Transparency note — This article draws on recent work (Evans et al., arXiv:2603.20639, March 2026) whose level of evidence remains exploratory. The conclusions presented are research hypotheses, not results confirmed by the scientific community.

In brief

A paper by researchers from Google and the University of Chicago (Evans, Bratton, Agüera y Arcas, March 2026) challenges the classic vision of the singularity: a monolithic AI spiraling toward isolated superintelligence. Their thesis: if an intelligence explosion occurs, it will be plural, social, and hybrid — carried by billions of interacting human and artificial agents. Agents do not replace LLMs alone; they reproduce, at a new scale, a mechanism that evolution has always used to amplify collective intelligence.

In short: forget the Hollywood image of a machine self-improving in a loop until it surpasses humanity. Instead, picture a city that becomes more intelligent not because its inhabitants improve one by one, but because it invents the street, the public square, the printing press, the university. Intelligence has never been individual — it has always been a property of collective organization. AI agents, in this thesis, are merely a new layer of that architecture.

The central thesis: intelligence has always been social

The starting point of Evans et al. is not technological — it is evolutionary. Each major transition in the history of intelligence (language, writing, institutions) did not consist of making individuals more powerful, but of creating a new collective unit of cognition. The primate brain evolved with the size of social groups, not with the difficulty of habitat [Dunbar, 1998]. Human language produced what Tomasello calls the “cultural ratchet”: a transgenerational accumulation of knowledge that no one needs to reconstruct alone [Tomasello, 1999]. Writing and institutions externalized social intelligence into durable infrastructure — a Sumerian scribe managing an agricultural accounting system did not understand his macroeconomic function; the system was functionally more intelligent than he was.

LLMs fit into this sequence. They are trained on the accumulated output of human social cognition — the cultural ratchet made computationally active. What migrates into the model’s parameters is not abstract reasoning — it is externalized social intelligence [Farrell, Gopnik, Shalizi, Evans, Science 2025].

The implication: if intelligence is fundamentally social, the path toward more powerful AI does not run through a monolithic oracle, but through richer social systems — and these systems will be hybrid.

The “society of thought” inside a model

The first level of the thesis concerns what happens inside a single reasoning model. Evans et al. cite a recent study (Kim, Lai, Scherrer et al., arXiv:2601.10825, 2026) on DeepSeek-R1 and QwQ-32B: these models do not improve their performance simply by “thinking longer.” They simulate multi-agent interactions inside their own chain of thought — what the authors call a “society of thought.”

The observed mechanism: the model spontaneously generates internal debates between distinct cognitive perspectives that argue, question, verify, and reconcile. This conversational structure would causally explain the precision advantage on difficult reasoning tasks — the authors demonstrated this by explicitly amplifying multi-party exchanges.

What makes the result notable: none of these models were trained to produce societies of thought. When reinforcement learning rewards only reasoning accuracy, multi-perspective conversational behaviors emerge spontaneously. The models rediscover, through optimization pressure alone, what centuries of epistemology have suggested: robust reasoning is a social process [Mercier & Sperber, 2011].

External agents: centaur configurations

The second level concerns external multi-agent systems. The authors describe the emergence of “centaur” configurations — composite actors that are neither purely human nor purely machine. A human directing many agents; an agent serving many humans; shifting mixed configurations. Each of us may transition between these configurations several times a day.

An agent facing a complex problem can now instantiate copies of itself, differentiate them, assign them subtasks, then recombine the results. An emerging perspective encountering a sub-problem beyond its scope can spawn its own subordinate society — a recursive descent into collective deliberation that unfolds when complexity demands it and closes when the problem resolves.

This mechanism recalls that of the major evolutionary transitions [Szathmáry & Smith, 1995]: new units of cognition emerge through social aggregation, not through individual improvement.

What this changes for alignment and governance

The social vision of intelligence has direct implications for alignment. The dominant model — RLHF (Reinforcement Learning from Human Feedback) [Christiano et al., NeurIPS 2017] — resembles a parent-child model of correction: fundamentally dyadic, it cannot scale to billions of agents.

In short: current RLHF is like a parent correcting their child, one feedback at a time. This dyadic logic works at small scale. For billions of agents in continuous interaction, you cannot multiply the “parents” — you have to switch to institutions: a tribunal, a protocol, a shared norm that does not depend on who occupies it. That is the move from “correcting the child” to “building the school”.

Evans et al. propose an alternative model: institutional alignment. Just as human societies do not rest on individual virtue but on persistent institutions — courts, markets, bureaucracies — large-scale agent ecosystems will need digital equivalents. An agent’s identity matters less than its capacity to fulfill a role protocol. A court functions because “judge,” “lawyer,” and “jury” are well-defined slots, regardless of who occupies them [North, 1990; Ostrom, 1990].

Social and organizational sciences have spent a century studying how team size [Wuchty, Jones, Uzzi, Science 2007], composition, hierarchy, role differentiation, and conflict norms shape collective performance. Almost none of this research has yet been applied to AI reasoning systems [Xu et al., arXiv:2501.09686, 2025]. This is precisely what the authors identify as an open design space.

For governance: when AI systems are deployed in high-stakes decisions (hiring, sentencing, benefit allocations), the question of who audits the auditors becomes unavoidable. The answer could be constitutional — governmental AIs with explicitly invested values (transparency, fairness, due process) that check and balance private-sector AIs, and vice versa. The alternative would be, for example, leaving the SEC to recruit business graduates armed with Excel spreadsheets to face high-dimensional collusion on AI-augmented trading platforms.

Governance, in the cybernetic sense, must be embedded in systems themselves as they grow more complex — decision verification protocols, procedural delegation of tasks, reliable scaffolds for delicate inter-agent collaborations. These technical protocols may have as much real-world effect as laws.

Limits of the thesis

The thesis of Evans et al. is stimulating, but several points merit cautious examination.

The evolutionary generalization is speculative. The analogy with the major evolutionary transitions (eukaryotic cells, multicellular organisms, human societies) is heuristically useful, but remains an analogy. Evolutionary mechanisms operate on timescales and selection pressures very different from artificial systems. The claim that AI agents constitute a “new major transition” is a hypothesis, not an empirical result. [UNVERIFIED by an independent source]

The cited platforms are embryonic. The authors mention OpenClaw and Moltbook as “embryonic glimpses” of the future. These platforms remain marginal, and the citation barely supports a thesis of civilizational transformation.

Society of thought: causation or correlation? The study on DeepSeek-R1 and QwQ-32B shows a correlation between conversational structure and precision. The claim that this structure “causally explains” the precision advantage is strong — the exact mechanisms remain to be established by further work.

Institutional alignment lacks specification. The idea of digital institutions analogous to courts or markets is conceptually appealing. But the authors do not specify how these institutions would be designed, governed, or updated — nor who would have authority to create them. The reference to Elinor Ostrom on commons management [Ostrom, 1990] is pertinent but underdeveloped.

The humans-in-the-loop hypothesis deserves testing. The authors assert that humans remain in the loop. But as agentic systems grow more complex and accelerate, effective human oversight may become nominal rather than real — precisely the problem the thesis seeks to resolve.

Key takeaways

  • Evans, Bratton and Agüera y Arcas (Google/UChicago, 2026) reject the monolithic singularity: if an intelligence explosion occurs, it will be plural, social, and hybrid, carried by billions of interacting agents.
  • Frontier reasoning models (DeepSeek-R1, QwQ-32B) spontaneously generate internal multi-perspective debates (“society of thought”) when trained on accuracy — emergent behavior, not programmed.
  • Agentic AI extends a documented evolutionary sequence: every major intelligence transition has consisted of creating new collective units of cognition, not making individuals more powerful.
  • Alignment at the scale of billions of agents requires digital institutions (roles, norms, protocols) rather than dyadic corrections — institutional alignment as an alternative to classical RLHF.
  • The thesis rests on stimulating but empirically unverified evolutionary analogies; the mechanisms of society of thought and the concrete forms of agent institutions remain to be specified.