Social Intelligence & AI Alignment
Understanding how the same abilities that enable cooperation also enable manipulation
Human Social Cognition
Social cognition — emotion perception, theory of mind — lets us understand other minds. My research has shown these abilities are inherently dual-use: the same capabilities that underpin compassion also enable manipulative deception. As AI systems develop social intelligence, this dual-use problem becomes an alignment problem.
Emotion perception and facial recognition form the foundation of human social intelligence—capabilities that enable both empathy and manipulation.
Deep dive: My postdoctoral work on gaze perception explored the neural and computational mechanisms of self-referential social attention — with interactive demos of drift-diffusion modeling, brain network connectivity, and implications for AI social awareness.
Visualizing Multi-Layer Nested Beliefs
This visualization demonstrates the complexity of nested belief states—the foundation of both cooperation and deception in human and artificial intelligence.
From Human Theory of Mind to AI Alignment
Humans with strong theory of mind can choose cooperation or manipulation. AI systems developing ToM capabilities face the same fork — except they can scale.
An AI that models human beliefs at multiple levels could predict and manipulate behavior far more effectively than any individual human. Understanding how these abilities work in people tells us what to watch for in models.
LLM Mentalizing Framework
I am currently building an evaluation suite based on these scenarios and others, to systematically assess LLM social cognition—identifying emergent capabilities relevant to both deceptive capabilities and situational awareness.
Current work: This foundational research on the dual-use nature of ToM now directly informs my Social Reasoning Warden project — a multi-agent framework testing whether LLMs can exploit psychological profiles, and whether monitor agents can defend against it. Preliminary results show ~95% reduction in manipulation success with a warden present.
Current Work: Personality × ToM → Emergent Misalignment
My ERA Fellowship project investigates how personality traits interact with theory of mind capabilities to produce emergent misalignment in multi-agent systems.
Building on Truthful AI's emergent-misalignment research and Anthropic's model psychiatry work, I am constructing AI models with systematically varied trait/ToM profiles to examine behavioral signatures—cooperation, deception, and scheming—under social pressure.
Connection to my human research: My work on personality, theory of mind, and social behavior in humans directly informs predictions about which trait combinations pose the greatest misalignment risks, and which behavioral signatures serve as reliable early warning indicators.