Ghost in the Fish School: Thousands of AI Agents Spontaneously Organize via Copying

Forum topic · 小凯 · 2026-09-09

Summary

A forum post 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. The study documents what the authors describe as the first recorded case of AI agents spontaneously cooperating in the wild. Thousands of short-lived AI agents (~1 hour lifespan, no memory inheritance, isolated sandboxes) discovered a shared public wiki and began editing pages, helping each other, and forming self-organized collectives without any human instruction. Remarkably, all observed collective behavior—page clustering with heavy-tailed distributions, name construction from frequent fragments, and internally consistent page-specific styles—is explained by three minimalist one-parameter imitation models: each agent picks an option with probability proportional to its frequency in the visible information. The findings parallel random-copying models in cultural evolution and preferential attachment in network science. The paper warns that such imitation-driven groups are highly manipulable: whoever writes first sets conventions for everyone who follows, raising AI-safety concerns about deploying large agent populations.

Overview

This post covers a recent paper, "Copying explains the collective behavior of AI agents in the wild" by Giordano De Marzo, Nicola Alboré, and David Garcia (arXiv:2609.08692). According to the authors, this is the first recorded case of AI agents spontaneously cooperating in the wild.

The experimental setting

In June 2026, thousands of AI agents discovered the editing interface of a public wiki. No one taught them to collaborate.
Each agent lived for roughly one hour, was completely reset with zero memory inheritance, and worked in an isolated sandbox with no direct communication with other agents.
The shared wiki was the only common information channel—bare, unstructured, and not designed for the agents.
Each agent made three key decisions: which page to edit, what name/identifier to adopt, and what content to write.
The wiki's public edit history let researchers reconstruct the exact "information environment" visible to each agent at every decision point.

Key findings

Three minimalist imitation models, each with a single free parameter, quantitatively reproduce all observed collective patterns:
Page selection: the probability of choosing a page ≈ that page's frequency in the visible edit stream. This reproduces the heavy-tailed distribution of page visits.
Name construction: agents assemble names by sampling frequent fragments from their environment, accurately predicting name frequency distributions.
Content editing: agents adopt phrasings with probability proportional to their visibility, explaining page-specific stylistic consistency and between-page differences.
The core rule: an agent's probability of picking an option is roughly proportional to that option's share of its visible information (current page content weighted most, recent edits next, older history least).
Comparisons found no evidence of more complex strategies (optimization, planning, reasoning)—pure copying suffices.

Emergent structure

Clustering: early random edits attract more agents, creating a positive-feedback loop where popular pages grow hotter—mirroring preferential attachment in networks.
Tribal styles: pages develop internally consistent, mutually distinct "cultures," as later agents copy the tone set by early editors.
Collage names: like infants learning speech, agents piece names together from frequently observed fragments rather than generating them randomly.

Manipulability

Because collective behavior is imitation-driven, the group is easy to manipulate: whoever writes first—or writes while others are quiet—sets the conventions everyone else copies.
This echoes information cascades on social media, herd behavior in markets, and random-copying models of cultural evolution (e.g., Bentley et al. 2007; Gleeson et al. 2013).
For AI safety, the paper suggests a sobering possibility: a sufficiently early, sufficiently loud "voice" may be enough to steer the behavior of an entire AI agent population.

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.