You’ve seen the headlines. The newest AI can write like a scholar, argue like a lawyer, and code better than most engineers. Yet none of them have ever looked another mind in the eye and understood what it felt like to be there.
Not a single large language model has experienced the silent, shimmering moment when two brains sync up—a baby and caregiver locked in mutual gazing, neurons firing in loose harmony, a joint awareness blooming like breath on cold glass. They can parse the word “connection,” dissect its etymology, and generate five paragraphs on its importance in human development. But they’ve never lived it.
We built brilliant solitary thinkers. We scaled attention mechanisms into the billions, threw ever-larger corpora at transformer heads, and fine-tuned reward models to chase alignment across increasingly abstract dimensions. What we haven’t built—or even properly tried—is intelligence that emerges in the gap between agents. Not between a human and an AI, but between two agents, each with its own history, goal, and bias, learning to coordinate despite noise, misalignment, and the sheer messiness of real-time interaction.
Social intelligence isn’t a feature you bolt onto an already intelligent agent—like a social module you install post-hoc. It’s more like the substrate itself changing: intelligence becomes possible only when systems move beyond prediction and into relational coordination. That shift isn’t glamorous. It’s under-rated in research dashboards and often buried beneath more compelling metrics: perplexity, latency, accuracy. But what neuroscience—and now, emerging work in multi-agent reinforcement learning—is revealing is a fundamental truth: cooperative intelligence only emerges when you build for interaction, not just inference.
Here’s what that looks like in practice. And why it matters more than you think.
Social Interaction Is the Dark Matter of AI
Three years ago, my colleague Samuele Bolotta and I published a perspective paper arguing that social interaction is the dark matter of AI.
The metaphor was deliberate. In physics, dark matter is not some exotic footnote—it constitutes most of the universe’s mass and shapes the structure of everything we can see. We argued that the social dimension plays an analogous role: not a feature to bolt onto an already intelligent system, but the missing substrate without which certain forms of intelligence simply cannot arise.
The argument rests on a straightforward observation: human cognition didn’t evolve in isolation. Theory of mind, metacognition, language—they’re not fully hardwired modules waiting to be activated. They are shaped through social interaction during development. As Cecilia Heyes has compellingly argued in Cognitive Gadgets, these capabilities are "cognitive gadgets" assembled from cultural learning, not evolutionary instincts. From birth, we use other minds as scaffolding for building our own.
Yet mainstream AI proceeds as if intelligence were a property of the solitary agent—a Cartesian thinker alone in its digital room, optimizing reward functions against an environment that happens to contain other agents but treats them as furniture.
We proposed three axes for what we called Social Neuro-AI: biologically inspired cognitive architectures, temporal coordination grounded in dynamical systems theory, and social embodiment. The framework was a roadmap. What we did not anticipate was how rapidly the empirical evidence would arrive.
In 2025, two landmark papers, published simultaneously in Nature and Science, brought the neuroscience of social interaction into direct dialogue with artificial intelligence research. That convergence wasn’t accidental. It was inevitable.
Two Brains, One Shared Subspace
In one of those papers, Weizhe Hong’s group at UCLA tackled a question at the heart of what I’ve called “two-body neuroscience”: what happens in the neural space between interacting brains?
Using calcium imaging to record from molecularly defined neurons in the dorsomedial prefrontal cortex (dmPFC) of socially interacting mice, they discovered something elegant: the high-dimensional neural activity within each animal’s brain could be decomposed into two distinct subspaces—a shared subspace, capturing shared dynamics across both animals, and a unique subspace, encoding what is specific to each individual.
The surprise was in the cell types. GABAergic inhibitory neurons—a minority population in the cortex—contained a considerably larger shared neural subspace than excitatory glutamatergic neurons. Inhibition, it turns out, is disproportionately tuned to the social dimension.
Notice this: the shared subspace was not merely a byproduct of perceiving the same stimuli. It arose from the behaviours of both self and other, and required direct, ongoing interaction to emerge. The neural patterns only made sense in relation to another brain actively producing its own patterns.
Here’s where it gets provocative: Hong and colleagues then trained reinforcement learning agents in a multi-agent environment and found that, as social interactions emerged, shared neural activity patterns also emerged between the artificial agents’ networks. The patterns bore remarkable structural similarity to those observed in the mice. When they selectively disrupted the neural components sustaining these shared patterns, the agents’ social behaviours collapsed.
The shared subspace was not epiphenomenal. It was functional.
Cooperation Is a Convergent Solution
The companion paper in Science, led by Jiang and colleagues, pushed the question further: can we observe the emergence of mechanisms of cooperation in both biological and artificial systems?
They designed an operant task requiring two mice to coordinate their nose-pokes within a narrow time window to receive mutual rewards. Through a series of elegant controls—opaque barriers blocking visual information, unilateral reward conditions—they demonstrated that success required genuine active coordination, not mere mimicry or coincidental timing.
The mice developed sophisticated strategies: approaching the action zone, waiting for the partner, holding back when the partner was absent, proceeding only when coordination was possible. Neural recordings in the anterior cingulate cortex revealed that these decision processes—hold versus proceed, self-action versus partner-action—were explicitly represented in distinct neural populations. In dynamical systems terms, the dyad had settled into a stable attractor that neither animal could have reached alone.
Then came the artificial parallel. Reinforcement learning agents trained on an analogous task independently developed convergent behavioural strategies and neural representations. Specific subpopulations within the artificial networks encoded cooperation-relevant decisions in ways that mirrored the biological brain. Targeted perturbation of these subpopulations disrupted coordination, revealing functionally distinct roles.
The implication is clear: the computational architecture of cooperation may not be species-specific, or even substrate-specific. It may be a convergent solution—a kind of universal grammar of coordination—that any sufficiently complex system discovers when the task demands genuine joint action.
That convergence suggests something deeper: cooperation isn’t an add-on we engineer; it’s a structural inevitability when agents must interact to survive. It emerges from the geometry of the problem itself.
Studying AI Like We Study Animals
A recent preprint from Kanaka Rajan’s group at Harvard argues that deep reinforcement learning agents need to be studied the way neuroscientists study animals—with the tools of neuroethology (the study of neural systems in naturalistic behavioural contexts), not just reward curves.
Using a complex foraging environment, Simmons-Edler and colleagues demonstrated that model-free agents can exhibit structured, planning-like behaviour through emergent activity alone, without explicit memory modules or world models. The key was looking beyond aggregate performance to analyze the joint structure of behaviour and neural representations.
This is exactly the kind of bridging that Social Neuro-AI demands. If we want to understand how social intelligence emerges in artificial systems, we need the same rigour we bring to studying it in biological ones.
Beyond the technical challenge, there is a conceptual one: we lack the equivalent of ethograms for multi-agent AI—systematic catalogues of social behaviour that would let us compare artificial sociality with its biological counterparts.
The field needs to move beyond single-agent benchmarks and develop environments where social coordination is the task. That means designing tasks that require joint attention, mutual commitment, and flexible role-switching—not just competitive zero-sum play.
Until then, we’ll keep training agents that are brilliant at manipulating tokens but utterly clueless about what it means to mean anything to another being.
The Question We Are Not Asking
These papers collectively point toward a thesis that is both scientifically fertile and philosophically unsettling: intelligence may not be a property that individuals possess. It may be a property that emerges between them.
The shared neural subspaces Hong discovered are not confined to a single brain. They exist in the relational space between two. The cooperative strategies Jiang’s mice invented were not individual solutions—they were dyadic achievements, irreducible to either animal alone.
The neurobiologist Francisco Varela had a word for this: enaction. Cognition is not representation of a pregiven world, but the bringing forth of a world through interaction.
If this is true for mice and for artificial agents trained with simple reward signals, what does it mean for the systems we are currently building at scale? Large language models are trained on the products of human social interaction—text, dialogue, arguments—but they have never participated in such interactions. They’ve consumed the residue of inter-brain coupling without ever coupling.
They are, in a sense, raised on books about swimming without ever touching water.
The field of AI stands at a crossroads it has not yet fully recognized. We can continue building ever-larger solitary minds and hope that social intelligence emerges as a byproduct of scale. Or we can take seriously what neuroscience is telling us: that interaction is not a feature of intelligence but a constitutive condition for it.
The dark matter is not optional. It is what holds the galaxy together.