As AI systems are increasingly integrated into from virtual assistants and customer service agents to counseling and AI companions, an understanding of social neural dynamics is essential for both scientific and technological progress. A new study from UCLA researchers shows biological brains and AI systems develop remarkably similar neural patterns during social interaction.

The study, recently published in the journal Nature, reveals that when mice interact socially, specific brain cell types create synchronize in “shared neural spaces,” and artificial intelligence agents develop analogous patterns when engaging in social behaviors.     

The new research represents a striking convergence of neuroscience and artificial intelligence, two of today’s most rapidly advancing fields. By directly comparing how biological brains and AI systems process social information, scientists can now better understand fundamental principles that govern social cognition across different types of intelligent systems. The findings could advance understanding of social disorders like autism while simultaneously informing the development of more sophisticated, socially  aware AI systems.  

This work was supported in part by , the National Science Foundation, the Packard Foundation, Vallee Foundation, Mallinckrodt Foundation and the Brain and Behavior Research Foundation.

Examining AI agents’ social behavior

A multidisciplinary team from UCLA’s departments of neurobiology, biological chemistry, bioengineering, electrical and computer engineering, and computer science across the David Geffen School of Medicine and UCLA Samueli School of Engineering used advanced brain imaging techniques to record activity from molecularly defined neurons in the dorsomedial prefrontal cortex of mice during social interactions. The researchers developed a novel computational framework to identify high-dimensional “shared” and “unique” neural subspaces across interacting individuals. The team then trained artificial intelligence agents to interact socially and applied the same analytical framework to examine neural network patterns in AI systems that emerged during social versus non-social tasks.

The research revealed striking parallels between biological and artificial systems during social interaction. In both mice and AI systems, neural activity could be partitioned into two distinct components: a “shared neural subspace” containing synchronized patterns between interacting entities, and a “unique neural subspace” containing activity specific to each individual.

Remarkably, GABAergic neurons — inhibitory brain cells that regulate neural activity —showed significantly larger shared neural spaces compared with glutamatergic neurons, which are the brain’s primary excitatory cells. This represents the first investigation of inter-brain neural dynamics in molecularly defined cell types, revealing previously unknown differences in how specific neuron types contribute to social synchronization.

When the same analytical framework was applied to AI agents, shared neural dynamics emerged as the artificial systems developed social interaction capabilities. Most importantly, when researchers selectively disrupted these shared neural components in artificial systems, social behaviors were substantially reduced, providing the direct evidence that synchronized neural patterns causally drive social interactions.

The study also revealed that shared neural dynamics don’t simply reflect coordinated behaviors between individuals, but emerge from representations of each other’s unique behavioral actions during social interaction.

“This discovery fundamentally changes how we think about social behavior across all intelligent systems,” said Weizhe Hong, professor of neurobiology, biological chemistry and bioengineering at UCLA and lead author of the new work. “We’ve shown for the first time that the neural mechanisms driving social interaction are remarkably similar between biological brains and artificial intelligence systems. This suggests we’ve identified a fundamental principle of how any intelligent system — whether biological or artificial — processes social information. The implications are significant for both understanding human social disorders and developing AI that can truly understand and engage in social interactions.”

Continuing research for treating social disorders and training AI

The research team plans to further investigate shared neural dynamics in different and potentially more complex social interactions. They also aim to explore how disruptions in shared neural space might contribute to social disorders and whether therapeutic interventions could restore healthy patterns of inter-brain synchronization. The artificial intelligence framework may serve as a platform for testing hypotheses about social neural mechanisms that are difficult to examine directly in biological systems. They also aim to develop methods to train socially intelligent AI.

The study was led by UCLA’s Hong and Jonathan Kao, associate professor of electrical and computer engineering. Co-first authors Xingjian Zhang and Nguyen Phi, along with collaborators Qin Li, Ryan Gorzek, Niklas Zwingenberger, Shan Huang, John Zhou, Lyle Kingsbury, Tara Raam, Ye Emily Wu and Don Wei contributed to the research.

If someone offers to make an AI video recreation of your wedding, just say no. This is the tough lesson I learned when I started trying to recreate memories with Google’s Gemini Veo model. What started off as a fun exercise ended in disgust.

I grew up in the era before digital capture. We took photos and videos, but most were squirreled away in boxes that we only dragged out for special occasions. Things like the birth of my children and their earliest years were caught on film and 8mm videotape.