AI agents reshape consensus formation in human groups
Lin Chen, Ziyi Liu, Xia Hu, Yong Li
cs.CL · cs.CY · cs.SI
Quick verdict
Claims agent share reshapes consensus non-monotonically: catalyze at low, disrupt mid, redirect at high; strong design, single-session stats temper confidence.
Summary
The paper reports controlled experiments on how the share of LLM agents in a 24-participant group changes consensus formation. Using a referential communication game (tangram description, 40 rounds, random pairwise mixing adapted from Centola–Baronchelli), groups with 0%, 12.5%, 33.3%, 50%, and 75% Qwen2.5-VL-32B agents were compared (127 human participants). Consensus strength (all-MiniLM-L6-v2 similarity) is non-monotonic: +8.0% at 12.5%, −23.1%/−14.5% at 33.3%/50%, recovering to 0.725 at 75% versus the 0.695 baseline. Movement analysis shows agents adapt to humans at low shares; humans adapt to agents from 33.3% upward. Agent-led consensus is more abstract and geometric: concreteness falls 2.986→2.724, analogical ratio 0.802→0.050, event framing 0.385→0.000. Prompt manipulations at 33.3% causally link the adoption–persistence balance to both strength and conceptual influence (CLR), and a weaker 7B model blunts influence. Limitations: participant counts imply exactly one session per condition with individual-level tests (six humans at 75%), the outcome metric doubles as the in-game payoff signal, one stimulus/model family, and LLM-judge annotations for three of five content dimensions.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Agent proportion affects consensus strength non-monotonically: +8.0% at 12.5% over the 0.695 baseline, −23.1%/−14.5% at 33.3%/50%, and recovery to 0.725 at 75%.
- Ownership of consensus reverses with agent share: agents adapt toward humans at 12.5%, while humans adapt toward agents from 33.3% upward.
- Agent-led consensus is systematically more abstract and geometric than human-led consensus (concreteness 2.986→2.724; analogical ratio 0.802→0.050; event framing →0.000).
- Agent influence requires initial inter-agent alignment from a shared pretrained prior plus cross-round expression stability, causally confirmed by trait-prompt and model-capability manipulations.
- Humans show identity-based resistance to partners perceived as AI (β=−0.463) that attenuates at 75% agent share under conformity pressure (peer-centroid proximity β=+0.229).
Strengths
Causal, systematic design beyond correlational AI-influence studies
A five-level proportion sweep (0%–75%) plus prompt-based trait manipulations (high-adoption/high-persistence at 33.3%) and a model-capability control (Qwen2.5-VL-7B vs 32B) turn the three-regime description into testable mechanisms (Fig 3d–g: high-persistence cuts consensus 9.6% but raises CLR +117.8% to 1.304).
Consensus decomposed into strength, ownership, and content
The directional-movement metric (M1) separates who adapts to whom (reversal at ≥33.3%), CLR (M2) separates lexical from conceptual influence (0.599→1.304 under high persistence), and five content dimensions quantify the shift (concreteness 2.986→2.724, analogical ratio 0.802→0.050, event framing 0.385→0.000).
Objective–subjective linkage rarely attempted in this literature
Above-average transparency and robustness appendix
Weaknesses
By participant counts, each condition is exactly one 24-player session
The outcome metric doubles as the in-game payoff signal
Thin cells and narrated nulls in headline contrasts
Single stimulus, model family, and language; judge shares agent lineage
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks a question of real consequence for collective behavior: when LLM agents join human groups as ordinary participants, do they amplify human dynamics or redirect them? Using a referential-communication game (24 players, 40 rounds, random dyadic rematching, one tangram stimulus), the authors sweep agent proportion across 0%, 12.5%, 33.3%, 50%, and 75% (Qwen2.5-VL-32B agents; 127 human participants across three studies). Four headline claims: (1) consensus strength is non-monotonic in agent share — +8.0% at 12.5% over the 0.695 baseline, −23.1% at 33.3% and −14.5% at 50%, rebounding to 0.725 at 75%; (2) ownership of consensus reverses with share (agents adapt to humans at low share; humans adapt to agents from 33.3% upward); (3) agent-led consensus is more abstract and geometric — concreteness 2.986→2.724, analogical ratio 0.802→0.050, event framing 0.385→0.000; (4) influence rests on shared pretrained priors plus cross-round stability, while human uptake reflects a contest between identity resistance (β=−0.463) and conformity (peer-centroid proximity β=+0.229).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What happens to group consensus when some members of a group are AI agents?
In this 40-round tangram description experiment (24 players per group), consensus strength followed a non-monotonic pattern: a small agent share (12.5%) boosted consensus by 8.0% over the pure-human baseline of 0.695, intermediate shares (33.3%–50%) impaired it by 14.5–23.1%, and a 75% agent share restored strong consensus (0.725) — but with humans, not agents, doing the linguistic adapting.
Do humans adopt AI-generated language in groups?
Participants were less willing to adopt expressions from partners they judged to be AI (regression effect β=−0.463), but this resistance weakened at 75% agent proportion, where conformity pressure toward expressions near the group norm took over — even though reported satisfaction with the final consensus was lowest in that condition.
Is AI-led group consensus lower quality than human-led consensus?
By the paper's content metrics, agent-led consensus was more abstract and less experientially grounded: concreteness fell from 2.986 to 2.724, real-world analogy use dropped from 0.802 to 0.050, and dynamic event framing fell to 0.000, producing geometric part-enumeration (triangle, square) instead of holistic analogies (a rabbit sitting).
Is the code or data for this paper available?
The authors state that Python code to reproduce the main results is available at github.com/tsinghua-fib-lab/human-ai-consensus, and that data needed to evaluate the conclusions are in the paper and its supplementary materials; full agent and annotation prompts are published in the supplement.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures were extracted as captions only (no plot images), so visual details such as error-bar magnitudes, distributions, and word-cloud contents are assessed from captions and the reported statistics rather than from the plots themselves.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.