Building blocks of social cognition: Mirror, mentalize, share?

Abstract

During the past decade, novel approaches to study social interaction have expanded and questioned long-standing knowledge about how humans understand each other. We aim to portray and reconcile the key psychological processes and neural mechanisms underlying navigation of the social environment. Theoretical accounts mostly revolved around either abstract inferences or embodied simulations, whereas experimental studies mostly focused on theory of mind or mentalizing, empathy, and action imitation. The tension between theories of and experiments on social cognition is systematically revisited to foster new theoretical and empirical studies in the fields. We finally retrace differential impairments in social capacities as a means to re-conceptualize psychopathological disturbance in psychiatry, including schizophrenia, borderline personality, and autism.

Introduction

During recent years, our understanding of the cognitive processes and neurobiological mechanisms underlying human capacities in social dynamics have been continuously challenged and diversified. For instance, many investigators believed the ability to understand the motivation behind others' behavior to emerge around the age of 4. More recent evidence suggests its presence in infants as young as 2 years (for a review, see Scott & Baillargeon, 2017), and certain nonhuman primates could make use of comparable abilities (Krupenye and others 2016). Additionally, the use of the umbrella term “empathy” denoting multiple affective, cognitive, and motivational processes is increasingly criticized as an impediment to scientific progress and communication (Bloom, 2017).

First, we juxtapose the major theoretical accounts of how individuals understand the emotions, beliefs, and intentions of others, which mostly focused either on abstract inferences or automatic simulations of other people's mental states. Second, we then examine the key experimental approaches to the study of social interaction, which essentially referred to operationalisations of theory of mind or mentalizing, empathy, and action imitation. Lastly, we aim at clarifying the possible mapping between these two theoretical accounts and at least three experimental research streams that remains a conundrum in social neuroscience until today.

Successful social interactions require adequate exploration and modeling of other individuals' mental or “inner” states, including their emotions, beliefs, or intentions behind action (Frith and Frith, 1999, Mitchell, 2009, Alcalá-López, 2017). Two main theoretical frameworks have emerged to explain how an adequate understanding of others' behavior may be achieved. On the one hand, the theory theory (TT) proposes that individuals draw on a collection of abstract principles about human behavior, acquired through life experience, that allows to interpret and predict the mental states and behavioral patterns of others (Carruthers, 2009, Gopnik and Wellman, 1994). Such inferences based on abstract emulation enable predictions about what individuals are thinking about others, despite the fact that their mental states can never be directly observed or explicitly confirmed (Gopnik and Wellman, 1992, Perner, 1991, Wellman, 1990). As defined by Saxe (2005), the TT account captures humans' lay theory of psychology constructed “from observation, inference and instruction, and then deployed to predict or explain another person's inference, decision or action”.

On the other hand, the simulation theory (ST) proposes that humans impersonate others and automatically imitate their mental states (Gallese and Goldman, 1998, Goldman, 2000). This neurocognitive reinstantiation of the other's mental content aims to understand what oneself would experience in the other's place (Gallagher, 2001). According to the ST account based on embodied simulation (Gallese & Sinigaglia, 2011), grasping the mental states of others means to “purposely pretend to be in the other's ‘mental shoes’ and use our own mind as a model for the mind of others” (Gallese, 2003). As such, the ST framework rejects the need to assemble abstract models to emulate others' behavior –as proposed by the TT– since humans already have a working model of how it feels to perceive and act in a given environment: one's own inner experience.

In the following, previous behavioral and neuroscience research inspired by TT and ST will be portrayed, as well as more recent theoretical accounts (e.g. Gallagher & Hutto, 2008). We will then discuss both conceptual frameworks with respect to their neural mechanisms.

The TT received extensive empirical support from experimental studies in developmental and comparative psychology focusing on perspective-taking tasks that require deception and false-belief detection (Baron-Cohen et al., 1985, Leslie, 1987, Leslie, 1994, Premack and Woodruff, 1978, Wimmer and Perner, 1983). For instance, Premack and Woodruff (1978) originally referred to the awareness that humans and non-human primates may have of others' mental states as theory of mind (ToM). Frith, Morton, and Leslie (1991) later introduced the term mentalizing to refer to the ability of belief attribution in humans and to include spontaneous and non-inferential capacities, as the term “theory” could lead to the misunderstanding that ToM is a fully developed theoretical account about the behavior and experience of others.

Whether non-human primates can adopt, at least to some extent, the perspective of other conspecifics has been debated at length in the past decades (Call and Tomasello, 2008, Premack and Woodruff, 1978). Tomasello, Call, and Hare (2003) described how previous research during the 1990s led many investigators to believe that such capacity would be a defining feature of the human species. However, more recent findings since the beginning of the century have challenged this view, and they rather entailed a continuous refinement of the comparative difference in ToM capacities in monkeys and humans. For instance, Hare and colleagues designed a series of experiments in which two chimpanzees, one dominant and one subordinate, had to compete for food. By strategically hiding the food in locations to which only one or both chimpanzees had visual access, the authors investigated whether the subordinate was aware of what the dominant could or could not perceive at the moment (Hare and others 2000) or in past situations (Hare and others 2001). This is in line with a recent behavioral study in three different species of apes (Krupenye and others 2016) that used an anticipatory looking measure as a rudimentary proxy to test for false-belief understanding. The behavioral experiments showed that apes could not only infer the goals and intentions of others' (external) actions, but also behaved in alignment with their (internal) mental states that were incongruent with the external reality (i.e., false beliefs). Nevertheless, different species of non-human primates show different degrees of sophistication in their mentalizing abilities. Devaine and colleagues (2017) recently showed that monkeys can learn simple trial-and-error heuristics to solve strategic social interactions, while apes are also capable of more complex forms of mentalizing to account for their own influence on the behavior of others as described by Krupenye and others (2016). These authors argued that a similar evolutionary gap in mentalizing sophistication separates apes from humans (Devaine et al., 2017). Given this apparent progression in mind reading capacities across species, it is becoming increasingly difficult to precisely demarcate the difference between humans and other species in their ability to understand fellow individuals.

Many experimental psychology studies in humans exemplified the capacity to attribute intentions and beliefs to others (i.e., TT) by using experimental paradigms based on the concept of ToM, such as false beliefs tasks (Baron-Cohen and others 1985), advocated by many as a fundamental mechanism underlying social interaction (Carruthers and Smith, 1996, Frith and Frith, 2003). Traditional experimental paradigms prompting participants to indicate the beliefs of others found that ToM emerges around the age of 4 years (Wellman and others 2001). This capacity was thus regarded as a relatively advanced form of social cognition. However, in the past decade novel approaches have shifted the focus away from probing how children answer specific test questions about their reaction to experimental situations (Baillargeon and others 2016). For instance, Buttelmann and others (2009) taught in a classical false-belief task a group of infants younger than 2 years of age how to unlock a pair of boxes with a pin. Afterwards, one of two experimenters would put a toy in one of the boxes and leave the room with the box unlocked. The second experimenter would then put the toy into the other box and lock both boxes. The first experimenter on his return would try to get the toy from the original box in which he had previously put the toy, while the infants were then prompted to help the experimenter. The authors found that the infants succeeded in helping to retrieve the toy by unlocking the box where the toy had been hidden, indicating awareness of the experimenter's false belief. Based on similar experiments, several recent studies found children younger than 2 years of age already behaving according to others' false beliefs (for a review see: Scott & Baillargeon, 2017). Many authors have argued that the diversity in children's mentalizing skills across studies may be driven by preceding differences in the maturation of domain-general cognitive abilities including linguistic and executive performance (Apperly, Samson, & Humphreys, 2009) the developmental trajectory of which may further be culture-dependent (Vogeley & Roepstorff, 2009). This view is congruent with a recent behavioral study that conducted a massive online poll of people's cognitive skills, including their ToM ability (Klindt, Devaine, & Daunizeau, 2017). The results of this study showed that variations in executive skills predicted differences in ToM abilities starting from the emergence of adulthood through the lifespan. These aspects together lend support for the possibility that the maturation of general, basic cognitive abilities likely scaffolds the development of complex mentalizing skills.

At the other pole of the lifespan, there has been an increasing interest in the study of age-related differences in ToM abilities. Despite initial findings of a better performance in ToM tasks in a group of older compared with younger participants (Happé and others 1998), subsequent studies have found contradicting results. For instance, Maylor and others (2002) reported a decline in ToM abilities in healthy aging in a series of behavioral experiments. This study included a replication attempt using the stimulus material and procedures described by Happé and others (1998), as well as an extended experiment that accounted for possible confounds due to age-related performance deficits in general cognitive ability. Said reported decline in ToM abilities is congruent with a neuroimaging study in which older participants performed worse compared with young participants even when explicitly prompted to infer the mental states of others (Moran and others 2012). Moreover, these authors found that such social-cognitive deficit in normal aging was associated with decreased neural activity of the dorsal mPFC during mentalizing tasks. Consistently, Henry and others (2013) recently performed a meta-analysis of 23 behavioral studies that included different types of experimental paradigms, visual and verbal stimuli, as well as distinct presentation modalities. The results of this meta-analysis showed an overall moderate deficit (r = −.41) in ToM performance in older adults, irrespective of task and presentation modalities. Furthermore, such age-related deficits were larger in magnitude than corresponding deficits in matched control tasks. This led the authors to suggest that ToM may be a domain-specific process, which declines with age disregarding perceptual or linguistic capacities.

Taking the behavioral findings to the neural level, the brain imaging literature on the TT framework has also frequently relied on ToM tasks. Such neuroimaging studies have consistently revealed that a set of brain regions (Fig. 1) including the medial prefrontal (mPFC) and posterior cingulate (PCC) cortices, as well as the bilateral temporo-parietal junction (TPJ), robustly increase in neural activity when participants undergo perspective-taking tasks probing ToM performance (Gallagher and Frith, 2003, Saxe and Kanwisher, 2003, Saxe et al., 2004, Vogeley et al., 2001). A virtually identical set of brain regions is, however, also known to increase its activity during the retrieval of autobiographical memory, spatial navigation from a first-person perspective, or prospection into the future (Buckner and Carroll, 2007, Spreng et al., 2009, Vogeley and Fink, 2003, Vogeley et al., 2004). In spite of its involvement in diverse experimental conditions, this set of regions was previously found to decrease in neural activity during many other tasks and was therefore called the default mode network (DMN), mostly active during idling mind sets (Gusnard and Raichle, 2001, Raichle et al., 2001). There is a considerable overlap between the neural correlates of the default mode network, particularly active at resting baseline, and those brain locations exhibiting increased activity during ToM tasks. This has led some authors to speculate that social information processing may be what is processed during resting states (Schilbach et al., 2008, Vogeley, 2017). Being aware of the dangers of “reverse inference” (Poldrack, 2006), we simply restate this previously made association of the default mode network activity and social information processing as a testable hypothesis.

Despite extensive empirical support in favor of TT, some authors have denied that a mechanism dedicated to abstract emulation is a necessary condition to grasp and represent others' subjective experience (Perner & Kühberger, 2005, pp. 174–189). Instead, the ST account proposes that individuals automatically mimic or intuitively impersonate in a covert fashion the behavior of others, even when simply observing them (Fogassi and Ferrari, 2007, Umilta et al., 2001). ST proposes that this reinstantiation of observed behavior provides access to the internal mental state of the other, thus enabling action understanding (Gallese et al., 2004, Keysers and Gazzola, 2007, Uddin et al., 2007). The ST framework for grasping other humans' minds has often been used as a conceptual basis to interpret experimental studies on empathy tasks. Consistently, Preston and De Waal (2002) proposed that witnessing others' social-affective behavior inevitably triggers one's own internal representation of that same behavior. Most researchers probably agree on a working definition of empathy as consciously experiencing an affective mental state that is congruent or very comparable to that of an observed individual (De Vignemont and Singer, 2006, Decety and Chaminade, 2003).

Developmentally, simpler forms of affective sharing were suggested to precede the onset of full-fledged empathy capacities in infants (Singer & Lamm, 2009). Concretely, mimicry and emotional contagion are already present in newborns (Piaget, 1945), before the onset of ToM (for a review, see Meltzoff & Moore, 1989). Dimberg and Öhman (1996) for instance investigated facial expressions using electrophysiological measures to show triggering of corresponding facial gestures (e.g., smiling or frowning) when perceiving others' affective expressions. Such a tendency to automatically reproduce the externally visible manifestations of internal affective states (i.e., mimicry) has been suggested as a possible low-level mechanism, elaborated on by more complex forms of empathy (Hatfield et al., 1993, Singer and Lamm, 2009). In emotional contagion, another proto-form of empathy, an individual synchronizes with and converges to others' affective mental states (Hatfield and others 2009). In contrast to full-fledged empathy, emotional contagion occurs without awareness of the observing individual (Decety and Jackson, 2004, de Waal, 1999), and has been identified in other species, including rodents (for a review, see Meyza, Bartal, Monfils, Panksepp, & Knapska, 2017). Although simple, such proto-form of empathy can turn out to be fundamental for social interactions. For instance, Langford and others (2006) demonstrated both an emotional contagion of pain as well as a social modulation effect on pain behavior in mice. Comparative studies across species are of special interest since they provide direct evidence of candidate “building blocks” of social cognition that may have evolved in isolation to more complex forms of social understanding (Decety & Svetlova, 2012).

Simulation mechanisms independent from affective states have been suggested to underlie other types of social behavior. For instance, the term herding refers to the behavior of an individual when she imitates or mirrors a group, as opposed to acting independently (Baddeley, 2010, Raafat et al., 2009). Such social behavior involves convergent thoughts, beliefs, or goals between the individual and the group or collective to whom she refers to. Nevertheless, herding typically occurs automatically and without awareness, and does not necessarily involve understanding the mirrored behavior (Baddeley, Pillas, Christopoulos, Schultz, & Tobler, 2007). Other studies have focused on social influence processes by which an individual revises and adapts her own beliefs as a consequence of social interactions. In a behavioral study, Lorenz, Rauhut, Schweitzer, and Helbing (2011) asked participants to estimate quantities regarding geographical facts and crime statistics in subsequent estimation trials, and provided some of them with information regarding others' estimates. Their results supported that participants used this information to update their own beliefs, which eventually narrowed the variance between individual estimates. This is congruent with recent studies measuring online behaviors such as voting on content (Muchnik, Aral, & Taylor, 2013) that show a similar tendency to adapt one's own decisions and preferences to comply with those of online friends. Similarly, a promising line of research draws upon Bayesian learning models to study how an individual aligns her subjective attitudes towards those of others. In an experimental paradigm similar to the one used by Lorenz and colleagues (cf. above), Moutoussis, Dolan, and Dayan (2016) asked participants to express their preferences for smaller but immediate rewards compared to larger but delayed rewards. These authors argued that such preferences might represent uncertain beliefs, and that individuals would take into consideration the preferences of others to reduce said uncertainty. Their empirical results were congruent with such view: the more uncertain a participant's preference, the more she adapted to those of others, as a normative Bayesian inference would predict (Moutoussis et al., 2016). In another behavioral study, Devaine and Daunizeau (2017) tested a computational model of how individuals learn about the lazy, impatient, or prudent attitudes of others based on Bayes-optimal information processing. Their results showed that participants behaved in accordance to the predictions of the model in a way that even reproduced well-known social influence effects such as false-consensus and influence biases. Thus, these authors concluded that learning about others' attitudes requires sophisticated forms of mentalizing. The sum of these studies provides evidence that individuals sometimes tend to naturally mirror or simulate the behavior of others, consistent with a ST account for understanding others. However, they can also learn and adapt their own behavior or beliefs about the intentions, preferences, or attitudes of others in a fashion that concurs with a TT account.

On the neural level, empathic state-matching reaction to others' affective behavior has consistently been associated with a brain network including the anterior insula (AI) and anterior mid-cingulate cortex (aMCC). This so-called saliency network is recruited, for instance, both when a participant receives painful stimulation as well as when perceiving others in pain (Decety, 2010, Fan et al., 2011, Lamm et al., 2011, Singer et al., 2004). A majority of studies in the neuroimaging literature on empathy tasks performed such comparison between the neural activity elicited by the observation of others in pain and by experiencing pain oneself (Decety and Lamm, 2006, Singer and Lamm, 2009, Singer and Leiberg, 2009, pp. 971–984). This concurs with the idea that social cues can elicit partial synchronization of neural activity patterns both in the agent and the observer (Adolphs, 2003, Decety and Grèzes, 1999, Gallese, 2003).

We argue for a broader notion of ST going beyond affective sharing. Automatic simulation of affective mental states according to the ST framework has largely focused on empathy and mechanisms of sharing emotion in the social neuroscience literature. However, affect- and emotion-independent mechanisms for simple action observation can be readily viewed as another flavor of internally reenacting others' behavior. There are hence few arguments against a simulation-based account for understanding others, in concordance with the above-mentioned original description of the ST. This is suggested by invasive experimental findings in monkeys that showed existence of neuron populations, so-called mirror neurons, that fire in response to both executing and observing the same goal-directed action (Di Pellegrino et al., 1992, Fogassi et al., 1998, Gallese et al., 1996).

By in-vivo recording single–neuron activity in the ventral premotor cortex of macaque monkeys, Rizzolatti and others (1996) found a subset of neurons that discharged when the monkeys grasped, held, or placed an object, as well as when it was the experimenter who was performing such actions. Consistently, Kohler and others (2002) found in recordings in macaques that this matched firing pattern could not only be evoked by visual, but also auditory stimuli. Further, mirror neurons seem to discharge specifically during goal-directed hand actions (e.g., grasping, tearing, and holding), but not in response to goal-free muscle contractions (Gallese and others 1996). In these experiments, the mere observation of a goal-directed action that is unrelated to mimicry, emotion contagion, or empathy, triggers the neural activity pattern responsible for the execution of that same action in the observer's brain. These empirical findings in monkeys have frequently enticed speculation that humans understand the actions of conspecifics because a human mirror-neuron analogous mechanism estimates possible outcomes of observed actions (Gallese and others 2004).

In humans, neuroimaging techniques allowed for a noninvasive exploration of whether identical brain regions are recruited during passive perception and active execution of particular actions (Buccino et al., 2001, Iacoboni et al., 1999, Nishitani and Hari, 2000). Direct evidence for mirror neurons in humans is, however, rare due to ethical constraints around electrophysiological recordings in healthy participants (Mukamel and others 2010). Nevertheless, it was suggested that a putative “mirror neuron system” (MNS) in humans could contribute to understanding others' actions and their underlying causes by internal neural simulation (Gallese and Goldman, 1998, Gallese et al., 2004, Rizzolatti et al., 2001). Etzel, Gazzola, and Keysers (2008) provided fMRI evidence for similar neural activity patterns in the premotor cortex during the execution and perception of an action. The authors used multivariate learning algorithms to classify the neural activity pattern of the premotor cortex when participants had to discriminate between the sound of a hand or mouth action in a similar task to that described by Gazzola, Aziz-Zadeh, and Keysers (2006). Once trained, the classification algorithm could determine whether the participant was executing a hand or mouth action at a later moment of the experiment. Taken together, electrophysiological and neuroimaging experiments emphasize the link between the perceived actions of others and their automatic reproduction in the observer, which invigorates the idea that the MNS allows to understand others by subliminal or subconscious re-experiencing or re-instantiating their behavior, compatible with the ST account.

Although the MNS was originally suggested to account for simulation of motor actions in single-cell recording experiments in monkeys (Chersi et al., 2011, Fogassi et al., 2005), evidenced neural simulations of emotion-unrelated motor action have often been extended to also explain a variety of social-affective psychological phenomena such as ToM and empathy (Goldman, 1992, Goldman, 2006, Gordon, 1986). Indeed, several authors have recently extended the MNS-based account of neural simulation to include empathic processes (Gallese et al., 2004, Keysers and Gazzola, 2009, Pfeifer et al., 2008) due to previous findings showing that empathy for pain involves at least some of the components of pain perception in an fMRI study (Singer and others 2004). Additionally, the ToM system again comes into play as soon as perceived movements can no longer be interpreted on the basis of expectations, but are different from what was anticipated (Georgescu and others 2014). This can be taken to suggest that the TT-related mentalizing system can be recruited to supplement the MNS and other instances of ST-related processes like empathy –two closely related and intertwined systems that are not mutually exclusive.

Even though parsimonious in principle, the ST has received numerous critics and revisions (Brass et al., 2007, Jacob and Jeannerod, 2005, Kilner, 2011, Newen and Schlicht, 2009). For instance, Mitchell, Macrae, and Banaji (2006) pointed out that a simulation mechanism would be necessarily limited to real-time social interactions during which one can perceive the other's current physical states. However, those mental states that derive from previous knowledge of attitudes or long-term dispositions cannot be inferred from observed external behavior (Mitchell and others 2006). Similarly, certain social behaviors may be less dependent on inferences from real-time sensory input. Umilta et al. (2001) reported that half of the mirror neurons they recorded in the premotor cortex in monkeys would fire not only in response to action observation, but also when the final part of the movement was blocked. The authors speculated that this subpopulation of mirror neurons would represent actions even if no actual movements are perceived in the environment. However, this animal experiment was not performed during a total absence of sensory input, but only during a limited time window. Therefore, the extent to which the MNS can simulate others' behavior without sensory input of others' ongoing motor action currently remains unclear.

Moreover, authors supporting a TT view have argued that ST cannot account for the systematic errors children make when attributing mental states to others (Nichols and Stich, 2003, Saxe, 2005). For instance, Ruffman (1996) showed in a behavioral study a set of beads grouped by color in different bowls to a group of children. He put a bead in a box, and then asked the children to guess what color would another individual think the bead was if he or she could not see from which bowl he took the bead. The results showed that children being 4 years old or younger would more often erroneously ascribe false beliefs to other individuals. Such finding suggests that children develop naïve rules about others' beliefs, such as “perceiving entails knowing”.

Congruently, Ramnani and Miall (2004) found evoked neural activity in the dorsal premotor cortex (PMd) during the preparation of responses in a Pavlovian associative task in which arbitrary visual cues determined future actions. In contrast, activity related to the anticipation of the responses of another individual did not activate the PMd, but the dorsal mPFC, TPJ and connected parts of the premotor cortex instead. Unlike other studies, Ramnani and Miall (2004) did not ask participants to explicitly attribute mental states to others. Rather, they provided explicit, simple rules that participants had to learn and their application was evaluated during the experimental condition. Thus, the authors ensured that participants could certainly anticipate the actions of other individuals. Despite the simple nature of the task, the activity in the dorsal mPFC and TPJ is consistent with a TT-related mechanism to anticipate others' behavior given that these regions have been persistently associated with ToM in perspective-taking tasks (Gallagher and Frith, 2003, Saxe and Kanwisher, 2003, Saxe et al., 2004).

Therefore, depending on the nature of the task, the ST appears inadequate to fully understand and predict the actions of another individual; ST is a necessary but not sufficient mechanism to understand others. This becomes especially clear in situations in which we can no longer predict or anticipate the outcome. Unexpected outcomes that do not match our assumptions plausibly require TT, rather than ST, mechanisms.

The TT and ST frameworks were introduced and treated as conceptual and empirical opponents based on evidence from developmental psychology, functional neuroimaging, and theoretical reasoning. This led many investigators in experimental psychology and neuroscience to take sides with either the ST or TT position (Carruthers, 1996, Goldman, 1992). However, behavioral and neural findings have encountered difficulties in settling whether the ST or TT is the predominant mechanism for explaining and predicting other individuals' behavior and their corresponding brain manifestations. In an fMRI study by Grèzes, Frith, and Passingham (2004), neural activity latency in the PMd and TPJ was higher when perceiving the actions of others compared to those of oneself. This is congruent with the ST account given that the neural mechanism involved in perceiving and simulating an action is the same when the observer and the agent of a movement refer to the same person. However, the authors also found activity in ToM-related regions when the observer inferred the agent's false beliefs about motor action, more coherent with TT than ST. Collectively, these results cannot unequivocally support either ST or TT as a unique explanation of understanding others' and one's own actions, suggesting that humans make combined use of both systems in everyday-life social dynamics.

The question whether the ability to attribute mental states to others is implemented by realizing simulation of (i.e. ST) or making inferences about (i.e. TT) others' behavior raised substantial interest in the last years from the computational perspective. Since there is a large tradition of computational approaches to study reward learning and decision making processes (O'Doherty et al., 2004, Schultz et al., 1997), many authors have designed experimental paradigms in which human participants are required to learn the contingencies of rewards from the observation of others' behavior. For instance, Behrens and others (2008) presented a learning game in which participants had to ascertain the likelihood of a reward's location as well as the reliability of a partner's advice regarding said location. A simple associative learning model could explain how participants updated their beliefs on both the location and the reliability of their partner. Moreover, the authors found that neural activity patterns within the pSTS/TPJ, a part of the brain bearing close relation to the ToM-related network (Saxe and Kanwisher, 2003, Van Overwalle, 2009), correlated with social predictions and, at a later stage, with prediction-error signals.

Beyond social reward learning, recent computational approaches have focused on the study of predictions about goal-directed actions, about other people's beliefs (i.e. ToM), or even about personality traits (for a review, see Koster-Hale & Saxe, 2013). For instance, Hampton and others (2008) designed a strategic game to investigate how well different computational models could explain the neural activity patterns observed in ToM-related brain regions. During the game experiment, participants alternated between the role of an employer and an employee. While the former could decide whether to inspect the employee or not, the latter could either work or avoid working at the risk of getting caught. These authors found that neural activity in the mPFC at the time of choice correlated with the prediction each participant made about their opponent's intentions. Furthermore, activity in the STS/TPJ at the time of the outcome correlated with the deviation of each participants' behavior from the prediction that their opponent had made (i.e. the prediction error). These results support the notion of ToM as an inference-based mechanism to understand the intentions of others in accordance with TT. In a similar fashion, Yoshida and others (2008) used a multi-player game where a human participant could either cooperate with the opponent (a computer agent) to hunt a large prey together, or to hunt a smaller prey on his own. In this ‘stag-hunt’ game, players must predict the other's goals to optimize their own behavior and, in doing so, they must consider that the opponent follows a similar strategy as they do. That is, the two players need to mutually adjust their choice behavior by considering the predictions of the interaction partner. The authors compared different computational models of choice behavior using simulated and real data. Their results showed that a model that considered that players made inferences about each other could better predict how players would behave in subsequent games, again in concordance with the TT framework. The importance of this recursive nature of the inferences we make to understand others' intentions has been further supported in more recent experiments (Devaine et al., 2014, de Weerd et al., 2015). Overall, computational models to probe mental state attribution have so far been congruent with an inference-based mechanism (i.e. TT) for understanding others when individuals engage in highly demanding tasks. As defined by Baker and others (2011), “ToM inferences come surprisingly close to those of an ideal rational model, performing Bayesian inference over beliefs and desires simultaneously”.

Besides the now classical ST and TT accounts, many alternative, often integrative theoretical accounts have emerged over the past decade to accommodate the shortcomings of the two classic views. Gallagher (2008) proposed a perceptual mechanism based on the premise that others' external behavior is a direct expression of their mental states. The so-called direct perception theory implies that when perceiving socially relevant environmental information (e.g., faces or body movements), a pattern-matching mechanism dependent on previously learned stimuli configurations would detect the behavioral fingerprint linked to a specific mental state. As such, neither abstract inference nor bodily simulation would be required to understand other individuals' intentions, in contrast to the TT and ST accounts.

Another alternative account for social cognition, referred to as narrative practice hypothesis (Hutto, 2008), accredits a fundamental role in the understanding of another's behavior to how humans acquire knowledge about other conspecifics through narrative stories. In a later redefinition, both authors combined their previous acquaintances to state that not only individuals understand from direct observation of other individuals' behavior (as originally suggested by Gallagher), but also get involved in interactive situations with them. From these interactions, humans learn about others by narratives which, in turn, provide abstract background knowledge for future social interactions (Gallagher & Hutto, 2008). Overall, a common feature of recent accounts for social interaction is that individuals do not appear to depend on a single mechanism to understand others, but on more than a set of complementary mechanisms. It is their elusive nature what remains a topic of debate: low-versus high-level simulation (Goldman, 2006), intuition versus inference-based understanding (Gallagher, 2001), or implicit versus explicit modeling (Newen, 2014).

There is hence lack of solid experimental psychology or neuroscientific evidence that could tip the balance between different candidate mechanisms to explain social behavior. Some authors are therefore calling into question the long-assumed opposition between TT and ST by pointing out that understanding others might involve two different, yet synergistic and complementary mechanisms that simply have two different functional roles in interacting and communicating with others (Apperly, 2008, De Lange et al., 2008, Kilner, 2011, Van Overwalle and Baetens, 2009). Brass, Schmitt, Spengler, and Gergely (2007) concluded from a functional brain-imaging study that an MNS-related mechanism did not mediate action understanding when the observed action was novel or when it was hard to understand. Instead, as stated before, a mental-state inference mechanism may be required. The authors further argued that it is the contextual plausibility that determines whether the observer can map the target's behavior based on own motor schemes in stereotypical, highly familiar actions, or they would need to explicitly infer the purpose of an unusual action in novel contexts. In line with this, Apperly (2008) underscored that TT and ST may reflect two different shades of ToM, and argued against interpreting the results of neuroimaging work within either of the two theoretical frameworks. This perspective receives support from authors warranting that such distinction may hinder further progress in our understanding of social-cognitive processes (Gallese & Goldman, 1998).

In another neuroimaging study, Santos et al. (2010) showed that gradual induction of a sense of animacy (via biological movement) recruited different key regions of the social brain. Abstract, inert objects were perceived as animated only on the basis of changing motion parameters that were highly suggestive of personal agents “behind” the movements. While the evaluation of actually present animacy signals recruited the vmPFC, AI, STS, FG, and HC, the mere disposition to detect socially salient movements was associated with increased neural activity in the superior parietal lobe and ventral PM, thus resembling MNS activation. In line with the social Bayesian concept of ToM discussed above, this might point to a putative gradient of complexity between TT-associated mechanisms of abstract emulation and ST-associated mechanisms of embodied simulation underlying social interaction. Thus, while understanding simple motor actions performed by others might involve a specific, lower-level neural mechanism, the more hidden the intention behind the motor act becomes, the more a higher-level neural mechanism would be needed (Vogeley, 2017).