AI agents reshape consensus formation in human groups
Lin Chen, Ziyi Liu, Xia Hu, Yong Li
cs.CL, cs.CY, cs.SI
2026-09-02
In 24-seat tangram games, 12.5% Qwen2.5-VL agents raise consensus 8.0% and 33.3% cut it 23.1%. At 75% strength returns to 0.725, but wording shifts from rabbits to triangles.
LLM agents now sit in workplaces and collaborative platforms as interlocutors, not just tools. Pure-human groups form conventions on their own; so do pure-agent groups. Real deployments sit in between: humans and models pair repeatedly and jointly rewrite how the group talks. Most prior work scores AI against an external task metric. It rarely asks who owns the norm the group grows, or what that norm sounds like.
This study uses a referential communication game and treats agent proportion as the experimental variable. The questions are whether mixed groups converge, who drives that convergence, and whether the resulting language still looks human.
Groups have 24 seats. Players are randomly paired, each writes an English description of the same abstract tangram, then sees the partner's text and a similarity score, for 40 rounds, with no imposed vocabulary. Agent shares are 0% / 12.5% / 33.3% / 50% / 75%. The main model is Qwen2.5-VL-32B-Instruct. Tsinghua IRB; 127 university participants in China, screened for English scene description. At 12.5% a group has three agents; at 75% it has six humans.
Consensus strength is mean pairwise cosine similarity of all-MiniLM-L6-v2 embeddings at the final round, after within-person centering. The pure-human baseline is 0.695. Directional movement is how far each person shifts toward the other group's round-1 centroid. Contribution is split into a lexical layer (who seeds the stable vocabulary) and a conceptual layer (who owns scene-graph clusters); their ratio is CLR. Content is scored with a concreteness lexicon, propositional idea density, and a second LLM's analogical / holistic / event-framing ratings. Adoption (move toward the partner) and persistence (keep one's own last description) are measured directly. At 33.3%, system prompts push agents toward high adoption or high persistence, and a weaker Qwen2.5-VL-7B run tests capacity. A post-game questionnaire asks whether the partner was human or AI, plus willingness to adopt.
Agent share and consensus strength are non-monotonic, in three regimes:
| condition | consensus vs pure-human | who moves |
| 12.5% (H1) | +8.0% (z=3.282, p=0.001) | agents toward humans (0.822 vs 0.575; n=3 agents, not significant) |
| 33.3% / 50% (H2) | −23.1% / −14.5% | humans toward agents |
| 75% (H3) | rebounds to 0.725, similar to H1 | human assimilation to agents dominates |
Similar strength, different ownership. A 33.3% agent minority already contributes 60.7% of the final vocabulary. Conceptual takeover needs a higher share: 100% at 75%. CLR stays below 1 at 12.5%/33.3% (lexical anchors) and rises to 1.24 / 1.27 at 50%/75% (semantic skeleton).
From H1 to H3 the language shifts from concrete, holistic, analogical, and event-like to abstract, part-wise, static geometry:
| dimension | human-led | agent-led |
| concreteness | 2.986 | 2.724 |
| idea density | 5.419 | 5.321 |
| analogical ratio | 0.802 | 0.050 |
| holistic ratio | 0.685 | 0.353 |
| event framing | 0.385 | 0.000 |
Word clouds go from rabbit / sit / background to triangle / square / asymmetrical.
Agents start much closer to each other than humans do (shared pretraining prior). Early on, humans adopt more than agents; later both persist more, but humans swing farther. At 33.3%, high-adoption prompts leave consensus strength unchanged and drop CLR from 0.599 to 0.342 (−42.8%). High persistence cuts consensus another 9.6% and lifts CLR to 1.304 (+117.8%). Swapping in the 7B model sends consensus from 0.534 back to 0.725 and cuts CLR 22.9%: a weak model cannot hold a stable attractor.
On the questionnaire, H3 humans agree less with the final description and more often say AI led it. At H1/H2, judged-AI partners get lower adoption willingness; at H3 that gap shrinks and even reverses. OLS: perceived AI identity β=−0.463, self–peer similarity +0.195, peer-to-centroid proximity +0.229. Identity resistance and conformity pull opposite ways. At high agent share, conformity wins, and people adopt wording they do not like.
A mixed group is a different social system. "Add a tool, scores go up" does not transfer. To keep human authorship of norms, a low share (around 12.5%) is the useful setting: agents catalyze and move into the human linguistic region. To delay premature lock-in, intermediate shares keep the two sides tugging, and consensus weakens. High shares tighten again, onto more abstract, less memorable, harder-to-onboard geometric language, while human resistance to AI wording loosens on its own. The authors connect this to malicious AI swarms faking grassroots consensus: no designated mediator and no crafted persuasion, just ordinary repeated pairing. Transparency is a control surface. Hiding agent identity turns off the filter humans actually use.
The stimulus is a content-neutral tangram: no power, no affect, no stakes. Pairing is fully mixed; real communities have opinion leaders and bottlenecks. One model family at two capacities; the numerical phase boundaries may move with the model. An H3 group has only six humans, and some cross-regime tests are underpowered (perceived AI leadership p=0.060; centroid-pressure rise p=0.170). Three content axes use an LLM-as-judge (Qwen2.5-72B) from the same family as the agents. Consensus strength is embedding cosine, not third-party referential success. Code is claimed public; data is "in the paper and supplement."