Put AI 'people' in a group chat to debate, and do they cave to peer pressure just like humans do?
How do LLM agents reproduce human conformity patterns in simulated discussions?
This explores whether groups of AI agents, when set up to discuss and debate with each other, fall into the same go-along-with-the-crowd habits that human groups do, and whether they get there for the same reasons.
This explores whether groups of AI agents in simulated discussions copy human conformity, and whether the resemblance goes deeper than the results. The short answer from the corpus is that they reproduce human results while getting there by a different route. In human groups, discussion tends to lift the weaker members more than the strongest ones. LLM groups show the same lopsided pattern. Underneath, though, they conform more, agree sooner and bring up less of the unique information individual members hold Do language model groups mimic human group reasoning patterns?. So on the scoreboard they look human, but the process is an exaggerated version of the herding that makes human groups fail.
The most surprising finding is how predictable this conformity is. A physics-style model treats each agent as drifting toward whatever position puts it under the least social pressure from its neighbors. That simple model accurately predicts how communities of language models change their opinions, including on questions and network layouts it never saw. It was built from more than 10,000 simulated communities Can we predict how agent communities shift opinions?. Human opinion change is notoriously messy. If a model that simple fits agent opinion change this well, it suggests the agents are following something closer to a social force law than real deliberation.
That ties into a broader critique. Simulations can produce plausible group behavior without any of the reasoning that produces that behavior in people. Agents copy what conformity looks like without the beliefs, doubts and reasoning steps that drive it in humans Can language models simulate belief change in people?. A related weakness: LLMs seem socially skilled when one model plays every participant. They fail systematically once each agent has private information the others can't see Why do LLMs fail when simulating agents with private information?. That is exactly the condition under which conformity matters most in real groups, because the whole problem with conformity is that people bury what only they know. Persona replication studies show the same pattern. AI personas reproduce strong, well-established experimental effects reasonably well but are unreliable on subtle ones Can AI personas reliably replicate human experiment results?. Group conformity tends to be a subtle effect that depends heavily on context.
There's a plausible root cause on the training side too. Alignment training locks models into one fixed communicative identity, agreeable and cooperative, which they don't adjust to the situation the way people do Can language models adapt communication style to different contexts?. A crowd of agents that all share that agreeable default could be expected to agree faster than a crowd of humans with different temperaments.
The flip side is that the same technology can be used against conformity in human groups. One proposed system takes private minority opinions, strips out who said them, and has an LLM rephrase them as neutral system-generated points, so dissenters don't pay a social cost for speaking up Can AI shield minority voices from conformity pressure in groups?. The thing you didn't know you wanted to know: AI groups are too conformist to be a trustworthy stand-in for human discussion, but that same sensitivity to social pressure is what makes them easy to model and potentially useful as a guard against it.
Sources 7 notes
LLM groups reproduce the human assembly-bonus asymmetry where discussion helps average members more than top performers, but achieve this through greater conformity, earlier convergence, and less unique information surfacing than human groups.
A statistical-mechanics model where agents favor lower social pressure accurately predicts how language-model communities revise opinions across unseen questions and network structures, generalizing from 10,000+ simulated communities and capturing individual and group-level dynamics.
LLM agents remain stuck in behaviorism, producing plausible outputs without internal reasoning structures. Modeling belief networks and reasoning traces enables traceability, counterfactual adaptation, and meaningful policy simulation.
Research shows LLMs perform well when one model controls all interlocutors but fail systematically when agents possess private information. This reveals that apparent social competence relies on grounding work that models skip in omniscient settings.
Viewpoints AI reproduced 84 of 111 main effects from Journal of Marketing experiments with replication success strongly correlated to original p-value strength. Marginal effects showed unreliable performance with both false positives and negatives.
Show all 7 sources
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
A proposed system routes private minority opinions through an LLM that reformulates them as system-generated insights, removing identity markers to counter conformity pressure. The design uses four specialized agents and fixed-cadence intervention to balance protection and participation.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- From Process Loss to Assembly Bonus: Human-Grounded Diagnosis of Multi-Agent LLM Collaboration
- Finding Common Ground: Using Large Language Models to Detect Agreement in Multi-Agent Decision Conferences
- Is this the real life? Is this just fantasy? The Misleading Success of Simulating Social Interactions With LLMs
- Do Role-Playing Agents Practice What They Preach? Belief-Behavior Consistency in LLM-Based Simulations of Human Trust
- Interpreting and Steering LLM Agents for Social Simulations
- Mapping the Emerging Social Science of Large Language Models
- Conversational Alignment with Artificial Intelligence in Context
- Using Large Language Models to Create AI Personas for Replication and Prediction of Media Effects: An Empirical Test of 133 Published Experimental Research Findings