Theory of Mind (ToM) is the ability to reason about other people's mental states [1]. It is also commonly referred to as perspective-taking or mentalising and has a central role in human social interaction. Much of human communication is not explicitly verbalised [2]. Instead, people use their abilities to infer others' beliefs, intentions, and feelings to make sense of social events.
Since 1978, when Premack and Woodruff coined the term, research has greatly evolved. There are various theories explaining this ability and the application of the concept has vastly expanded in recent years. There is now research on the development, social relevance, influencing factors, cultural differences, individual differences, and neuroscience of ToM.
Beyond psychological research, Theory of Mind has increasingly been studied in the field of Artificial Intelligence (AI). In the context of the development of AI that fits into human society, the demand for social intelligence in artificial agent grows [3]. Currently, the implementation of typical human traits in artificial intelligence is exploding as a research field.
However, there is a fundamental difference in perspective on intelligence and behaviour from psychological and computational research. The two fields are highly complementary and have the potential to greatly inform each other but there is limited exchange of concepts across them. For example, there is evidence that humans use their ToM abilities very selectively [4]. Computational models of ToM, in contrast, often are built on the assumption that we create rich models of people all the time [5]. In a world as complex as ours, this strategy would be impossible to realise. The heuristics and shortcuts the brain uses to perform ToM as flexibly as psychologists have observed are largely understudied and overlooked in the field of Social AI.
In this paper we discuss ToM at the intersection of Psychology and Computer Science. We illustrate the complexity of ToM and highlight the challenges in generating computational models. We then present an experimental study exploring human ToM use to understand the heuristic mechanisms underlying ToM, towards developing a more efficient computational model. The results suggest a strong influence of egocentric information in predicting others' behaviour. The possible ramifications for computational models are discussed.
This section considers the question of Theory of Mind reasoning from both a psychological perspective and a computational standpoint.
The field of ToM is vast. Most work is developmental research studying if, when, and how ToM develops in young children. The focus of the present work, however, is on adults, assuming a typically functioning ToM system. The emergence of different theories of ToM have led to a variety of understandings of what the ability entails. Generally, ToM is understood as the ability to reason about others' mental states. However, where this ability starts and ends is not clearly defined [6]. Furthermore, it has been argued that ToM tasks measure different aspects of ToM [7].
Ultimately, ToM is an umbrella term for different phenomena. A person's model of another human could consist of any of, and is not limited to, the following:
• explaining a person's behaviour by modelling their goals and/or moral values • predicting a person's behaviour by modelling their goals and/or moral values • modelling a person's beliefs by interpreting their behaviour • inferring a person's intentions by considering one's own experience • deciding on one's own behaviour by modelling another person Moreover, mental models can vary greatly in accuracy and depth. The ability to use ToM is also characterised by a large variance in ToM use across individuals and situations [8].
Psychological research suggests that humans do not always infer others' mental states, even when it would be helpful to do so [4]. Studies of cultural differences in ToM suggest an element of individual learning and experience in shaping those tendencies [9]. Furthermore, there is evidence that the activity of another person, as opposed to them being passive, facilitates perspective-taking [10]. Moreover, the nature of the other appears to be relevant for ToM use, i.e. whether it is a human or an inanimate object [11]. Research also suggests that mood and emotion are related to the extent to which an individual may be egocentric in their perspective [12].
Furthermore, researchers have distinguished between different types of ToM. The distinction between explicit and implicit ToM [13], for example, is in line with other dual-system approaches to psychological phenomena [14,15]. Proponents of these accounts generally distinguish between a fast system and a slow system. The fast system uses implicit and automatic processing, based on previously established beliefs and reaction patterns, which we are not consciously aware of. It is characterised by heuristics and biases, which make it fast but also difficult to intervene with or execute direct control over. The slow system is believed to be deliberate, operating within the scope of our conscious awareness.
Apperly and Butterfill [13] strongly argue in favour of a dual-system account of mindreading, suggesting that the fast and implicit component of Theory of Mind develops early, whereas the explicit, slow, and more cognitively demanding element develops only later-on. They propose that young children fail at the explicit but not the implicit mind-reading skills [16]. Some ToM research focusses primarily on explicit ToM and some primarily on implicit ToM, whereas other researchers discount the distinction altogether. Burge [17], for example, argues that there are very powerful other explanations for the evidence suggesting Theory of Mind in very young children, which would not require mind-reading, such as learning based on associations or reinforcement.
The focus of the present work is on explicit ToM. This is not because the work argues against implicit ToM requires perceptual measures [18], which are not applied here. Instead, this work specialises on the explicit content of mental models and the computations by which it is selected. Subsequent sections of this paper will, however, suggest a more complex relationship between implicit and explicit processes. Specifically, implicit processes may shape a person's explicit interpretations of behaviour.
In response to increasing demands of AI to complete social tasks [19], especially in collaboration with other humans, multi-agent research has explored a range of techniques for agents to model other agents. These include but are not limited to the use of action observations to reconstruct action strategies, modelling group relations to predict agents' decision making, or using hierarchical action descriptions to determine plans and goals. See Albrecht & Stone [20] for a detailed review. Here we briefly cover the work relevant for the current study.
With growing insight that the brain operates like a "prediction machine" [21], researchers have increasingly used Bayesian approaches to model ToM. Baker and colleagues [5], for example, developed the Bayesian Theory of Mind (BToM), which is based on probabilistic backward inferences. The model computes which mental state is most likely given the observed behaviour. BToM accurately represents human responses on different tasks. However, the domains are very restricted by the design of the study. Instead of forcing a flexible reaction to a rich environment or situation, the search space at hand is tailored to the model from the start. Models of others have been described as higher-order beliefs [22], i.e. beliefs about beliefs, with varying levels of depth. Computationally they have been represented as nested beliefs with a recursive structure [23]. Applying this principle, for example, PsychSim [24,25] is a system which models multiagent sequential decision-making tasks. IPOMDPs (interactive partially observable Markov decision processes) generate forward inferences to predict others' behaviours.
Especially when several levels of depth are involved, computing all possible behaviours a person may engage in (considering several steps of projection into the future and a variety of possible events or behaviour by others), to then choose the most likely option, can be extremely intractable. Considering that humans operate in very complex environments, it seems highly unlikely that a person could realise this mechanism explicitly, given short-term memory constraints, at least to the full extent of searching through all possible mental states. This therefore poses the question when exactly human ToM is performed and to what degree. What are the computational shortcuts human brains take to speed up processing?
In order to create a realisable computational model of ToM it is required to work around this problem of tractability. Towards that end, we can ask what decides when humans use ToM, how they use it, as well as when they do not use it. What features of the situation, experience and social interaction determine the activation of human mentalising abilities?
The potentially ToM-relevant feature illuminated in this work is the presence and timing of an egocentric cue. Zeng et al. [26] present a neuroscience-informed model, recognising the different brain areas that have been shown to be involved in ToM. They distinguish between perspective-type and belief-type information impacting on ToM use. Thereby, the perspective is an early element in mentalising, composed of information from lower-level brain areas, whereas beliefs are influential later-on, and the product of higher-level processing (Figure 1). This model is consistent with dual process approaches to cognition, distinguishing between more automatic, fast, less controlled processes and more deliberate, slow, and conscious processes, and with models which distinguish between implicit and explicit ToM.
Zeng et al. [26] distinguish between the self-experience learning pathway, the motivation understanding pathway, the reasoning about one's own belief pathway, and the reasoning about other people's belief pathway. Based on their model, the present study seeks to distinguish between subjects use of previously established higher-level ToM beliefs about the observed agent knowledge and low-level egocentric cues available to the participant. Specifically, participants are asked to predict the behavior of agent looking for a misplaced book that subjects are told is hidden from the agent. The experiment manipulates whether and when the subject gets a cue as to the book's location. Thus the present study distinguishes between previously established beliefs and momentarily available egocentric cues. The following hypotheses were proposed:
H1: With no visual cue present, there is no egocentric bias (belief unbiased and perspective unbiased).
H2: With an early and partial visual cue present (belief biased but perspective unbiased), there is no egocentric bias.
H3: With a full visual cue present, there is an egocentric bias towards the target direction (belief biased and perspective biased).
The results of this study will show whether the distinction between lower-level and higherlevel influences is helpful in modelling ToM.
This study was approved by the ethics board of the Psychology department at the University of Glasgow. 60 participants were recruited with the online platform Prolific. They were from the United Kingdom with a 95% approval rate on Prolific and at least 10 previously completed studies. 21(35%) of the participants were male, 38(63.33%) female, and 1(1.67%) other. The mean age was 39.43(SD=14.90) years with a minimum of 19 and a maximum of 67.
The maze videos used in this study are created with the tool used by Pöppel et al. [27]. A virtual agent moved along a pre-determined trajectory through a 2D space, searching for a book (see figures 2-4). Mazes were designed to rule out asymmetries in orientation as confounding variables. The target direction (egocentric cue present) was kept constant as the red option.
Videos were 10 seconds long, with the first static frame with the agent at its starting point and instructions (and cue if applicable) shown for 3 seconds, the agent remaining at the starting point without instructions (and cue in the second condition) for 1 second, and finally the agent moving around the space for 6 seconds.
The study had a 1x3 between-groups design with the independent variable cue visibility (no cue vs partial cue vs full cue). Figures 2-4 show snapshots of the different conditions. The dependent variable direction was measured on a nominal scale with the categories red, black, and "I don't know". Another dependent variable explanation was measured qualitatively by asking each participant why they chose their respective response.
The target stimulus included a short video of the mazes, with the cue visibility according to the condition participants were in. They were told that the agent was looking for the book and saw it move through the maze until it stopped at the bottom. At this junction where it turned either left or right, the participant was asked what they thought where the agent would go next. Thereby, one direction was marked in red, the other in black. Participants selected their answer from three choices: black, red, or "I don't know" (Figure 5).
The potentially confounding variables maze orientation (left vs right) and colour (black & red) were randomised for all conditions.
None of the demographics were found to significantly affect participant responses.
Figure 6 shows which direction participants predicted the agent to go. When no cue was present, they equally chose between red and black, whereas the cue invoked favouring the red direction, regardless of when it was shown. A Chi-Square analysis was conducted and shows that the difference between the three groups is not significant, X2(4, N=60) = 7.76, p = 0.101.
We asked participants why they chose the answer they selected. As depicted in the framework by Zeng et al. [26], qualitative responses were analysed with regards to the distinctions between faster, more intuitive lower-level (perspective) pathways vs slower, more rational higher-level (belief) pathways, and self (egocentric) vs other (altercentric).
1. EP = Egocentric Perspective (self-experience learning, i.e. own feeling, intuition, etc.) "it just felt right" "I always read and choose left to right" 2. AP = Altercentric Perspective (motivation understanding, i.e. prediction of other's actions based on the experience of their previous actions) "Because he kept turning right on the search" "it always went to its left" 3. EB = Egocentric Belief (reasoning about one's own belief) "closest to the book" "It was where the book was"
"There is no way to know" "because it's a 50/50 chance, I don't know"
We analysed whether participants in the different cue visibility conditions chose different strategies to predict the agent's next move. The Chi-Square test shows that they did, X2(6, N=60) = 12.84, p = 0.046. There was no significant difference by cue visibility in the source of information (belief vs perspective) for participants' prediction of where the agent would go, X2(2, N=60) = 0.45, p = 0.798. (Figure 7). Interestingly, perspective-based reasoning was more common than belief-based reasoning in all three conditions. Participants did, however, differ significantly in what viewpoint they took depending on whether they saw a cue or not, X2(2, N=60) = 10.85, p = 0.004 (Figure 8). When no cue was visible, participants largely considered the agent's perspective (A) but when a cue was there, the own perspective (E) dominated, regardless of whether the cue was only shown early or throughout the video.
It was hypothesised that there is an egocentric bias with a visual cue fully present (H3) and no egocentric bias when the cue is only partially present (H2) or absent (H1). The descriptive statistics suggest an egocentric bias in both the early cue and full cue condition and not in the no cue condition (Figure 6). Statistical analyses do not support the significance of these differences. However, the result is close to significance and possibly underpowered with only 20 participants in each condition. We therefore recommend to follow this study up with a larger sample size. The result did not reach significance, but the trends are consistent with the phenomenon shown in previous literature that egocentric information is very dominant in social interactions (Keysar et al., 2003).
Analyses of participants' verbal statements show that whether participants based their choices on belief-based or perspective-based explanations did not differ across conditions (Figure 7). Perspective-based reasoning was more common than belief-based reasoning in all three conditions. Zeng et al. [26] suggest that this pathway is faster, which suggests that it may be favoured by the brain to save costs and energy. However, the dual process approach applied here is not without its criticism. While the framework has gained a lot of popularity, researchers have also highlighted its limitations. Evidence suggests that many psychological findings require more interaction between the two respective processes than the theory suggests. Researchers have argued that neither "fast" nor "slow" ToM can occur in isolation of the other [16]. This highlights the difficulty in focussing on either explicit or implicit elements of ToM, as mentioned earlier. The present paradigm and results demonstrates the variety of influences on ToM and the difficulties in untangling them.
There are significant differences in egocentrism vs altercentrism across groups, supporting hypotheses 1 (no cue) and 3 (full cue), but not hypothesis 2 (early cue). Whether the cue was shown throughout the video or only early-on, participants largely considered their own knowledge rather than the other's. When there was no cue, the agent's knowledge was considered more (Figure 8). This suggests a high sensitivity to egocentric information in the context of ToM, regardless of when it is shown, consistent with previous literature [4]. The results suggest a preference for egocentrism over perspective-taking regardless of when an egocentric cue is available.
It is noteworthy that the explanations of participants' choices differed significantly across conditions while the decisions themselves did not differ significantly. However, as mentioned above, the quantitative comparison was not far from reaching the significance level.
A computational model based on the data and observations presented above is currently in development. The model will conceptualise the switch between whose beliefs, actions, and goals are cognitively represented at all and/or drawn upon for reasoning about another's mental states. A key goal for the model is to explore ways to constrain the recursive aspects of ToM that require predictions of predictions (Figure 9). The deeper the level of ToM, the more possibilities need to be considered when a prediction is generated (Figure 10). This computation is even more complex in a sequential decision-making setting over multiple steps in time. The present data suggests a bias towards using lower-order information even when next-order information is available. For example, even when it was possible for them to reason about the agent's beliefs or previous actions, many participants used their own tendencies or strategies to predict the agent's actions. Interestingly, there is also tendency by participants to make at least some prediction rather than indicating that they did not know where the agent would go.
A key implication for modelling ToM is therefore that the distinction between a fast and a slow system appears less fundamental than considering the information a human has previously gathered. Participant responses greatly reflect their own prior experiences. This highlights the inherent relevance of subjective elements, such as priming effects or biases. Interestingly, these have not been applied in a majority of ToM models to date [20]
ToM use largely varies across situations and individuals. Like with other aspects of cognition, humans develop strategies and habits which can differ considerably from one person to the next. For example, there are large differences in societal expectations across the world [28], and languages differ in the extent to which they include the vocabulary needed to communicate ToM concepts [29]. Evidence suggests that factors such as mood contribute to the extent of perspective-taking. For instance, uncertainty has been associated with more self-focussed reasoning [12].
All people are biased in their beliefs and perspectives, shaped by their own experiences. These biases are at the core of the very mechanism which makes human cognition so fast and efficient. But the speed and efficiency of human cognition doesn't mean that it is flawless. Biases come at the cost of errors which are often very difficult to detect [30]. This highlights an important question to consider as part of this research: Are the mechanisms required to develop robust, efficient, and dynamic artificial ToM worth the cost of the errors that are a fundamental part of human cognition?
On the other hand, considering the biases and flaws that are part of human ToM is fundamental to AI successfully and accurately predicting human behaviour. Furthermore, these insights may assist current efforts to reduce biases and teach critical, reflected thinking in social contexts. Working beyond the biases in ToM may result in strategies helpful to both humans and artificial agents in understanding and managing social situations. Modelling human strengths and weaknesses is critical to realising AI powered assistive technology, particularly effective human-AI teamwork.
This study has indicated the extent to which 1) people can employ very different reasoning to come to the same conclusion and 2) the value of qualitative rather than quantitative research to identify these subtle differences. Rather than thinking about ToM as a binary ability, which is either present or absent, the present results show the possible variance in mental model complexity. We suggest that mental model depth and complexity is an important factor contributing to flexible and efficient ToM use. Future research will need to explore more deeply how mental models are structured, how or when the depth of mental models changes, and how exactly more slow, elaborate ToM differs from faster, biased ToM.
The authors are especially grateful to Jan Pöppel, Bielefeld University, Germany. This study builds on his research on Theory of Mind and his maze tool, ideas for designing the study, and general advice have been most helpful for collecting the data presented here.
⋆ This work was supported by the UKRI Centre for Doctoral Training in Socially Intelligent Artificial Agents, Grant Number EP/S02266X/1