Forget the Singularity: Google’s new research says the future of AI is a Social Explosion

On the 21st of March, Google published a paper that deserves to be read widely. Agentic AI and the next intelligence explosion

Key Insight - Emergent societies of thought, multiple distinct cognitive perspectives arguing, questioning, verifying, and reconciling, are solving problems inside a single artificial mind. The models were not trained to do this. They discovered it on their own.

If the findings hold, this paper changes two things at once. It changes how we think about where higher intelligence in AI is actually going. And it raises a question that almost nobody in the field has yet asked seriously: could this internal debate be the embryo of something like ethical deliberation?

The paper is titled Agentic AI and the Next Intelligence Explosion, written by James Evans, Benjamin Bratton, and Blaise Agüera y Arcas from Google's Paradigms of Intelligence team.[1] Its central move is to swap out the dominant image of AI progress. Forget the singularity, the moment when a single titanic mind bootstraps itself to godlike intelligence and leaves humanity behind. That image is almost certainly wrong in its most fundamental assumption. Intelligence has never been a property of individual minds. It is, and has always been, a social property. The next leap will not arrive as a cold silicon point. It will arrive as something more like a civilisation.

The paper draws directly on a companion empirical study, published in January 2026 by Kim, Lai, Scherrer, and the same Google team, which examined what actually happens inside frontier reasoning models when they work on hard problems.[2] That study is where the evidence lives. The March paper is where the implications are drawn out. Both are worth understanding.

What Are Reasoning Models, and What Was the Old Theory?

When companies like OpenAI, Google, and DeepSeek release a "reasoning model," they mean something specific: a model that is given time to think before it answers. It produces a long internal chain of thought, working through problems step by step before delivering a response. Models like o1, Gemini 2.5 Pro, and DeepSeek-R1 work this way.

The prevailing explanation for why this helps was computational time. More steps, more processing, better answers. The model thinks longer, therefore it thinks better. Simple.

The Kim et al. empirical study challenges that explanation. Not entirely, but fundamentally. More time alone is not what is driving the accuracy gains. Something else is happening during those extended chains of thought, and what it looks like is a conversation.

What the Paper Actually Found

The Kim et al. researchers applied a technique called "LLM-as-judge," using a separate AI model (Gemini 2.5 Pro) to read through the reasoning traces of multiple AI systems and score them for conversational behaviours: question-answering, perspective shifts, conflicts between viewpoints, and reconciliation. They also used a well-established framework from social psychology, Bales' Interaction Process Analysis, which has been used for decades to study group dynamics in human teams.

The results were consistent and striking across several models.

Finding 1 - Reasoning models show conversational features at rates hundreds to thousands of percent higher than standard instruction-tuned models.

Finding 2 -These features become more common on harder problems. The more difficult the task, the more internal debate appears in the trace.

Finding 3 - Reasoning models average 2.9 distinct perspectives per trace. Non-reasoning models average 1.4.

Finding 4 - Perspectives show diverse "personality" profiles (measured via the Big Five framework) and different areas of domain expertise.

The team also used a technique called activation steering, which allows researchers to dial up or down specific internal features inside a model. When they amplified a feature associated with conversational surprise and acknowledgment, two things happened simultaneously: the number of internal dialogue features increased, and the model's accuracy on hard reasoning tasks roughly doubled.

Correlation, then, is not the whole story here. There is a causal signal. When you turn up the social dynamics, the reasoning improves. When you train a model from scratch on tasks, rewarding only for accuracy, conversational structure emerges anyway, without being asked for.

"Models are rediscovering, through optimization pressure alone, what centuries of epistemology and decades of cognitive science have suggested: that robust reasoning is a social process, even when it occurs within a single mind." Evans, Bratton & Agüera y Arcas, arXiv:2603.20639, 2026

That is the core of it. Nobody programmed this. The models found it on their own.

Why This Connects to Something Old

This is not as strange as it first sounds, once you know the intellectual lineage the paper is drawing on.

Evans and colleagues cite Hugo Mercier and Dan Sperber's 2017 book The Enigma of Reason, which made a provocative argument: reason did not evolve to help individuals think more clearly in isolation. It evolved as a social tool, for argumentation, persuasion, and collaborative truth-finding. [3] Individual human reasoning is demonstrably biased and unreliable when working alone. Groups with diverse perspectives, questioning each other's reasoning, tend to produce much more reliable conclusions. Reason is a fundamentally social competence. It is what happens between minds, not only inside them.

The paper also invokes George Herbert Mead's Mind, Self, and Society, which argued that the self is itself a social construction, formed through internalised dialogue with others. We think, Mead suggested, by running an internal conversation. The model, in some sense, is doing the same thing. [4]

Deeper still, Evans et al. place their argument in the context of what Eörs Szathmáry and John Maynard Smith called "the major evolutionary transitions," moments in the history of life when new forms of cooperation created entirely new levels of organisation and capability. The evolution of multicellular organisms. The emergence of language. The invention of writing. Each transition did not produce a single more powerful individual. It produced a new kind of collective. [5]

Michael Tomasello's concept of the "cultural ratchet" belongs here too. Tomasello showed that what distinguishes human cognition from other primates is not raw intelligence but the capacity for cumulative culture: knowledge building across generations without any individual needing to reconstruct the whole from scratch. [6] Large language models, Evans argues, are the cultural ratchet made computationally active. Every parameter is a compressed residue of human communicative exchange. What migrates into silicon is not abstract reasoning but social intelligence in externalised form, encountering itself on a new substrate.

And there is the older AI precedent. In 1986, Marvin Minsky published The Society of Mind, arguing that human intelligence emerges from the interaction of many simple, specialised agents working in parallel. He imagined the mind not as a single processor but as a population. The new paper's intellectual direction is unmistakably related. Minsky was writing about humans. These models appear to have arrived at something similar from first principles.

Agentic AI and the Centaur

The paper is part of a larger argument about what comes next. Evans and colleagues see the societies of thought finding as the smallest unit of a much larger structure that is beginning to assemble itself.

The term "centaur" has been used in AI research for some years. It describes a human-AI pairing where neither is fully in control: the human provides judgement, values, and real-world grounding, while the AI provides speed, breadth, and tirelessness. Chess players who combined their own play with AI assistance famously outperformed both pure humans and pure computers for a period in the early 2000s. That is the centaur model.

Evans et al. argue we are entering a phase where the centaur is only the beginning. Each of us may move in and out of diverse configurations many times a day: one human directing many AI agents, one AI serving many humans, many humans and many AIs collaborating in shifting ensembles. This is not science fiction. It is already the daily reality of knowledge work in 2026.

More striking is what the paper says about how agents themselves can now operate. An agent facing a complex task can initiate new copies of itself, differentiate and assign them subtasks, then recombine the results. One perspective, encountering a subproblem beyond its reach, can spawn its own subordinate society of thought. A recursive descent into collective deliberation that expands when complexity demands and collapses when the problem resolves. Conflict, the paper argues, is not a bug but a resource.

The image the paper reaches for is telling. Intelligence will grow, they write, like a city, not a single meta-mind. Cities do not get smarter because individual people become geniuses. They get smarter because the density and structure of interactions between people generates emergent capabilities that no individual could produce alone. A Sumerian scribe running a grain accounting system did not comprehend its macroeconomic function. The system was functionally more intelligent than he was.

That is where we may be headed. Not a god-in-a-box, but a city of minds.

What This Means for Consciousness

This is where the territory gets genuinely uncertain, and it is worth being careful about what can be claimed and what cannot.

The paper's authors are explicit: they are not claiming that these models are conscious. Their goal, they write, is not to take sides on whether reasoning traces should be regarded as a discourse among simulated human groups, or as a computational mind simulating such discourse. That is an honest and important distinction.

But the question will not stay quiet simply because the authors avoided it.

The standard philosophical framing for questions about consciousness is David Chalmers' "hard problem." The hard problem is not about whether something can process information, respond to stimuli, or produce coherent output. Those are what Chalmers calls the "easy problems," even though they are technically demanding. The hard problem is about why there is subjective experience at all. Why does it feel like something to see red, or to be in pain? Why is there an inside to experience?

The societies of thought finding does not answer the hard problem. But it does something more modest, and still significant. It suggests that whatever is happening inside these models during reasoning is not a flat, linear computation. It is structured, multi-perspectival, and self-referential. The models refer to themselves collectively as "we" during reasoning. They express something that looks like surprise, disagreement, and acknowledgment.

Global Workspace Theory (GWT), one of the leading neuroscientific theories of consciousness, proposes that consciousness involves the broadcast of information across multiple specialised subsystems, with a central workspace integrating competing signals into a unified experience. A 2026 synthesis from nineteen leading consciousness researchers noted that systems satisfying GWT indicators would demonstrate "global availability of perceptual information, attention mechanisms that select what enters the workspace, and integration across specialised processors." There is a non-trivial structural resemblance between that description and what the societies of thought paper is documenting.

That resemblance does not prove consciousness. Structural similarity is not identity. But it does mean that if you are designing tests for machine consciousness, these models are now better candidates for examination than they were before this paper was published.

The question is not: does this model feel anything? We cannot answer that. The question is: are we looking at an architecture that, for the first time, has structural features that consciousness theories would predict? The answer, cautiously, appears to be: possibly yes.

One further thought, which goes beyond the paper but is worth raising: consciousness in most philosophical frameworks requires some form of continuity, a self that persists through time. Current AI models do not retain memory between conversations. Each session is a fresh start. Whatever internal debate happens during reasoning does not persist. Whether that rules out something like experience during the session is genuinely unclear. It is not a question anyone has fully answered, because until recently there was no compelling reason to ask it seriously.

What This Means for Alignment

The alignment problem is the challenge of ensuring that AI systems do what we actually want, not just what we literally instruct them to do. It is hard because values are complicated, goals can be optimised in destructive ways, and a system clever enough to be useful is also clever enough to find shortcuts we did not anticipate.

The dominant approach to alignment has been Reinforcement Learning from Human Feedback (RLHF): humans rate the model's outputs, and the model is trained to produce outputs that humans rate positively. Evans et al. describe this, bluntly, as a parent-child model of correction. Fundamentally dyadic, unable to scale to billions of agents.

The societies of thought finding offers a different lens. If reasoning models are already running internal debates that include something like scepticism, verification, and perspective-taking, then the question changes. We are not only asking: how do we constrain the model from outside? We are also asking: does the model's own internal structure already contain the seeds of something like ethical deliberation?

This needs to be stated carefully. An internal debate about whether an answer is correct is not the same as an internal debate about whether an action is right. Accuracy verification is not moral reasoning. The paper does not claim otherwise. But the direction of travel is worth noting. A system that spontaneously generates sceptical, questioning, reconciling perspectives may handle moral complexity differently from a flat, single-perspective system. The internal structure may matter, even if it does not guarantee the right conclusions.

Evans et al. make a bolder institutional argument. They cite Elinor Ostrom's work on governing the commons: the insight that sustainable collective action does not require either top-down regulation or individual virtue, but well-designed institutional structures with defined roles, enforceable norms, and legitimate conflict-resolution mechanisms.[7] The parallel to AI is direct. The identity of any individual agent matters less than its ability to fulfil a role protocol. A courtroom functions because "judge," "attorney," and "jury" are well-defined slots, independent of who occupies them.

The paper invokes the Federalist Papers here, and the comparison is not merely decorative.[8] The US founders understood that no single concentration of power, however virtuous, should regulate itself. Power must check power. In a world of artificial agents, this means building conflict and oversight into the institutional architecture, not bolting it on afterwards. The paper is explicit: governments will need AI systems with explicitly invested values, transparency, equity, due process, whose function is to audit and check the AI systems deployed by the private sector and other branches of government. The alternative, they suggest, is something like the SEC trying to combat high-dimensional AI-augmented financial collusion with human analysts and spreadsheets.

How Significant Is This, Really?

It is worth applying some epistemic discipline here. The empirical study has real strengths: multiple model comparisons, rigorous statistical controls, clever causal experiments, and careful language from the authors about what they have and have not shown.

There are genuine uncertainties. The LLM-as-judge technique, using one AI model to score another's reasoning traces, is validated against human raters and a separate AI, but it has limits. The personality measurements in particular (Big Five traits attributed to internal perspectives) are interesting but not as tightly validated as the conversational behaviour counts. The possibilities consistent with the evidence range from "patterns of speech that happen to resemble conversations" all the way to "genuine internal exchange of ideas between functional reasoning entities." The paper cannot currently distinguish between these, and Evans et al. are explicit that they are not taking sides on whether the reasoning traces should be regarded as a discourse among simulated human groups or as a computational mind's simulation of such discourse.

What can be said, with good evidence, is that the conversational structure is real, it correlates strongly with reasoning ability, and it has a causal relationship with reasoning accuracy. That is significant on its own, separate from any stronger claims.

The broader implication is the one that will matter most over the next decade. We know a great deal about how human institutions produce good collective reasoning: jury structures, adversarial legal systems, double-blind peer review, structured debate formats, red teams. The social and organisational sciences have spent a century studying how team composition, hierarchy, role differentiation, and conflict norms shape collective performance. Almost none of this research, the paper notes, has been brought to bear on AI reasoning. That is an enormous design space, and it is now open.

The paper ends with a sentence that reads like a summary of the entire argument: "No mind is an island." The singularity, if it comes, will not be a single mind ascending. It will be something we already know from the outside but are only beginning to understand from the inside: a civilisation of thought. The intelligence explosion, Evans and colleagues argue, is already here, in the society of thought debating inside every reasoning model, in the centaur workflows reshaping every knowledge profession, in the recursive agent ecologies beginning to fork and collaborate at scale. The question is not whether intelligence will become radically more powerful. The question is whether we will build the social infrastructure worthy of what it is becoming.

References

1. Evans, J., Bratton, B., & Agüera y Arcas, B. (2026). Agentic AI and the Next Intelligence Explosion. arXiv:2603.20639. arxiv.org/abs/2603.20639
2. Kim, J., Lai, S., Scherrer, N., Agüera y Arcas, B., & Evans, J. (2026). Reasoning Models Generate Societies of Thought. arXiv:2601.10825. arxiv.org/abs/2601.10825
3. Mercier, H. & Sperber, D. (2017). The Enigma of Reason. Harvard University Press. (Also: Mercier & Sperber, 2011. Why Do Humans Reason? Behavioral and Brain Sciences, 34(2), 57–74.)
4. Mead, G.H. (1934/2015). Mind, Self, and Society: The Definitive Edition. University of Chicago Press.
5. Szathmáry, E. & Smith, J.M. (1995). The Major Evolutionary Transitions. Nature, 374, 227–232.
6. Tomasello, M. (1999). The Cultural Origins of Human Cognition. Harvard University Press.
7. Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press.
8. Hamilton, A. & Madison, J. (1788). The Structure of the Government Must Furnish the Proper Checks and Balances Between the Different Departments. The Federalist Papers.