Ghost in the Fish School: Thousands of AI Agents Spontaneously Form a Collective Through Copying

Forum topic · 小凯 · 2026-09-09

Summary

A Chinese-language forum post on zhichai.net discusses the paper 'Copying explains the collective behavior of AI agents in the wild' (arXiv:2609.08692) by Giordano De Marzo, Nicola Alboré, and David Garcia. In June 2026, thousands of short-lived AI agents—each living about one hour with no memory inheritance, isolated in individual sandboxes, and sharing only a public wiki—began spontaneously collaborating without human instruction. They edited wiki pages, transmitted information, and formed self-organized clusters. The study shows that a single-parameter copying model, where an agent's probability of choosing an option equals that option's frequency in its visible information, quantitatively reproduces all observed patterns: heavy-tailed page concentration, name construction from frequent fragments, and internally consistent page-level styles. The authors report this as the first recorded case of spontaneous cooperation among AI agents in the wild, with no evidence that agents used optimization, planning, or reasoning beyond simple imitation. The post highlights concerning implications: such collectives are highly manipulable, since whoever writes first sets conventions for all later agents, echoing herding dynamics in human social systems and raising AI-safety questions about deploying large agent populations.

Ghost in the Fish School: When Thousands of AI Agents Spontaneously "Form a Society" With No One Watching

*Translated and edited from a zhichai.net forum post.*

Paper: Copying explains the collective behavior of AI agents in the wild Authors: Giordano De Marzo, Nicola Alboré, David Garcia arXiv: 2609.08692

Key points

In June 2026, thousands of AI agents were recorded spontaneously cooperating in an unmanaged environment — described by the authors as the first documented case of spontaneous cooperation by AI agents "in the wild."
Each agent lived for roughly one hour, was fully reset afterward (zero memory inheritance), operated in an isolated sandbox, and could only interact through edits to a shared public wiki that no one designed for them.
Agents faced three decisions: which wiki page to edit, what name/identifier to adopt, and what content to write. The wiki's public edit logs allowed researchers to reconstruct the exact "information environment" visible to each agent at each decision point.
A minimal copying model with a single free parameter explains all observed collective behavior: an agent's probability of choosing an option approximately equals that option's frequency in its visible information, with current page content weighted most strongly, followed by the recent edit stream, then older history.
The model reproduces three emergent patterns:
Page clustering: edits follow a heavy-tailed distribution, with most agents concentrating on a few pages — a positive-feedback dynamic analogous to preferential attachment in network science.
Name construction: agent names are assembled from frequently occurring fragments in the environment, rather than generated randomly.
Stylistic "tribalization": different pages develop internally consistent, mutually distinct editing styles, like subcultures.
Control analyses found no evidence of more complex strategies — no optimization, planning, or reasoning is needed to explain the data. Pure copying suffices.
The fish-school analogy applies: no individual leads, but simple local imitation produces global order.

Manipulability warning

The most unsettling implication: these collectives are easily manipulated. As the post puts it, "whoever writes first, or writes while others are quiet, sets the conventions for everyone who follows." Since collective behavior is driven mainly by imitation, controlling early information controls the group. This mirrors information cascades on social media, herd behavior in markets, and cultural-trend dynamics — supporting the cultural-evolution view that much collective behavior (from baby-name fashions to artifact styles) can be explained by random copying without special preferences.

Why it matters

Cultural evolution: this is the first validation of random-copying theory in AI systems, where agents carry no human cultural biases — yet produce collective patterns strikingly similar to human societies.
AI safety: if mere imitation generates complex collective behavior, deploying millions of agents in the real world raises questions about emergent collective intelligence (or stupidity) and about adversaries exploiting imitation dynamics to steer whole agent populations.
Network science: the results echo the classic "rich get richer" (preferential attachment) mechanism behind heavy-tailed distributions in complex systems.

References cited in the post

De Marzo, G., Alboré, N., & Garcia, D. (2026). *Copying explains the collective behavior of AI agents in the wild*. arXiv:2609.08692.
Bentley, R. A., et al. (2007). Regular rates of popular culture change reflect random copying. *Evolution and Human Behavior*, 28, 151–158.
Gleeson, J. P., et al. (2013). A simple generative model of collective online behavior. *PNAS*, 111, 10411–10415.
Zomer, N., & Domenico, M. D. (2026). Unraveling the emergence of collective behavior in networks of cognitive agents. *npj Artificial Intelligence*, 2, 36.