural selection has happened in the environment where certain elements were reliably present and so whatever developmental process it has selected will implicitly presuppose their presence in ontogeny. In other words, the phylogenetic heritage determines what can be possibly experienced by the child, but it is the context of growth and experience therein that determine what will be experienced (cf. Nelson 2007, p. 249).

The kind of interactive experiences available to the child will naturally differ depending on the time and place of development. Nelson’s (2007, p. 19) model of constraints on cognitive development captures this fact nicely. She identifies six kinds of constraints: evolved, embodied, ecological, socially embedded, encultured, and that of past experience.¹³ What is experienced at a given time in development is jointly determined by the constraints. The view on development that we get here is therefore necessarily scaffolded—starting from the interactions available to an infant, she actively “acts her way up,” establishing first basic action-based representations (e.g. sensorimotor contingencies), and then using them to engage in new interactive experiences and establish further, more complex, representations (e.g. social competence).¹⁴ The basic representations will be fairly invariant across environmental contexts as they will rely on largely universal patterns of physical interaction with the environment and caretakers; higher representations will be grounded in more complex social and cultural interaction, which will naturally differ greatly across social and cultural contexts.

5.3 Socio-cognitive development in an action-based framework

Coming back to the problem at hand, two constraints on experience are especially relevant for the emergence of socio-cognitive mental structure—non-linguistic social interaction and linguistic social interaction. Below we show why,

Carpendale and Lewis (2004, 2006) stress the role triadic interaction with parents and objects has for cognitive development. Naturally, the kind of interactive experience the child gets while interacting with other people will establish socially embedded representations of interactive potentials: The child will have developed certain expectations of how people behave in a range of situations and how they react to the child’s own actions. Such largely behavioral interaction competence will not, however, involve abstract concepts of minds or beliefs. The action-based representations established will be derived from purely physical interaction with other people, not understanding their minds. As argued by others (e.g. Fenici 2012, 2015; Gallagher 2008), such embodied interactive competence can readily explain early social cognition and spontaneous–response test results.

Things change when language enters the developmental system. Conversational situations constitute a new kind of interactive context that enables representation of unobservable minds. Carpendale and Lewis claim that language is acquired by learning patterns of interactions for which “it is appropriate to use a particular term, for example, mental or emotional, or dealing with pain, and so forth” (Carpendale and Lewis 2004, p. 88). Linguistic interaction is part of the external environment and as such provides interactive potentials embedded in language, otherwise unavailable. This fact—that language enables later social cognition based on abstract notions of mind and belief—has also been argued by others on various grounds (e.g. Fenici 2012; Gallagher and Hutto 2008; Hutto 2008; Nelson 2005, 2007). An action-based principle could provide a more detailed model of why this is so. The bottom line is that linguistic interaction potentials can go a long way in explaining how children pass the standard FBT, without the need to posit theoretical knowledge (cf. Gopnik and Wellman 1992).¹⁵

5.4 Empirical data within an action-based framework

It becomes clear that the view is generally consistent with other views that proclaim the embodied nature of early social cognition, and linguistic nature of later mindreading (e.g. Andrews 2012; Fenici 2012; Fenici and Garofoli 2017; Gallagher and Hutto 2008; Newen 2015; Zawidzki 2013). However, the exact views on the nature of cognitive processes involved espoused by these approaches could differ greatly from the action-based ontogeny we have presented.