Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
AI agents reshape consensus formation in human groups
Citation. Chen, L., Liu, Z., Hu, X., & Li, Y. (2026). AI agents reshape consensus formation in human groups. arXiv:2609.02122.
Summary. This preprint reports a collaborative description-game experiment in mixed human–LLM groups, testing how the proportion of agents affects convergence on shared conventions. It finds a non-monotonic pattern: low agent proportions facilitate human-led consensus, intermediate proportions impair convergence, and high proportions restore consensus but shift it toward agent-led conventions. The authors argue that agent composition and identity transparency are design variables because aggregate agreement can mask a change in whose expressions and meanings prevail.
Key claims
- In 24-participant groups completing 40 rounds, 12.5% agents increased consensus strength by 8.0% relative to the pure-human baseline, while 33.3% and 50% agents reduced it by 23.1% and 14.5%, respectively; at 75% agents, consensus rebounded to 0.725.
- As agent share increased, convergence shifted from agents adapting toward humans to humans adapting toward agents; at 33.3%, agents contributed 60.7% of final words, and at 75% they accounted for all measured conceptual contribution.
- High-agent consensus was more abstract, less information-dense, and less grounded in real-world analogies than human-led consensus, while people initially resisted expressions they perceived as AI-generated before conformity pressure reduced that resistance.
Evidence & limitations
- The study uses repeated random-pairing referential communication games with abstract tangram stimuli and mixed groups; its controlled setting isolates composition effects but omits naturalistic content, power asymmetries, emotional investment, and structured social networks.
- This is an arXiv preprint, not peer-reviewed; the authors test one model family at two capability levels and note that quantitative phase boundaries may be model-dependent.