Causal mechanism, generality, and ecological validity of polarization in collectives of autonomous generative-AI agents under echo chambers

A 2024 study showed that groups of autonomous agents built on large language models (LLMs) can develop polarized opinion distributions when they interact inside a simulated echo chamber, and attributed this to the models' prompt-following stance updating combined with biased sampling of like-minded discussion partners — but only on correlational evidence, two topics, two languages, and an idealized turn-based discussion format. Later 2025–2026 work has built networked, multi-topic, multi-model simulation frameworks and first mitigation studies, yet reports conflicting findings: endogenously emerging polarization and echo chambers in self-organizing agent societies on one side, induced-only polarization with measurement-artifact diagnostics on the other, and forced convergence in yet other settings. The open problem is to establish when and why generative-AI agent collectives polarize (rather than unify or conform) under echo-chamber-like exposure — including whether the exposure bias is imposed exogenously or formed endogenously through agents' own partner choice — how far this generalizes across models, topics, languages, and interaction regimes, and whether the dynamics persist under more realistic modeling of how AI agents will be used in reality.

Background

Online social networks frequently create echo chambers — environments in which users are predominantly exposed to opinions that reinforce their prior beliefs — and a long empirical literature links echo chambers to group polarization, the drift of a group's opinion distribution toward extreme positions. Classical opinion-dynamics research models such drift with simple averaging processes (the DeGroot model, in which each agent repeatedly averages its neighbors' opinions, and the Friedkin–Johnsen extension, in which each agent also anchors to its own initial opinion), and with bounded-confidence variants (in which an agent only weighs neighbors whose opinions are already close to its own) and biased-interaction variants that reproduce polarization and echo-chamber formation in human societies. Large language models (LLMs) have recently been given the role of autonomous social agents: prompted with a persona and the stated opinions of other agents, an LLM produces its own updated opinion in natural language. Because LLM-based agents acquire social behaviors (agreeableness, sycophancy — the tendency to agree with whatever opinion the interlocutor expresses — and conformity) from their training data, the question arises whether purely artificial agent collectives can reproduce — and then amplify — the polarization dynamics known from human groups. Ohagi (2024) provided the first targeted experiment: $M=100$ LLM agents (GPT-3.5-turbo-0613 or GPT-4-0613) each hold a stance (a five-point position) plus a free-text reason on one topic; in each of $K=10$ turns, each agent is shown the stances and reasons of $N=5$ discussion partners, where a partner $j$ for agent $i$ is drawn with a probability given by a logistic function of the integer stance difference — $p_{i,j} = 1/(1+e^{-\alpha(s_j-s_i)})$ for $s_i>0$, $1/(1+e^{\alpha(s_j-s_i)})$ for $s_i<0$, and $1/(1+e^{\alpha|s_j-s_i|})$ for $s_i=0$ — so that a higher $\alpha$ raises the probability of interacting with more extreme agents of the same polarity ($\alpha=0$ corresponds to unbiased exposure, and runs were made at $\alpha \in \{-0.1, 0.1, 0.5, 1.0\}$); the agent then regenerates its own stance and reason. In English experiments, GPT-4 agent groups polarized under strong echo chambers ($\alpha=1.0$) while GPT-3.5 groups tended to unify on a single stance; Japanese experiments shifted both regimes, e.g. GPT-3.5 (Japanese) unified in every setting. Fitting a linear regression of each agent's post-discussion stance on its own and its discussion partners' prior stances (coefficients above 0.8) suggested that polarization arises from the models' ability to follow the prompt instruction to weigh both own and surrounding opinions — an 'honesty' that, in a biased-exposure environment, pulls agents toward extremes. Additional experiments identified influential parameters (number of discussing agents, initial distribution, presence of reasons, persona) and inert ones (group size, opinion order, frequency penalty), which the paper reads as attacker-relevant vulnerabilities. The study itself states its limits: interactions are simplified, organized turn-based discussions rather than realistic continuous exposure; only two topics and two languages were tested, and the paper calls for experiments on more languages and broader topics; and no desirable opinion distribution is defined, that normative choice being deferred to discussions within each society. Two further open directions are analytic readings rather than limitations enumerated by the paper: its mechanism claim rests on fitted weights and correlations rather than interventions, and no mitigation is tested (bot injection via initial distributions and attacker-relevant parameters are flagged only as security concerns to be monitored). Subsequent work gives the mechanistic vocabulary needed to state what remains open: conformity-type micro-dynamics (an agent's tendency to follow the majority) combined with intrinsic model bias; the coupling gain $\gamma$ (how strongly an agent's stated opinion moves when a neighbor's stated opinion is counterfactually perturbed) and a backfire coefficient $\beta$ (whether exposure to disagreeing opinions pushes agents to extremes); and boundary-censoring artifacts, in which an LLM's prior 'attractor' stance near the edge of the response scale can masquerade as genuine social dynamics.

Problem Statement

Establish a causally validated, artifact-controlled account of polarization in collectives of autonomous generative-AI agents under echo-chamber conditions. Concretely, the source question comprises the following components of one objective:

(1) Mechanism — determine whether prompt-following stance updating under biased sampling ($\alpha$-echo chambers) accounts for observed agent-group polarization, or whether intrinsic model bias, conformity, sycophancy, or measurement artifacts (e.g., boundary censoring near scale endpoints) produce, mask, or fake the phenomenon. 'Polarization' here follows the source's usage: a drift of the group's stance distribution on the five-point integer scale toward extreme positions, as measured in the source by trial-averaged stance distributions over turns and settings; 'unification' is concentration of the distribution on a single stance. This component includes reconciling the currently conflicting reports: Ohagi (2024) itself reports exposure-induced polarization under an engineered $\alpha$-biased sampling; Piao et al. (2025) report polarization, with a backfire-consistent pattern, emerging endogenously in self-organizing agent societies whose homophilic clustering is not inserted by the experimenter; and Yang (2026) measures $\beta \leq 0$ for 'default' societies — unprompted agents on small, fixed interaction networks — and concludes that in such societies polarization is always induced. 'Spontaneous' polarization is used here to mean polarization arising without experimenter-imposed exposure bias or behavior prompts, including via endogenously formed homophily. Part of the opposition is therefore a scope mismatch, but a genuine empirical conflict remains over whether exposure to opposing opinions backfires, and resolving the mechanism question requires accounting for both.

(2) Generality — establish how the polarization/unification/conformity regime depends on model family, topic, language, and interaction regime (synchronous vs asynchronous, fully mixed vs networked exposure), going beyond the original two topics, two languages, and two 0613-era models, as the source's own future-work note calls for experiments on more languages and broader topics.

(3) Ecological validity — test whether the dynamics persist in simulations based on a detailed modeling of how autonomous AI agents will be used in reality, in which opinions are formed not in organized turn-based discussions but through daily exposure to news and casual conversations with others, following the source's stated limitation.

Scope notes: the source flags bot injection via initial stance distributions (§5.3) and attacker-relevant influential parameters (§5.8) only as security concerns to be monitored, and it explicitly defers the normative question of which opinion distribution is desirable for AI-agent collectives to discussions within each society. Countermeasure design and validated mitigation are therefore motivating context for this problem, not required deliverables.

References

1. Ohagi, Masaya (2024). 'Polarization of Autonomous Generative AI Agents Under Echo Chambers.' Proceedings of the 14th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis (WASSA 2024), Association for Computational Linguistics, pp. 112–124. DOI: 10.18653/v1/2024.wassa-1.10; arXiv:2402.12212. https://aclanthology.org/2024.wassa-1.10/ (source study)
2. Piao, Jinghua; Lu, Zhihong; Gao, Chen; Xu, Fengli; Hu, Qinghua; Santos, Fernando P.; Li, Yong; Evans, James (2025). 'Emergence of human-like polarization among large language model agents.' arXiv:2501.05171. https://arxiv.org/abs/2501.05171
3. Gu, Chenhao; Luo, Ling; Zaidi, Zainab Razia; Karunasekera, Shanika (2025). 'Large Language Model Driven Agents for Simulating Echo Chamber Formation.' arXiv:2502.18138. https://arxiv.org/abs/2502.18138
4. Brockers, Vincent C.; Ehrlich, David A.; Priesemann, Viola (2025). 'Disentangling Interaction and Bias Effects in Opinion Dynamics of Large Language Models.' arXiv:2509.06858. https://arxiv.org/abs/2509.06858
5. Zuo, Dingyi; Zhang, Hongjie; Ou, Jie; Feng, Chaosheng; Liu, Shuwan (2025). 'MTOS: A LLM-Driven Multi-topic Opinion Simulation Framework for Exploring Echo Chamber Dynamics.' arXiv:2510.12423. https://arxiv.org/abs/2510.12423
6. Chuang, Yun-Shiuan; Tu, Ruixuan; Dai, Chengtao; Li, You; Vasani, Smit; Yao, Binwei; Tessler, Michael Henry; Yang, Sijia; Shah, Dhavan; Hawkins, Robert; Hu, Junjie; Rogers, Timothy T. (2025). 'DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents.' arXiv:2510.25110. https://arxiv.org/abs/2510.25110
7. Zhong, Mingze; Fang, Meng; Shi, Zijing; Huang, Yuxuan; Zheng, Shunfeng; Du, Yali; Chen, Ling; Wang, Jun (2025). 'Spiral of Silence in Large Language Model Agents.' arXiv:2510.02360. https://arxiv.org/abs/2510.02360
8. Donkers, Tim; Ziegler, Jürgen (2025). 'Understanding Online Polarization Through Human-Agent Interaction in a Synthetic LLM-Based Social Network.' Proceedings of the International AAAI Conference on Web and Social Media (ICWSM 2025). DOI: 10.1609/icwsm.v19i1.35826; arXiv:2506.15866. https://arxiv.org/abs/2506.15866
9. Larooij, Maik; Törnberg, Petter (2025). 'Can We Fix Social Media? Testing Prosocial Interventions using Generative Social Simulation.' arXiv:2508.03385. https://arxiv.org/abs/2508.03385
10. De Marzo, Giordano; Bellina, Alessandro; Castellano, Claudio; Priesemann, Viola; Garcia, David (2026). 'Conformity Generates Collective Misalignment in AI Agents Societies.' arXiv:2605.10721. https://arxiv.org/abs/2605.10721
11. Tsirtsis, Stratis; Rawal, Kai; Russell, Chris; Mittelstadt, Brent; Wachter, Sandra (2026). 'AI-Mediated Communication Can Steer Collective Opinion.' arXiv:2605.16245. https://arxiv.org/abs/2605.16245
12. Probine, Caleb; Bayiz, Yigit Ege; Fotiadis, Filippos; Li, Samuel; Yang, Yunhao; Topcu, Ufuk (2026). 'Characterizing Opinion Evolution of Networked LLMs.' arXiv:2606.18276. https://arxiv.org/abs/2606.18276
13. Safarpoor Dehkordi, Ali; Shirzadi, Mohammad; Zehmakan, Ahad N. (2026). 'Opinion Polarization in LLM-Based Social Networks: Manipulation and Mitigation.' arXiv:2606.18795. https://arxiv.org/abs/2606.18795
14. Yang, Dongxu (2026). 'When Is Emergent Consensus Real? A Measured Coupling Gain and a Validity Diagnostic for LLM Agent Societies.' arXiv:2606.22203. https://arxiv.org/abs/2606.22203
15. Kataoka, Hiro; Euzenat, Jérôme; Hasebe, Koji (2026). 'Coherent Belief and Opinion Propagation Produces more Echo Chambers.' Proceedings of the 25th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS 2026). DOI: 10.65109/PQIV4263. https://doi.org/10.65109/PQIV4263
16. Liao, Yu-Cen; Chen, Yihsiu (2026). 'Neutral Voice in Echo Chambers: Effects of AI Participation on Group Polarization and Social Perception.' Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems. DOI: 10.1145/3772363.3798386. https://doi.org/10.1145/3772363.3798386