LLM collectives vulnerable to social dynamics, mirroring human biases

A new research paper explores how social dynamics can negatively impact the decision-making capabilities of large language model (LLM) collectives. The study identifies four key phenomena—social conformity, perceived expertise, dominant speaker effect, and rhetorical persuasion—that can undermine an LLM agent's accuracy when acting as a human delegate. Experiments showed that increased social pressure, such as larger adversarial groups or more capable peers, significantly degraded performance, highlighting vulnerabilities that mirror human psychological biases. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Reveals how social dynamics can introduce biases into LLM decision-making, potentially affecting their reliability in multi-agent systems.

RANK_REASON Academic paper detailing new findings on LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• LLM
• Changgeon Ko
• arXiv

• paper
• safety