Knowing me, knowing you: theory of mind in AI

Fabio Cuzzolin, Alice Morelli, Bogdan-Ionuţ Cîrstea, Barbara J. SahakianPsychological 2020 · DOI · OpenAlex

Abstract

Artificial intelligence has dramatically changed the world as we know it, but is yet to fully embrace 'hot' cognition, i.e., the way an intelligent being's thinking is affected by their emotional state. Artificial intelligence encompassing hot cognition will not only usher in enhanced machine-human interactions, but will also promote a much needed ethical approach. Theory of Mind, the ability of the human mind to attribute mental states to others, is a key component of hot cognition. To endow machines with (limited) Theory of Mind capabilities, computer scientists will need to work closely with psychiatrists, psychologists and neuroscientists. They will need to develop new models, but also to formally define what problems need to be solved and how the results should be assessed.

Introduction: The Cold Reality of Modern AI

Current artificial intelligence excels at "cold cognition"—the domain of pure logic, pattern recognition, and closed-system optimization. Systems like AlphaGo demonstrate superhuman performance in environments with fixed rules. However, these systems fundamentally lack "hot cognition," or the ability to understand how emotional states and social contexts influence behavior. Central to this limitation is the absence of a "Theory of Mind" (ToM): the capacity to attribute mental states—such as beliefs, desires, and intentions—to oneself and others. For AI to transition from specialized tools to safe, effective partners in human environments, it must bridge this cognitive gap.

The "Ice Cream Van Problem": Why Observation is Not Enough

The failure of purely observational AI is best illustrated by the "ice cream van problem." Consider an autonomous vehicle observing a child on a sidewalk. A standard predictive model might extrapolate the child's path based on their current velocity and historical motion data. However, if an ice cream van appears across the street, the child’s internal motivation changes instantaneously.

Without a Machine Theory of Mind, the AI cannot infer this shift in mental state. It sees only the physical coordinates, not the desire for a treat. This reliance on "naive pattern recognition" is a primary factor in safety lapses within complex, human-centric environments. As the authors argue, it is mathematically impossible to recover a human’s true preferences from behavior alone without making foundational assumptions about their internal rationality and psychology.

Theoretical Foundations: Theory-Theory vs. Simulation-Theory

To address this, the paper proposes a roadmap inspired by two primary psychological frameworks:

  1. Theory-Theory (TT): This suggests that humans develop "scientific" theories about the minds of others, using logical rules to predict behavior.
  2. Simulation-Theory (ST): This posits that we understand others by using our own mental processes to simulate their perspective, a concept often linked to "mirror neurons" in the human brain.

The authors advocate for a hybrid approach. Rather than attempting to "brute-force" social intelligence through massive datasets, AI should be built using modular "blocks" of artificial ToM. These blocks would represent logical relations between mental states (e.g., the link between anger and aggression). By assembling these pre-learned modules, an AI could create a "bespoke simulation" of a specific individual in a specific context, allowing for more nuanced and accurate predictions.

The Data Bottleneck and the Efficiency Gap

A significant hurdle in this evolution is the lack of standardized benchmarks. While large-scale video datasets exist, they generally lack "mental state annotations"—labels that describe the underlying emotions or intentions of the subjects.

The quantitative disparity between human and machine learning is stark. To reach the goal-inference capabilities of a six-month-old infant, current deep learning models require approximately 32 million samples. In practical terms, an AI would need the equivalent of 175,000 labeled demonstrations every day for six months to match a human baby. This data inefficiency highlights the need for architectural changes rather than simply more data.

Conclusion: A Multi-Disciplinary Roadmap

Endowing machines with Theory of Mind is not a task for computer scientists alone. The path forward requires a "virtuous cycle" involving psychiatrists, psychologists, and neuroscientists. This collaboration is essential to define the formal problems of social cognition and to develop evaluation metrics that go beyond simple prediction accuracy. Moving toward "hot cognition" is not just a technical challenge; it is a prerequisite for creating AI that is truly ethical, safe, and integrated into human society.

Summary updated

Abstract

Artificial intelligence has dramatically changed the world as we know it, but is yet to fully embrace 'hot' cognition, i.e., the way an intelligent being's thinking is affected by their emotional state. Artificial intelligence encompassing hot cognition will not only usher in enhanced machine-human interactions, but will also promote a much needed ethical approach. Theory of Mind, the ability of the human mind to attribute mental states to others, is a key component of hot cognition. To endow machines with (limited) Theory of Mind capabilities, computer scientists will need to work closely with psychiatrists, psychologists and neuroscientists. They will need to develop new models, but also to formally define what problems need to be solved and how the results should be assessed.
BibTeX
@inproceedings{cuzzolin2020knowing,
  author = {Fabio Cuzzolin and Alice Morelli and Bogdan-Ionuţ Cîrstea and Barbara J. Sahakian},
  title = {{Knowing me, knowing you: theory of mind in AI}},
  booktitle = {Psychological},
  year = {2020},
  doi = {10.1017/s0033291720000835},
}
Figure