Towards a neurocomputational account of social controllability: From models to mental health

Abstract

Controllability, or the influence one has over their surroundings, is crucial for decision-making and mental health. Traditionally, controllability is operationalized in sensorimotor terms as one’s ability to exercise their actions to achieve an intended outcome (also termed “agency”). However, recent social neuroscience research suggests that humans also assess if and how they can exert influence over other people (i.e., their actions, outcomes, beliefs) to achieve desired outcomes (“social controllability”). In this review, we will synthesize empirical findings and neurocomputational frameworks related to social controllability. We first introduce the concepts of contextual and perceived controllability and their respective relevance for decision-making. Then, we outline neurocomputational frameworks that can be used to model social controllability, with a focus on behavioral economic paradigms and reinforcement learning approaches. Finally, we discuss the implications of social controllability for computational psychiatry research, using delusion and obsession-compulsion as examples. Taken together, we propose that social controllability could be a key area of investigation in future social neuroscience and computational psychiatry research.

1. Introduction

The world we live in is complex and uncertain. Controllability, defined as the influence one has over their surroundings and inherently linked to the concept of agency, allows an agent to use their actions to produce an intended outcome or avoid an unwanted one (e.g., sensorimotor responses, rewards, social goals; Dorfman & Gershman, 2019; Friston et al., 2013; Moore et al., 2010). For humans, our social world—due to its high degree of complexity and uncertainty— may represent the most relevant yet most challenging case for exerting behavioral control. As such, having control in social interactions (“social controllability”) is a critical ability for not only survival, but also for our mental wellbeing. For instance, both overestimating controllability of others (e.g., in delusion) (Na et al., 2022) and displaying enhanced aversion to unexpected outcomes when overexerting control in situations that are not controllable (e.g., in obsessive-compulsion) (Banker et al., 2022) may lead to misconstrued beliefs about the external world and negatively impact psychosocial functioning.

While social controllability (see Table 1 for Glossary) is difficult to manipulate or quantify in experimental settings, recent work in computational neuroscience provides a useful framework for quantifying behavioral and neural processes that support an agent’s ability to actively exert behavioral control to achieve desired outcomes in social settings (Huys & Dayan, 2009; Moscarello & Hartley, 2017). Furthermore, a rich social neuroscience literature has provided important insight into the underpinnings of human behavior under the influence of others (Chung et al., 2015; Lockwood et al., 2016; Lockwood & Klein-Flügge, 2021; Zhang & Gläscher, 2020). In this review, we will specifically highlight mechanisms pertaining an agent’s ability to actively exert influence over others from the behavioral economics (e.g., social preference models) and reinforcement learning literature after briefly reviewing the literature on how humans behave under influence. Finally, we will discuss how controllability computation might breakdown in psychiatric conditions, using delusion and obsession-compulsion as examples.

Table 1.

Glossary

Concept	Definition
Contextual controllability	The extent to which a context or environment allows an agent to exert control; exists in both nonsocial and social contexts
Perceived controllability	An agent’s estimation (i.e., internal model) of how much control or influence they could have; exists in both nonsocial and social contexts
Social controllability	The degree to which an agent’s actions can alter or influence other people (i.e., their actions, outcomes, beliefs, etc.) to achieve desired outcomes
Computational modeling	A mathematical formalization that characterizes or predicts a behavior or cognitive process, defined by parameters that vary across the population
Reinforcement learning	Learning behaviors that become optimized by maximizing expected outcomes or goals
Model-free (MF) learning	Trial-and-error learning that updates cached values of choices that become associated with a stimulus
Model-based (MB) learning	Learning that involves mental simulation of future states and representations of potential state transitions and outcomes (i.e., forming a cognitive map)
Cognitive map	A mental representation (“map”) of both spatial and non-spatial elements of the environment that depends on relational learning

2. Contextual versus perceived controllability

Controllability has been studied from drastically different perspectives in various fields such as psychology, sociology, management, and political science (Acemoglu & Robinson, 2012; Huys & Dayan, 2009; Magee & Galinsky, 2008; Maier & Seligman, 1976; Moscarello & Hartley, 2017; Schaerer et al., 2018). Briefly, these studies have examined two aspects of controllability: (1) controllability that emerges from a situation or context (“contextual controllability”) and (2) an agent’s internal model of how much control or influence they could have (“perceived controllability”). Contextual controllability is typically manipulated by the experimenter in laboratory studies and can be seen as the objective aspect of controllability, whereas perceived controllability is either manipulated by experimenters or subjectively estimated by participants. Each of these aspects of controllability exists in both nonsocial and social contexts and has been differentially linked to crucial wellbeing outcomes (Leotti et al., 2010). In this next section, we will synthesize the literature on controllability and how it influences mental wellbeing.

2.1. Contextual controllability

In social sciences, extensive research has identified situations (Fast et al., 2009; Fiske, 1993; Fleischmann et al., 2019; Kipnis, 1972), institutions (Acemoglu & Robinson, 2012; Schaerer et al., 2018), and cultures (Gelfand et al., 2011; Harrington & Gelfand, 2014), or social hierarchies (i.e., social class; Dubois et al., 2015; Magee & Galinsky, 2008; Rucker & Galinsky, 2017) that grant agents control (i.e., contextual controllability). In these contexts, the term ‘power’ or ‘authority’ have been commonly used to indicate the granted control. For example, control can emerge due to social constraints imposed on others (Gelfand et al., 2011), collective accesses to resources (Piff et al., 2018), or the distribution of the controllability in a society (Acemoglu & Robinson, 2012) and so on. People are able to appraise their level of control via one’s status (Magee & Galinsky, 2008), resources (Schilke et al., 2015), roles (Guinote, 2007), or relationships in social network (Sichani & Jalili, 2017).

In neuroscience laboratory studies, however, contextual controllability is usually operationalized as the extent to which animals can exert actions that have influence on the environment (“agency”), which is distinct from the human concepts of power and control. It is known that rodents exhibit stress-related behavioral, neural, and physiological responses in aversive situations that are uncontrollable (e.g., inescapable shocks) but not in aversive situations that are controllable (e.g., escapable shocks) (Maier et al., 2006). Prolonged exposure to uncontrollable stressors leads to learned helplessness, often used as an animal model for depression (Bland et al., 2003; Maier & Watkins, 2005). Work from human neuroimaging also indicates that people are averse to uncontrollable situations. For example, greater neural activation in the amygdala, insula, cingulate cortex, and caudate nucleus is observed during uncontrollable versus controllable events (Wang & Delgado, 2021). The medial prefrontal cortex (PFC) seems to play an important role in mediating the relationship between stressor controllability and reduced stress responses. For example, rats respond to escapable shocks as if they are inescapable if medial PFC is lesioned or deactivated (Amat et al., 2005, 2006). In the next section, we will discuss how positive effects of perceived controllability may depend on these neural regions.

In humans, having control is linked to physical and mental wellbeing (Fein, 1995; Fiscella & Franks, 1997; Sapolsky, 2005) and restricting control is linked to the opposite (Bandura et al., 1985; Morgan & Tromborg, 2007). Greater contextual controllability heightens people’s self-esteem (Wojciszke & Struzynska–Kujalowicz, 2007), increases confidence in decision-making (Fast et al., 2009, 2012), decreases advice-taking from others (Tost et al., 2012), and facilitates psychological resilience during fear extinction learning (Hartley et al., 2014). A recent cross-cultural study indicates that subjective wellbeing is also promoted among countries that highly value the autonomy to pursue personal goals (e.g., prosocial goals) (Rhoads, Gunter, et al., 2021). Indeed, having control enhances goal-orientation, granting people more means to achieve their goals and exert more control even in unrelated domains (Guinote, 2007, 2017). Having more resources and higher status is also linked to behaviors that promote others’ wellbeing (Gittell & Tebaldi, 2006; Hughes & Luksetich, 2008; Korndörfer et al., 2015; Post, 2005; Stamos et al., 2020), the effect of which seems to be moderated by feelings of power (Dubois et al., 2015). Thus, it is unsurprising that people exhibit strong preferences for power (Mason & Blankenship, 1987) and seeking control over outcomes (Wang & Delgado, 2019). However, it is not always the case that contextual controllability leads to positive outcomes—for example, wellbeing is negatively impacted by the fear of losing control (Kelly-Turner & Radomsky, 2020; Radomsky, 2022) as well as the fear of guilt that could result from behaving irresponsibly (Mancini & Gangemi, 2004; Salkovskis et al., 2000).

2.2. Perceived controllability

The true controllability of the external world is usually unknown, so people must generate an internal model of how much control they can have over other people to produce desired outcomes (i.e., subjective estimate of their own controllability). This estimation is not always accurate. Langer (1975) found that people tend to overestimate their control particularly for desirable outcomes (Langer, 1975). This ‘illusion of control’ resembles optimism, which in certain situations could be beneficial (Fontaine et al., 1993; Taylor & Brown, 1994), but might also reflect a cognitive (in)ability to learn causality or action-outcome contingencies (FentonO’Creevy et al., 2003). In contrast to the illusion of control, people sometimes underestimate controllability as well. For instance, a field study showed that people receive more ‘yes’ responses for their requests (e.g., to borrow their phone, sponsor their race) from strangers than expected, showing that people may underestimate their influence over others in certain situations (Bohns, 2016). It is important to note that the degree to which people may overestimate or underestimate their influence likely depends on the context.

Despite this variation in accuracy, people’s beliefs about their own controllability better explain motivation (Eitam et al., 2013) and behavior (Skinner, 1996) than the true contextual controllability that is set up in a laboratory, and are consistently associated with better physical and mental health. Perceived controllability over one’s life is highly associated with educational achievement (Daniels et al., 2014; Stupnisky et al., 2007), career performance (Brockner et al., 2004; C. Lee et al., 1990) and adaptability (Duffy, 2010; Zhou et al., 2016). For example, perceived control strongly predicts college students’ achievement even compared to self-esteem (Stupnisky et al., 2007) or goal-oriented motivation (Daniels et al., 2014). Perceived control affects performance indirectly as well. In a field study, survivors with higher perceived control who were laid off by employers felt the post-layoff experiences less threatening, which was associated with greater job performance (Brockner et al., 2004). Belief that one can make a difference rather than passively accept what happens is generally a productive coping strategy (Compas et al., 1991). Perceived controllability also acts as a buffer against negative outcomes later in life. Higher perceived control is related to reduced declines in health and cognition among older adults a decade later in life, even controlling for other objective factors such as demographics, health status, and physical risk factors (Lachman & Agrigoroaei, 2010; Robinson & Lachman, 2018). Also, stability in perceived control is associated with better physical health among old people (Chipperfield et al., 2004). The benefits of perceived controllability emerge even in the absence of contextual controllability or if one has the option to exert control but do not actually exert that control (Thompson, 1981).

Furthermore, people also have beliefs about controllability that generalize across contexts (i.e., global beliefs about controllability). A rich literature from clinical, psychological, and management science probed global beliefs based on “self-efficacy” (Bandura, 1982; Sherer et al., 1982), “sense of control” (Lachman & Weaver, 1998; Rodin, 1986), or “locus of control” (Rotter, 1966). These self-reported beliefs have shown significant connections with life outcomes such as career (Caliendo et al., 2015; Zhou et al., 2016) and health (Harrow et al., 2009; Lachman & Weaver, 1998; Maggio et al., 2019; Oi & Alwin, 2017; Southwick & Southwick, 2018). A longitudinal study with a dataset of ~10,000 participants at age 5, 10, 16, and 30 also showed that locus of control was longitudinally quite stable during this period and that childhood locus of control predicted educational and vocational qualification whereas education in childhood did not predict later locus of control (Oi & Alwin, 2017). Global beliefs about controllability are also associated with self-efficacy, self-esteem, or dominance (Bandura, 1982; Galvin et al., 2018; R. E. Johnson et al., 2015). While the basic need for controllability may be motivated by these various positive outcomes (Leotti et al., 2010), it is also likely that both perceived controllability and the preference to exert control are modulated by personal experience with control (Mineka & Hendersen, 1985). For example, learned helplessness (Maier & Seligman, 1976, 2016), which is a generalized belief that situations are uncontrollable, characterizes depression and is widely used in animal models of depression. Also, a loss of sense of control strongly contributes to burnout (Southwick & Southwick, 2018).

Perceived controllability is typically measured by self-reports in human subject experiments (in contrast to contextual controllability as systematically manipulated). It can also be experimentally manipulated, for example, by simply changing the task instructions (Stolz et al., 2020). Changing instructions alters people’s level of perceived controllability, which positively correlates with the level of medial PFC activation (Stolz et al., 2020). Having control can also inflate the value of a reward, and that inflated amount (i.e., the subjective value of control) is tracked by ventromedial PFC activation (Wang & Delgado, 2019). In addition, people with higher ventromedial PFC activation to controllable cues also exhibited a greater increase in avoidance behavior in the controllable compared to uncontrollable events (Wang & Delgado, 2021). While these previous studies pinpoint towards the medial PFC as a key region in supporting perceived controllability and its interaction with contextual controllability (Leotti et al., 2010; Stolz et al., 2020; Wang & Delgado, 2019), it remains poorly understood if the neural computations underlying perceived and contextual controllability are the same due the lack of computational tools or spatiotemporal precision of neuroimaging techniques used in these studies.

Controllability has significant implications on various aspects of our lives. Yet, it remains an abstract construct and difficult to measure in laboratory settings. In the studies described above, researchers either (1) measured or manipulated the level of control people have (e.g., social status, power) or (2) relied on participants’ self-reported perceptions of control. While the studies above suggest that similar brain regions may underlie the basis of both contextual and perceived controllability, it does not necessarily indicate that they are supported by the same computational accounts. Computational approaches can help further understanding the processes underlying active controllability, particularly during social interactions. The next section introduces relevant computational approaches that could provide clues for quantifying behavioral and neural processes that support actively exerting social controllability.

3. Modeling social controllability

Recent work from social neuroscience suggests that controllability over other people is an important computation implemented by the human brain and subsequently contributes to mental wellbeing. As such, social controllability, or the degree to which we can alter or influence the actions, outcomes, or beliefs of other people to achieve desired outcomes (i.e., our ability to influence other people), may constitute the exercise of the ultimate extent of agency for humans and entails both contextual and perceived facets.

3.1. Being influenced versus influencing others

Before discussing how humans exert social control, it is important to note that there is a rich literature examining how humans behave under social influence (Cacioppo et al., 2018; Chung et al., 2015; Falk & Scholz, 2018; Lockwood et al., 2016; Zhang & Gläscher, 2020). For instance, people are influenced by each other’s opinions similarly regardless of differences in their competence (Mahmoodi et al., 2015; Zonca et al., 2021), but may rely more on peers’ opinions when they are endorsed by other people (Zonca et al., 2021). In a gambling task, observing others’ choices of gambles influenced the subjective value (e.g., utility) of those gambles for the observer. This subjective value was encoded in the ventromedial PFC and indicated the degree to which observers’ gambles were influenced by others’ choices (Chung et al., 2015). One study demonstrated how people weigh tradeoffs between short-term incentives for an individual and long-term incentives for the groups (Park et al., 2019). The researchers found that ventromedial PFC encodes immediate expected rewards as individual utility while the lateral frontopolar cortex encodes the utility of the group. Crucially, they found that these utilities depended on updating beliefs about others’ decisions encoded in anterior cingulate cortex and the temporoparietal junction (TPJ). When competing to advise other people’s choices, TPJ activity tracks the level of influence people have while medial PFC activity tracks the level of performance relative to a rival (Hertz et al., 2017). Even in simple probabilistic learning tasks with financial rewards, people’s actions are influenced by others’ rewards (Lockwood et al., 2016; Rhoads et al., 2023) and even integrate information about others’ choices (and the group consensus) to optimize their own outcomes (Zhang & Gläscher, 2020). Neurophysiological coupling between influencers and those being influenced in key regions related to considering others’ mental states and subjective valuation may also underlie successful social control. For example, greater alignment of neural activity in anterior insula and anterior cingulate is observed between proposers and responders whose choices reflected greater reciprocal reactions during an multi-round ultimatum game (Shaw et al., 2018).

Fewer studies have focused on the mechanisms underlying how people actively exert influence over others. Among the studies that have focused on this topic, it is clear that neural regions underlying the ability to consider the internal states of other people (i.e., mentalizing about their abilities, beliefs, emotions, goals) plays a critical role for successfully influencing others (Baek et al., 2020; Baek & Falk, 2018). These regions typically include the medial PFC, TPJ, precuneus, superior temporal sulcus (STS), and temporal poles. For example, successfully convincing others to accept or reject TV show ideas is linked to greater TPJ activity during initial exposure to the ideas (Falk et al., 2013). Regions that overlap with the brain’s valuation system similarly play a key role—people exhibit increased medial PFC and ventral striatal activity when their advice is accepted by others (Mobbs et al., 2015). It has also been shown that these two mental processes—accounting for social influence in one’s own decision-making versus being able to exert social control—are also heavily intertwined. In one study (Mahmoodi et al., 2018), participants completed a visual perceptual decision-making task in which they estimated the spatial position of a previously displayed target stimulus before and after observing a peer’s estimate. This study found that people are more likely to be influenced by people that they have influenced previously.

3.2. Social preference and controllability in standard behavioral economic paradigms

Choices in social situations often depend on our preferences for concepts such as cooperation, altruism, reciprocity, inequity aversion, conformity, or fairness (Konovalov et al., 2018; D. Lee, 2008; Rhoads, Cutler, et al., 2021; Rilling & Sanfey, 2011; Sanfey, 2007). These studies have shown that complex human social preferences can be modeled and tested using economic games such as dictator or ultimatum games [Figure 1; (Camerer, 2011)]. In a typical dictator game (Figure 1a), the participant might be facing a “dictator” who unilaterally decides how to split a certain amount of money (e.g., $3 out of a total of $10). In this scenario, the level of controllability is lowest (i.e. none). In contrast, the ultimatum game (Figure 1b) bestows some level of control to the participant (responder) by allowing her to either accept or reject an offer. If the participant accepts the offer, both players split the money as proposed; if she rejects however, neither party receives anything. In this case, the participant has control over the immediate outcome, but not over other people in general as the offers are usually still random. Finally, in a modified version of the ultimatum game called the social controllability game (Figure 1c) (Na et al., 2021), a participant’s current decision will not only determine their current payoff, but also influence how much other players might propose in the future. In this paradigm, the human participant not only has control over her and her counterpart’s immediate outcomes, but also can estimate the degree of control she has over others and thus can influence their future behaviors. In other words, as the paradigm offers contextual controllability (manipulated by the experimenter) as well as perceived controllability (estimated by participants), the path to achieve one’s desired goals (i.e., exerting social influence to maximize rewards) must include a computation of the perceived effects one has on somebody else’s actions.

Figure 1. Probing social controllability with behavioral economic paradigms.

Behavioral economic paradigms are commonly used to measure social interactions that are of varied levels of contextual controllability. a) In a typical dictator game, a participant often plays the role of the responder, who faces a “dictator” that unilaterally decides how to split a certain amount of money (e.g. $3 out of a total of $10). The participant has no control over the interaction. b) In a typical ultimatum game, the participant often plays the role of a responder (left) who can either accept or reject an offer from the proposer. Here, although the responder cannot control the current offer from the proposer, they can control the immediate outcome by either accepting (split as is) or rejecting (both receive nothing) the offer. c) In the social controllability game (Na et al., 2021), the participant plays an iterated version of the ultimatum game with multiple proposers; importantly, they can influence future offers by accepting or rejecting the current offer as a social signal.

Mathematical models have also been developed to quantify these highly psychological concepts. For example, Fehr and Schmidt (Fehr & Schmidt, 1999) developed an inequity aversion model that explain why people do not always maximize financial rewards in exchange games such as the ultimatum game (Figure 1b). Specifically, the Fehr-Schmidt model suggests that people are aversive to unfairness—either advantageous or disadvantageous towards themselves—and that the degree of aversion varied among individuals (Charness & Rabin, 2002).

3.3. Model-free learning in social interactions

The first step for an agent to exert any level of control is to learn the contingency between one’s action and its consequences. A prevalent framework to explain such process is reinforcement learning (RL) (Sutton & Barto, 2018), which is widely used in cognitive science and artificial intelligence. Under the RL framework, the goal of behavioral control is to maximize rewards (and minimize losses), assuming that rewards are contingent on an agent’s own actions and/or states. (Here, states refer to different properties of a system being modeled, such as positions in space, sequences in time, events, environmental conditions, emotions, beliefs). To achieve that goal, an agent builds a value function that maps actions (or states) onto expected rewards and updates this value function based on errors that deviated from its expectations (“prediction errors”). At the brain level, RL signals are known to be computed by dopamine neurons in the mesolimbic circuitry (Deserno et al., 2015; Schultz et al., 1997).

RL can be broadly understood in terms of two behavioral control modalities: model-free (MF) and model-based (MB) (Daw et al., 2011; Gläscher et al., 2010). MF learning involves learning based on associations between specific choices (or states) and outcomes in order to guide future behaviors (“trial-and-error”). MF learning is quick to process since choices are immediately assigned values based on outcomes. However, it is inflexible, limited by the associations it has formed, and thus cannot easily adapt to changes in the environment (Atkeson & Santamaria, 1997; Dolan & Dayan, 2013). When action-outcome contingency is low or when the structure of the environment changes, MF control may produce suboptimal behaviors because actions cannot be tied to outcomes (Dorfman & Gershman, 2019).

MF learning is increasingly being used to characterize behavior under many different social contexts (Charpentier & O’Doherty, 2018; Konovalov et al., 2018; Rhoads & Gan, 2022). A recent study has also combined MF learning with social preference models to show that people dynamically adapt their internal criterion of fairness (Gu et al., 2015). Specifically, this study characterized how people respond to norm violations and update subsequent expectations about other people’s behaviors—an important feature of successful social controllability. This subjective norm adaptation is guided by norm prediction errors (i.e., the difference between one’s expectation of how people should behave and how people actually behave). Damage to anterior insula impaired how people updated their norm expectations (Gu et al., 2015). These studies demonstrate how MF learning mechanisms can be combined with behavioral economic models to account for the dynamics in human social behaviors.

MF control characterizes decisions based on a history of experiences and reward contingencies without considering the underlying structure that governs those experiences. This mode of behavioral control is useful in situations where one does not need to consider changes in complex social environments. However, suboptimal behavior can result from MF control when environments or goals change. For example, this strategy may not be useful when the task requires one to simulate the future actions of other people, as it tracks the association between sensorimotor responses and their outcomes but not others’ intentions over time (Camerer, 2011). This highlights the importance of incorporating a model-based framework, which additionally considers the underlying environment in which people behave.

2.2. Model-based control in social interactions

In contrast to MF control, MB processes rely on an in-depth value function which incorporates the relationship between actions and state transitions. State transitions in this context refer to changes in different properties of a system being modeled (e.g., positions in space, sequences in time, events, environmental conditions, emotions, beliefs). State transitions can be learned through state prediction errors, which are independent from observable rewards (Daw et al., 2011; Gläscher et al., 2010), and make up an internal model of an environment, commonly called a ‘cognitive map’ (Behrens et al., 2018). Notably, two-dimensional mental state spaces potentially relevant to cognitive maps have been shown to be mapped in hippocampus (O’keefe & Nadel, 1978; Tavares et al., 2015), entorhinal cortex (Stensola et al., 2012), and ventromedial PFC (Constantinescu et al., 2016). MB learning is not as fast as MF learning because an agent needs to mentally simulate multiple paths of future possible steps in the ‘map’ to reach the final output of the value function. However, it is more flexible in that it allows individuals to take changes in the environment into account and adjust behaviors accordingly (in contrast, MF learning may impede adaptive behavior if the environment changes and learned associations no longer hold true).

Because the complexity and continuity of social interactions often requires temporarily extended learning which must include an assessment of somebody else’s model of the environment, pattern of actions, goals etc., MB learning is also more suitable to explain the full exploitation of social controllability than MF mechanisms. Specifically, MB learning allows for forward planning or mental simulation of events that have not yet occurred, which relies on information about state transitions (Szpunar et al., 2014). In non-social contexts, converging neural and computational evidence support the notion that MB learning accounts for choice behaviors in tasks that require forward planning and mental simulation of the future (Daw et al., 2011; Dolan & Dayan, 2013; Doll et al., 2015; Gläscher et al., 2010). For instance, using a two-step task and various visual stimuli, one fMRI study demonstrated that MB learning was associated with visual representations of prospective stimuli (Doll et al., 2015). Although MB is considered more flexible compared to MF learning, computation becomes exponentially expensive as the sequence of events is extended (Huys et al., 2012, 2015), which is likely in social interactions.

MB mechanisms are a well-suited account for agency and controllability in social settings, as humans naturally mentalize about other people’s internal states or characters to predict subsequent behaviors (e.g., Theory of Mind). That is, we tend to build mental models of other individuals during iterated interactions with them or form mental models of groups of people even when interaction with a specific person is only one-shot (i.e., stereotypes). Again, behavioral economic paradigms and models provide initial insight into the mechanisms supporting social controllability (Camerer, 2011; Rusch et al., 2020). For instance, when playing a two-player strategic game in which opponents have competing goals, people consider how their actions will affect their counterpart’s mental model of them, rather than just considering the immediate prediction errors related to other peoples’ choices (Hampton et al., 2008; Hill et al., 2017). In other types of strategic games, such as bargaining, individuals differ drastically in their ability to manage their social images and exert influence on others, a behavioral phenomenon subserved by underlying neural differences in prefrontal regions (Bhatt et al., 2010). Furthermore, humans are able to use forward planning to mentally simulate future interactions during a multi-round trust game (Hula et al., 2015).

Finally, Na and Chung et al. (2021) proposed a computational account of behavior during a social controllability game (Figure 1c). Their computational model predicts that participants’ choices to accept or reject offers are guided by the mental simulation of the summed value of both current and future outcomes (i.e., forward planning; Figure 2a), where the number of planning steps is also incorporated into the model. On a given trial, the utility of a choice is a function of the current reward and internal norm expectations. The internal norm expectation is initialized as one’s subjective initial expectation of a fair offer (f₀) which is iteratively updated via norm prediction errors after each received offer. Norm prediction errors are weighed by the norm adaptation rate (ϵ), which governs the degree to which one updates their norm expectations. In the utility function, norm violations (defined as the difference between the actual offer received and one’s internal norm expectation of the offers) are weighed by a parameter (α) that indexes individuals’ aversion to norm violations. Importantly, the utility of future choices depends on the expected influence (δ) that individuals estimate they have over the interactions (i.e., how much the next offer will increase or decrease if they reject or accept the current offer). The probability of choices is determined using a softmax function that weighed the utility of choices by the inverse temperature (β), which indicates how much people base their choices on the estimated utility. This model thus incorporates five free parameters that govern social controllability and vary across individuals (Table 2).

Figure 2. Modeling social controllability and implications for computational psychiatry.

a) In the social controllability game, participants build a mental model of potential future states based on available actions (e.g., accept or reject). This forward thinking model is similar to the concept of cognitive map in nature and different individuals might engage different lengths/steps in their forward planning horizons. Modified from (Na et al., 2021). b) Individuals with high trait delusion demonstrated a heightened illusion of control as indexed by a higher “expected influence” parameter estimated from their choice behavior (left panel) as well as greater self-reported social controllability (right panel) during uncontrollable interactions. Modified from (Na et al., 2022). c) In contrast, individuals with high misophonia and obsessive-compulsive symptoms showed a greater disconnection between their perceived controllability and modelestimated parameter of “expected influence” (left panel), while also showing heightened sensitivity to norm violation (right panel) during uncontrollable interactions. Modified from (Banker et al., 2022). PDI: Peters Delusion Inventory. OC: obsession-compulsion. Miso: misophonia (a syndrome hallmarked by intolerance to sounds generated by other people). C: controllable interactions. U: uncontrollable interactions.

Table 2.

Description of the free parameters in the social controllability model and current findings related to computational psychiatry.

Free parameter	Name	Description	Range	Impacted in delusion	Impacted in Miso-OC
f0	Initial norm	An individual’s initial expectation of offers prior to beginning the task	$0 ≤ f0 ≤ $20
ϵ	Norm adaptation rate	Degree to which an individual updates their norm expectations	0 ≤ ε ≤ 1
α	Aversion to norm violations	Degree to which an individual is averse to norm violations	0 ≤ α ≤ 1		Higher α in uncontrollable condition
δ	Expected influence	Amount of the offer changes that an individual estimated would be induced by rejecting the current offer	−$2 ≤ δ ≤ $2	Higher δ in uncontrollable condition	Greater difference between perceived controllability (self-reported) and δ in uncontrollable condition
β	Inverse temperature	Degree to which an individual weighs the current utility of a choice	0 ≤ β ≤ 20

Despite robust evidence on the existence and the impact of perceived controllability in general (Hashimoto et al., 2015; R. E. Johnson et al., 2015; Lachman & Weaver, 1998; Oi & Alwin, 2017; Rodin, 1986), the mechanisms underlying these beliefs have been rarely examined through computational methodologies. This is possibly due to operational difficulties and extensive computation, since one must simulate future actions based on multiple models of the environment (one’s own and those estimated for the other agents involved). Given its relevance to self-concept, perceived controllability might also arise from metacognition (knowledge about one’s own cognition). Brain regions related to locus of control overlapped with the “self-as-object network” (Hashimoto et al., 2015; Sui & Gu, 2017), which might suggest that the brain has a metacognitive “critic’ of its own self (Sui & Gu, 2017). Fleming and Daw proposed a hierarchical Bayesian framework to explain metacognition (Fleming & Daw, 2017). They showed that unlike the first-order model in which action and belief are identical, the second-order model could explain error detection and belief adjustment following an action.

Collectively, this body of research indicates that controllability can be effectively and meaningfully modelled in social paradigms—noting that (1) MF and MB RL are just one collection of models that could characterize social controllability and (2) the dichotomy described here is useful at a conceptual level but humans likely use a mixture of MF and MB control in real-world interactions (see Collins & Cockburn, 2020). Modeling active and perceived controllability during social interactions may also further our understanding about how and when aberrant computations of social controllability impact mental wellbeing. In the next section, we will use illusion of control as an example to demonstrate how aberrant computation and perception of social control might go awry in clinical disorders.

4. Implications for computational psychiatry

As reviewed above, aberrant estimates of social controllability may also have negative impacts on various aspects of life, including but not limited to one’s social relationships, job prospects, financial resources, and ultimately, mental health. It has been, however, difficult to quantify controllability or understand its mechanics due to the lack of computational approaches.

4.1. Delusion

Illusion of control represents an example of altered controllability estimation that is widespread in the general population, and is characterized by the incorrect perception of a strong sense of control (D. D. Johnson & Fowler, 2011). Illusion of control has also been frequently associated with delusions (R. P. Balzan et al., 2013), or false beliefs representing a hallmark characteristic of schizophrenia (American Psychiatric Association & American Psychiatric Association, 2013). Although delusions have not been extensively studied in direct relation to social controllability, common types of delusions such as paranoia and suspicion are manifestations of inherently distorted social models prompting one’s fixed belief that they might be harmed or otherwise manipulated by another person, group, or organization (American Psychiatric Association & American Psychiatric Association, 2013; Griffin & Fletcher, 2017). Furthermore, considering a dimensional rather than diagnostic approach to psychosis, aberrant estimation of social controllability might be shared along a continuum that ranges from a personality trait in community samples to a psychiatric symptom in patients with schizophrenia (Van Os et al., 2009).

Previous work has attempted to understand how aberrant (non-social) beliefs might prompt delusional ideation. For instance, it has been shown that delusional individuals prioritize evidence that matches their beliefs (R. Balzan et al., 2013) and tend to discount evidence that contradicts their beliefs (Woodward et al., 2007). An inclination to ―jump to conclusions‖ has also been associated with delusion (Ward & Garety, 2019), although subsequent work has not consistently replicated this finding (McLean et al., 2020; So & Kwok, 2015) and the theoretical argument is challenged by findings that suggest greater propensity for information-seeking in severe schizophrenia (Baker et al., 2019). A bias against disconfirmatory evidence also exists across the psychosis continuum (Eisenacher & Zink, 2017). Further along the cognitive hierarchy, distinctive metacognitive beliefs exist across delusion spectrum in non-clinical populations (Larøi & Van der Linden, 2005; Stainsby & Lovell, 2014).

Using the aforementioned social controllability paradigm (Figure 1c) and modelling approach (Figure 2a), Na and Blackmore et al. examined illusion of control in an online sample of individuals with either high or low trait delusion (Na et al., 2022). They found that while intact in exploiting controllability during controllable interactions, people with high trait delusion attempted to exert a higher level of control (higher δ) and self-reported a stronger sense of perceived controllability in situations that they were not able to control (Figure 2b). Collectively, these results suggest that delusion is primarily linked to an illusion of control at both behavioral and self-report levels.

4.2. Obsession, compulsion, and misophonia

Altered perception or computation of controllability has also been linked to obsessive-compulsive (OC) symptoms and misophonia (a syndrome with aversion to sounds generated by other people), which have been recently shown to fall onto the same clinical spectrum (Banker et al., 2022). Seemingly different, both OC and misophonia are hallmarked by a intolerance to uncertainty and aversion to losing control (Natalini et al., 2020; Reuven-Magril et al., 2008; Schröder et al., 2013). Using same social controllability task, Banker and colleagues found that individuals with higher OC-misophonia scores demonstrated a greater discrepancy between their perceived controllability and behaviorally expressed controllability (i.e. δ as listed in Table 2) during uncontrollable interaction compared to those without these symptoms (Banker et al., 2022). This group also showed a higher sensitivity to norm violation (higher α), consistent with their clinical symptom of less tolerance to uncertainty. These results suggest that individuals with high misophonia/obsessive-compulsive symptoms have a heightened sense of control despite none actually being afforded to them, as well as aversive reactions when offers do not match expectations (Figure 2c). Taken together, these studies provide two examples of how aberrant computation and perception of social controllability may differentially contribute to different mental health symptoms.

4.3. Broader implications for gamification in digital health and dimensional views of mental health

These studies demonstrate how computational approaches may provide a nuanced view of the complex mapping between clinical symptoms and social deficits in the general population. One promising direction to expand these approaches to broader use is the gamification of these paradigms on digital platforms, which has thus far been primarily adopted in studies on reward and perceptual decision-making (Gillan & Rutledge, 2021; Hauser et al., 2022; Long et al., 2023). By incorporating elements of social cognition in these games (e.g., https://labs.icahn.mssm.edu/thesocialbrainapp), researchers can gather data on how individuals with various mental health conditions encode controllability and form social representations. Finally, gamification using digital health platforms allows for the collection of large datasets not typically available in traditional laboratory settings and enables data-driven exploratory analysis (Wise et al., 2022) of the complex relationship between social behaviors and mental health.

Our proposed paradigm and computational approach fit well within two recent frameworks that aim to develop dimensional approaches to improving the way mental health is studied and classified (Michelini et al., 2021), including the Research Domain Criteria (RDoC) initiative (Insel et al., 2010) and the Hierarchical Taxonomy of Psychopathology (HiTOP) (Kotov et al., 2017). In particular, our approach allows for a more nuanced understanding of the multilevel factors that might contribute to psychopathology in social interactions (e.g., internalized beliefs about controllability; exertion of social control). Future work should consider other areas that can improve the measurement of constructs previously difficult to quantify, like controllability. These types of studies can offer new insights into the underlying neurocomputational mechanisms of social controllability and its impact on mental health, paving the way for the development of more effective diagnoses, interventions, and treatments.

5. Conclusion

The literature reviewed so far identifies a crucial need for developing more sophisticated computational models and ecologically valid paradigms to study controllability in both social and non-social settings. We argue that social controllability affords people with a specific sense of control that is central to mental wellbeing. While existing work has examined how social controllability breaks down in people with high trait delusion or OC-misophonia, these findings are still preliminary and open further questions and opportunities to investigate how social controllability is impacted among individuals with other psychiatric conditions as well as how these aberrant computations might be treated via cognitive, pharmacological, or neuromodulatory interventions. It is also crucial to investigate social controllability across the lifespan to better understand the developmental trajectory of this important mental function in relation to mental health. Finally, the studies reviewed here are nascent examples of computational approaches applied to the construct of social controllability. Future research can build upon this methodological foundation to construct a more comprehensive characterization of the mechanisms involved in social controllability and its impact on mental wellbeing.

Highlights.

• Controllability is generally important for decision-making and mental wellbeing

• Humans exploit the controllability of their social interactions

• Social controllability is examined with economic paradigms and computational models

• Controllability goes awry in conditions like delusion and obsession-compulsion