Preprint
Review

This version is not peer-reviewed.

Collective Social Behaviours of AI Agents: A Survey of Causes, Effects, and Interventions

Submitted:

11 November 2026

Posted:

13 November 2026

You are already at the latest version

Abstract
Multiple AI agents, most of them built on large language models (LLMs), are collectively deployed to complete certain tasks. Their interactions produce collective behaviours known from human societies, such as conformity, polarisation, echo chambers, and capture by zealots. These collective behaviours cannot be fully predicted or observed by studying single agents alone. In this tutorial and survey paper, we discuss thirty-three behaviours and classify them as observed, emerging, or anticipated. Nineteen major observed behaviours are reviewed in six families, namely agreement, division, minority influence and manipulation, contagion and diffusion, cooperation and coordination, and structure and hierarchy. We review the causes, effects, and interventions of each. Interventions suppress a behaviour when its effects are harmful and promote it when its effects are beneficial. We also give examples to explain key models. A synthesis section summarises the drivers and interventions shared across families and presents a perspective on the fourteen remaining behaviours, which are either recently observed, emerging, or anticipated. Emerging and anticipated behaviours can affect the stability of societies and financial markets when LLM agents become prevalent. Moreover, we present a case study with a scalar model showing that similarity-based memory retrieval acts as an endogenous confidence bound. This retrieval can produce echo chambers among simulated agents with no network homophily. Also, we evaluate diversity-aware retrieval as an intervention. Finally, we discuss open research directions. The paper is multidisciplinary, bridging the humanities and social sciences, the study of social networks, and AI.
Keywords: 
;  ;  ;  ;  ;  ;  ;  ;  ;  
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.