How populations of LLM agents form, exchange and change opinions when they debate in a network, and how they influence humans.
Opinion dynamics models describe how individual views evolve through social interaction, giving rise to consensus, polarization or fragmentation. Classic models encode these updates as simple numerical rules, as in bounded confidence.
LLMs change the picture: agents can now argue in natural language, and their opinions emerge from what they read and write. We use LLM agents as the units of opinion dynamics simulations to understand what collective behaviour they produce, and what that reveals about the models and about their role in human debates.
LLM agents are not neutral discussants. Their opinion changes follow selective agreement rather than plain sycophancy, mixing models with different biases shapes the outcome, and populations tend to drift towards specific positions.
Who talks to whom matters. We embed agents in networks with different topologies, homophily and community structure, and compare the results with classic bounded-confidence dynamics.
Since opinions travel as text, we can analyse which arguments, rhetorical strategies and fallacies agents use, and link them to how LLMs persuade human users.
Define reproducible protocols for opinion dynamics simulations among LLM agents.
Measure biases, sycophancy and agreement drift across models, prompts and network structures.
Compare language-driven dynamics with classic opinion dynamics models.
Assess the persuasive power of LLMs on humans and the factors that make people susceptible to it.
@article{cau2026network,
title = {Network Effects and Agreement Drift in LLM Debates},
author = {Cau, Erica and Failla, Andrea and Rossetti, Giulio},
journal = {arXiv (Cornell University)},
year = {2026},
doi = {10.48550/arxiv.2604.11312},
url = {https://doi.org/10.48550/arxiv.2604.11312},
publisher = {Cornell University},
}@article{carrillo2026llms,
title = {LLMs can persuade only psychologically susceptible humans on societal issues, via trust in AI and emotional appeals, amid logical fallacies},
author = {Carrillo, Alexis and Citraro, Salvatore and Ardebili, Ali Aghazhadeh and Taietta, Enrique and Rossetti, Giulio and Ferrara, Emilio and Veltri, Giuseppe and Stella, Massimo},
journal = {arXiv (Cornell University)},
year = {2026},
doi = {10.48550/arxiv.2604.16935},
url = {https://doi.org/10.48550/arxiv.2604.16935},
publisher = {Cornell University},
}@article{cau2025selective,
title = {Selective agreement, not sycophancy: investigating opinion dynamics in LLM interactions},
author = {Cau, Erica and Pansanella, Valentina and Pedreschi, Dino and Rossetti, Giulio},
journal = {EPJ Data Science},
year = {2025},
volume = {14},
number = {1},
doi = {10.1140/epjds/s13688-025-00579-1},
url = {https://doi.org/10.1140/epjds/s13688-025-00579-1},
publisher = {Springer Nature},
}@article{cau2025language,
title = {Language-Driven Opinion Dynamics in Agent-Based Simulations with LLMs},
author = {Cau, Erica and Pansanella, Valentina and Pedreschi, Dino and Rossetti, Giulio},
journal = {arXiv (Cornell University)},
year = {2025},
doi = {10.48550/arxiv.2502.19098},
url = {https://doi.org/10.48550/arxiv.2502.19098},
publisher = {Cornell University},
}@incollection{cau2025bots,
title = {Bots of a Feather: Mixing Biases in LLMs’ Opinion Dynamics},
author = {Cau, Erica and Failla, Andrea and Rossetti, Giulio},
booktitle = {Studies in computational intelligence},
year = {2025},
pages = {166--176},
doi = {10.1007/978-3-031-82439-5_14},
url = {https://doi.org/10.1007/978-3-031-82439-5_14},
publisher = {Springer Nature},
}