Testing theory of mind in large language models and humans.

At the core of what defines us as humans is the concept of theory of mind: the ability to track other people's mental states. The recent development of large language models (LLMs) such as ChatGPT has led to intense debate about the possibility that these models exhibit behaviour that is indistinguishable from human behaviour in theory of mind tasks. Here we compare human and LLM performance on a comprehensive battery of measurements that aim to measure different theory of mind abilities, from understanding false beliefs to interpreting indirect requests and recognizing irony and faux pas. We tested two families of LLMs (GPT and LLaMA2) repeatedly against these measures and compared their performance with those from a sample of 1,907 human participants. Across the battery of theory of mind tests, we found that GPT-4 models performed at, or even sometimes above, human levels at identifying indirect requests, false beliefs and misdirection, but struggled with detecting faux pas. Faux pas, however, was the only test where LLaMA2 outperformed humans. Follow-up manipulations of the belief likelihood revealed that the superiority of LLaMA2 was illusory, possibly reflecting a bias towards attributing ignorance. By contrast, the poor performance of GPT originated from a hyperconservative approach towards committing to conclusions rather than from a genuine failure of inference. These findings not only demonstrate that LLMs exhibit behaviour that is consistent with the outputs of mentalistic inference in humans but also highlight the importance of systematic testing to ensure a non-superficial comparison between human and artificial intelligences.

Ascribing consciousness to artificial intelligence: human-AI interaction and its carry-over effects on human-human interaction.

The question of whether artificial intelligence (AI) can be considered conscious and therefore should be evaluated through a moral lens has surfaced in recent years. In this paper, we argue that whether AI is conscious is less of a concern than the fact that AI can be considered conscious by users during human-AI interaction, because this ascription of consciousness can lead to carry-over effects on human-human interaction. When AI is viewed as conscious like a human, then how people treat AI appears to carry over into how they treat other people due to activating schemas that are congruent to those activated during interactions with humans. In light of this potential, we might consider regulating how we treat AI, or how we build AI to evoke certain kinds of treatment from users, but not because AI is inherently sentient. This argument focuses on humanlike, social actor AI such as chatbots, digital voice assistants, and social robots. In the first part of the paper, we provide evidence for carry-over effects between perceptions of AI consciousness and behavior toward humans through literature on human-computer interaction, human-AI interaction, and the psychology of artificial agents. In the second part of the paper, we detail how the mechanism of schema activation can allow us to test consciousness perception as a driver of carry-over effects between human-AI interaction and human-human interaction. In essence, perceiving AI as conscious like a human, thereby activating congruent mind schemas during interaction, is a driver for behaviors and perceptions of AI that can carry over into how we treat humans. Therefore, the fact that people can ascribe humanlike consciousness to AI is worth considering, and moral protection for AI is also worth considering, regardless of AI's inherent conscious or moral status.

Predicting the attention of others.

As social animals, people are highly sensitive to the attention of others. Seeing someone else gaze at an object automatically draws one's own attention to that object. Monitoring the attention of others aids in reconstructing their emotions, beliefs, and intentions and may play a crucial role in social alignment. Recently, however, it has been suggested that the human brain constructs a predictive model of other people's attention that is far more involved than a moment-by-moment monitoring of gaze direction. The hypothesized model learns the statistical patterns in other people's attention and extrapolates how attention is likely to move. Here, we tested the hypothesis of a predictive model of attention. Subjects saw movies of attention displayed as a bright spot shifting around a scene. Subjects were able to correctly distinguish natural attention sequences (based on eye tracking of prior participants) from altered sequences (e.g., played backward or in a scrambled order). Even when the attention spot moved around a blank background, subjects could distinguish natural from scrambled sequences, suggesting a sensitivity to the spatial-temporal statistics of attention. Subjects also showed an ability to recognize the attention patterns of different individuals. These results suggest that people possess a sophisticated model of the normal statistics of attention and can identify deviations from the model. Monitoring attention is therefore more than simply registering where someone else's eyes are pointing. It involves predictive modeling, which may contribute to our remarkable social ability to predict the mind states and behavior of others.

Fundamental principles of cortical organization reflected in a new study.

In an exciting new finding by Gordon et al., the human motor cortex does not show a traditional body map. Instead, body-part-specific zones are separated by integrative zones, possibly reflecting the statistical structure of the human motor repertoire.

Conscious intention: New data on where and how in the brain.

How do we decide to act, and how do those decisions relate to conscious choice? A new study helps dissociate the neuronal mechanisms that choose, prepare, and trigger movement from our explicit reports of conscious intention.

Right temporoparietal junction encodes inferred visual knowledge of others.

When people make inferences about other people's minds, called theory of mind (ToM), a cortical network becomes active. The right temporoparietal junction (TPJ) is one of the most consistently responsive nodes in that network. Here we used a pictorial, reaction-time, ToM task to study brain activity in the TPJ and other cortical areas. Subjects were asked to take the perspective of a cartoon character and judge its knowledge of a visual display in front of it. The right TPJ showed evidence of encoding information about the implied visual knowledge of the cartoon head. When the subject was led to believe that the head could see a visual change take place, activity in the right TPJ significantly reflected that change. When the head could apparently not see the same visual change take place, activity in the right TPJ no longer significantly reflected that change. The subject could see the change in all cases; the critical factor that affected TPJ activity was whether the subject was led to think the cartoon character could see the change. We also found that whether the beliefs attributed to the cartoon head were true or false did not significantly affect activity in the present paradigm. These results suggest that the right TPJ may play a role in modeling the contents of the minds of others, perhaps more than it participates in evaluating the truth or falsity of that content.

A conceptual framework for consciousness.

This article argues that consciousness has a logically sound, explanatory framework, different from typical accounts that suffer from hidden mysticism. The article has three main parts. The first describes background principles concerning information processing in the brain, from which one can deduce a general, rational framework for explaining consciousness. The second part describes a specific theory that embodies those background principles, the Attention Schema Theory. In the past several years, a growing body of experimental evidence-behavioral evidence, brain imaging evidence, and computational modeling-has addressed aspects of the theory. The final part discusses the evolution of consciousness. By emphasizing the specific role of consciousness in cognition and behavior, the present approach leads to a proposed account of how consciousness may have evolved over millions of years, from fish to humans. The goal of this article is to present a comprehensive, overarching framework in which we can understand scientifically what consciousness is and what key adaptive roles it plays in brain function.

Consciousness is already solved: The continued debate is not about science.

A logical explanation of consciousness has been known for decades. The brain must construct a specific set of information about conscious feeling (theory-of-mind information), causing people to believe, think, and claim to have consciousness. Theories that propose an actual, intangible feeling are non-explanatory. They add a magical red herring while leaving unexplained the objective phenomena: the believing, thinking, and claiming.

The origin of smiling, laughing, and crying: The defensive mimic theory.

Why do we leak lubricant from the eyes to solicit comfort from others? Why do we bare our teeth and crinkle our faces to express non-aggression? The defensive mimic theory proposes that a broad range of human emotional expressions evolved originally as exaggerated, temporally extended mimics of the fast, defensive reflexes that normally protect the body surface. Defensive reflexes are so important to survival that they cannot be safely suppressed; yet they also broadcast information about an animal's internal state, information that can potentially be exploited by other animals. Once others can observe and exploit an animal's defensive reflexes, it may be advantageous to the animal to run interference by creating mimic defensive actions, thereby manipulating the behaviour of others. Through this interaction over millions of years, many human emotional expressions may have evolved. Here, human social signals including smiling, laughing and crying, are compared component-by-component with the known, well-studied features of primate defensive reflexes. It is suggested that the defensive mimic theory can adequately account for the physical form of not all, but a large range of, human emotional expression.

The attention schema theory in a neural network agent: Controlling visuospatial attention using a descriptive model of attention.

In the attention schema theory (AST), the brain constructs a model of attention, the attention schema, to aid in the endogenous control of attention. Growing behavioral evidence appears to support the presence of a model of attention. However, a central question remains: does a controller of attention actually benefit by having access to an attention schema? We constructed an artificial deep Q-learning neural network agent that was trained to control a simple form of visuospatial attention, tracking a stimulus with an attention spotlight in order to solve a catch task. The agent was tested with and without access to an attention schema. In both conditions, the agent received sufficient information such that it should, theoretically, be able to learn the task. We found that with an attention schema present, the agent learned to control its attention spotlight and learned the catch task. Once the agent learned, if the attention schema was then disabled, the agent's performance was greatly reduced. If the attention schema was removed before learning began, the agent was impaired at learning. The results show how the presence of even a simple attention schema can provide a profound benefit to a controller of attention. We interpret these results as supporting the central argument of AST: the brain contains an attention schema because of its practical benefit in the endogenous control of attention.

Attention, awareness, and the right temporoparietal junction.

The attention schema theory posits a specific relationship between subjective awareness and attention, in which awareness is the control model that the brain uses to aid in the endogenous control of attention. In previous experiments, we developed a behavioral paradigm in human subjects to manipulate awareness and attention. The paradigm involved a visual cue that could be used to guide attention to a target stimulus. In task 1, subjects were aware of the cue, but not aware that it provided information about the target. The cue measurably drew exogenous attention to itself. In addition, implicitly, the subjects' endogenous attention mechanism used the cue to help shift attention to the target. In task 2, subjects were no longer aware of the cue. The cue still measurably drew exogenous attention to itself, yet without awareness of the cue, the subjects' endogenous control mechanism was no longer able to use the cue to control attention. Thus, the control of attention depended on awareness. Here, we tested the two tasks while scanning brain activity in human volunteers. We predicted that the right temporoparietal junction (TPJ) would be active in relation to the process in which awareness helps control attention. This prediction was confirmed. The right TPJ was active in relation to the effect of the cue on attention in task 1; it was not measurably active in task 2. The difference was significant. In our interpretation, the right TPJ is involved in an interaction in which awareness permits the control of attention.

What makes us so certain that we're conscious?

In the attention schema theory (AST), having an automatically constructed self-model that depicts you as containing consciousness makes you intuitively believe that you have consciousness. The reason why such a self-model evolved in the brains of complex animals is that it serves the useful role of modeling, and thus helping to control, the powerful and subtle process of attention, by which the brain seizes on and deeply processes information.

Visual motion assists in social cognition.

Recent evidence suggests a link between visual motion processing and social cognition. When person A watches person B, the brain of A apparently generates a fictitious, subthreshold motion signal streaming from B to the object of B's attention. These previous studies, being correlative, were unable to establish any functional role for the false motion signals. Here, we directly tested whether subthreshold motion processing plays a role in judging the attention of others. We asked, if we contaminate people's visual input with a subthreshold motion signal streaming from an agent to an object, can we manipulate people's judgments about that agent's attention? Participants viewed a display including faces, objects, and a subthreshold motion hidden in the background. Participants' judgments of the attentional state of the faces was significantly altered by the hidden motion signal. Faces from which subthreshold motion was streaming toward an object were judged as paying more attention to the object. Control experiments showed the effect was specific to the agent-to-object motion direction and to judging attention, not action or spatial orientation. These results suggest that when the brain models other minds, it uses a subthreshold motion signal, streaming from an individual to an object, to help represent attentional state. This type of social-cognitive model, tapping perceptual mechanisms that evolved to process physical events in the real world, may help to explain the extraordinary cultural persistence of beliefs in mind processes having physical manifestation. These findings, therefore, may have larger implications for human psychology and cultural belief.

Attention control and the attention schema theory of consciousness.

In the attention schema theory (AST), the brain constructs a schematic, simplified model of attention. The model is associated with three cognitive processes: a model of one's own attention contributes to the endogenous control of attention, a model of the attention of others contributes to theory of mind, and the contents of these models leads to the common human claim that we contain a non-physical consciousness or awareness inside us. Because AST is a control-engineering style theory, it can make specific predictions in complex situations. Here, over six experiments, we examined interactions between attention and awareness to test predictions of AST. Participants performed a visual task in which a cue stimulus affected their attention, as measured by their reactions to a subsequent target stimulus. The task measured both exogenous attention drawn to the cue and endogenous attention directed to a target location predicted by the cue. When participants were not aware that the cue predicted the target, both exogenous and endogenous attention effects remained. In contrast, when participants were not visually aware of the cue itself, the exogenous attention effect remained and the endogenous effect was impaired. In an additional two experiments, when participants learned an implicit shift of attention, the learning generalized from trained spatial locations to adjacent, untrained locations. Each of these findings matched predictions of AST. The results support the interpretation that attention control relies partly on an internal model that is responsible for claims of awareness.

Consciousness and the attention schema: Why it has to be right.

This article describes some aspects of the underlying logic of the attention schema theory (AST) of subjective consciousness. It is a theory that distinguishes between what the brain actually, physically has, what is represented by information models constructed in the brain, what higher cognition thinks based on access to those models and what speech machinery claims based on the information within higher cognition. It is a theory of how we claim to have an essentially magical, subjective mind, based on the impoverishment and reduction of information along that pathway. While the article can stand on its own as a brief account of some critical aspects of AST, it specifically addresses questions and concerns raised by a set of commentaries on a target article.

Other people's gaze encoded as implied motion in the human brain.

Keeping track of other people's gaze is an essential task in social cognition and key for successfully reading other people's intentions and beliefs (theory of mind). Recent behavioral evidence suggests that we construct an implicit model of other people's gaze, which may incorporate physically incoherent attributes such as a construct of force-carrying beams that emanate from the eyes. Here, we used functional magnetic resonance imaging and multivoxel pattern analysis to test the prediction that the brain encodes gaze as implied motion streaming from an agent toward a gazed-upon object. We found that a classifier, trained to discriminate the direction of visual motion, significantly decoded the gaze direction in static images depicting a sighted face, but not a blindfolded one, from brain activity patterns in the human motion-sensitive middle temporal complex (MT+) and temporo-parietal junction (TPJ). Our results demonstrate a link between the visual motion system and social brain mechanisms, in which the TPJ, a key node in theory of mind, works in concert with MT+ to encode gaze as implied motion. This model may be a fundamental aspect of social cognition that allows us to efficiently connect agents with the objects of their attention. It is as if the brain draws a quick visual sketch with moving arrows to help keep track of who is attending to what. This implicit, fluid-flow model of other people's gaze may help explain culturally universal myths about the mind as an energy-like, flowing essence.