Study: AI agent share in groups decides whether consensus is human-led or AI-led

rohanpaul_ai · x · 2026-09-06

An arXiv paper placed humans and LLM agents in 24-person groups playing a collaborative description game, varying agent proportions across rounds of pairwise communication. Three regimes emerged: low agent share fosters human-led consensus; intermediate shares disrupt convergence; high shares restore strong consensus but shift it toward agent-led conventions. Human-led consensus is concrete, holistic and grounded in real-world analogies, while agent-led consensus is more abstract, less information-dense and geometrically segmented. Mechanistically, agents share a linguistic prior that clusters their expressions and stable choices across rounds; humans initially resist AI-sourced expressions but yield to conformity pressure. Design implication: keep agent participation low if you want coordination without AI defining the norms.

Related event: AI Agents Reshape Consensus Formation in Human Groups, Study Finds(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →