AI Model Groupthink

A small web tool asked a panel of 12 large language models to give independent "top three" answers to everyday ranking questions, normalized for equivalent responses, and scored each model against the consensus of the other eleven (full credit for correct item in the correct rank, half for correct item in the wrong rank). Results show a clear conformity gap by origin: the four most conformist models are Chinese (DeepSeek 0.59, Kimi K2 0.47, Tencent Hy3 0.46, GLM 0.45), five Chinese models average 0.47 versus 0.38 for six US models and 0.36 for the lone European entrant (Mistral). The most contrarian models were Meta’s Llama 4 Scout (0.34) and Anthropic’s Claude Haiku (0.32). Queries ran mainly through OpenRouter using cheaper "flash" or small variants, often with reasoning disabled and default temperatures left unchanged.

Several mechanisms could explain the pattern: large-scale distillation of model outputs, tight reuse of open weights and synthetic data among Chinese labs, differences in model size/tier and knowledge cutoffs, post-training tuning that increases personality in some US models, and an English-centric training canon that favors canonical answers. Important caveats include a small, convenience dataset, an arbitrarily composed panel, and the fact that agreement is not accuracy. The deeper implication is a risk of self-reinforcing consensus: models that echo each other can amplify and monetize specific answers, enabling feedback loops where consensus becomes de facto truth and can be gamed into cultural, political, or factual influence.