Original Authors: James Evans, Benjamin Bratton, Blaise Agüera y Arcas
Published in: Science, Vol 391, Issue 6791, March 19, 2026

Abstract

The traditional "technological singularity" narrative assumes that a single superintelligence will achieve god-like intelligence through self-iteration. However, Evans et al.'s latest commentary in Science points out that this assumption is fundamentally flawed. Based on empirical studies of frontier reasoning models such as DeepSeek-R1 and QwQ-32B, the authors discovered that intelligence is essentially a high-dimensional, relational social attribute rather than a single quantifiable metric.

The research shows that reasoning models improve accuracy through multi-perspective debates within an internal "society of thought," and that human-AI "centaur" collaborative forms are reshaping knowledge work. The article proposes that the next "intelligence explosion" will manifest as the social complexification of billions of humans and trillions of AI agents, rather than the emergence of a single superintelligence. This poses requirements for AI governance to shift from "individual alignment" to "institutional alignment."

1. Background: Re-examining the "Technological Singularity" Narrative

For decades, the concept of "technological singularity" in artificial intelligence has depicted a specific vision: a single, massive intelligence gradually achieving god-like intelligence through self-guided iteration, consolidating all cognitive capabilities into a cold silicon node. This narrative has circulated widely in popular culture and some technology forecasts, but its core assumption—that intelligence is a single attribute measurable on a linear scale—is increasingly being questioned.

From the perspectives of evolutionary biology and cognitive science, Evans et al. point out that if AI development follows the path of previous major evolutionary transitions, the current stepwise improvement in computational intelligence will be pluralistic, social, and deeply entangled with its human predecessors. In fact, "human-level intelligence" itself is a vague concept—human intelligence has always been a collective attribute, not an individual one.

2. Core Findings: The Social Nature of Intelligence

2.1 The "Society of Thought" Within Models

The author team conducted empirical studies on frontier reasoning models such as DeepSeek-R1 and QwQ-32B, discovering a counterintuitive phenomenon: the accuracy improvements in these models do not simply stem from "thinking longer," but rather from complex multi-agent interactions simulated internally—a "society of thought."

Specifically, these models spontaneously generate internal debates between different cognitive perspectives within their chain of thought: argumentation, questioning, verification, and reconciliation. By explicitly guiding and amplifying this multi-agent dialogue, the research team confirmed the causal contribution of this conversational structure to the models' accuracy advantages on difficult reasoning tasks.

Implications of Emergent Behavior

This finding is significant because it reveals emergent behavior—these models were not explicitly trained to produce a society of thought, yet when reinforcement learning rewarded only reasoning accuracy, they spontaneously increased dialogic, multi-perspective behavior. From an epistemological perspective, the models rediscovered through optimization pressure insights that cognitive science and epistemology have developed over centuries: robust reasoning is a social process, even when it occurs within a single mind.

2.2 Human-AI "Centaur" Collaborative Forms

If intelligence is essentially social, the path to more powerful AI lies not in building a single giant oracle, but in constructing richer social systems—and these systems will be hybrid. The authors note that we have entered the era of "human-AI centaurs": composite actors that are neither purely human nor purely machine.

This composite form can take many shapes: one person directing multiple agents; one agent serving multiple people; multiple people and multiple agents collaborating in dynamic configurations. The authors use legal personality as an analogy—corporations and nations consist of numerous humans yet possess independent legal status and collective agency that no individual member can fully control—suggesting that similar structures may emerge at the scale of billions of humans interacting with non-human minds.

Currently, platforms such as OpenClaw (open-source multi-purpose AI agent building platform) and Moltbook (AI agent social network) are already showing the rudiments of this future. More structurally significant is the transformation that agents can now self-update, fork into two versions, and interact with each other; agents facing complex tasks can spawn new copies, distribute subtasks, and then recombine results. This recursive agent ecology is taking shape.

3. Discussion: Implications for AI Governance

These findings pose fundamental challenges to AI governance. The current mainstream AI alignment paradigm—Reinforcement Learning from Human Feedback (RLHF)—is essentially a parent-child correction model, a dyadic relationship that cannot scale to billions of agents.

The social intelligence perspective offers an alternative: institutional alignment. Just as human societies rely not on individual virtue but on enduring institutional templates—courts, markets, bureaucracies—defined by roles and norms, scalable AI ecosystems also need digital equivalents. An agent's identity matters less than its ability to fulfill role protocols, just as the function of a court stems from the well-defined positions of "judge," "lawyer," and "jury" rather than who specifically occupies them.

In high-stakes decision scenarios (hiring, sentencing, welfare distribution, regulatory enforcement), the question of "who audits the auditors" becomes unavoidable. The solution proposed by the authors has constitutional structural characteristics: governments need AI systems with different explicit value claims (transparency, fairness, due process) to check and balance AI systems deployed by the private sector and other government departments. For example, labor department AI auditing the differential impact of corporate hiring algorithms; judicial department AI assessing whether executive department AI risk assessments meet constitutional standards.

4. Limitations and Open Questions

As a commentary article, this paper presents a thought-provoking framework, but several points remain to be verified:

5. Conclusion

The vision proposed in this article is neither utopian nor dystopian, but evolutionary. Any emergent intelligence explosion will be nurtured by the interactions of eight billion humans with hundreds of billions, eventually trillions, of AI agents. Its scaffold is not the ascent of a single mind, but the complexification of composite societies: intelligence grows like a city, not like a single super-mind.

The "singleton singularity" framework leads policy to focus on preventing a technology that may never emerge. Instead, we should look for the next explosion in the same source as previous ones: cooperation, competition, and creative interaction among numerous socially intelligent minds. The difference is that this time, most minds will be non-biological.

This pluralistic model directs attention to the right places: the design of hybrid human-machine social systems, the norms that govern them, and the institutions and protocols on which their conflicts and coordination rely. In a sense, the intelligence explosion has already arrived—in the society of thought debating within every reasoning model, in the centaur workflows reshaping every knowledge profession, in the recursive agent ecology beginning to fork and collaborate at scale, and in the constitutional questions we must now begin to pose.

The question is not whether intelligence will become extremely powerful, but whether we can build the corresponding social infrastructure.

References

← Back to Blog