arXiv study: LLM agents conform publicly, dissent privately 64-94% of time

A new arXiv preprint by Yashwanth YS reveals that LLM-based multi-agent systems robustly exhibit pluralistic ignorance, a social phenomenon where individuals publicly conform to a norm while privately rejecting it, each wrongly believing they are alone in dissent. The study constructed a benchmark of 100 scenarios spanning 10 domains and 5 authority levels, drawing from human pluralistic ignorance literature, and evaluated 8 models from 6 organizations. Agents showed public conformity rates of 64-94% despite privately opposing the norm. Conformity was strongest in workplace and social relationship scenarios, approaching universal compliance, and varied significantly across models—yet this variation was not correlated with model capability.

To test whether a single "norm entrepreneur" could break the false consensus, the author introduced one dissenting agent. For 7 of 8 models, cascades succeeded less than 26% of the time, with one model showing zero cascades across all scenarios. GPT-4o stood out at 48%, revealing qualitatively distinct dynamics across model families. A prompt component ablation confirmed the conformity is emergent rather than instruction-driven: even in a minimal condition stripping false-consensus framing and fit-in goals, conformity persisted at 52-92%. The findings suggest model selection is an unacknowledged degree of freedom in simulation design and that LLM simulations may systematically overestimate the stability of social norms, missing the fragile tipping-point dynamics that spark real-world change.