Why AI models favor their makers' countries—and when they do it most

Seven major LLM labs' base models show minimal geopolitical bias, but their chat variants developed by the same labs—which undergo alignment and instruction-tuning—dramatically favor their developer's country or region. Alibaba's Qwen shifts from neutral on China to strongly pro-China. The bias is language-dependent: Mistral favors France only when prompted in French. This reveals that geopolitical preferences are baked in during post-training choices, not inherited from internet text.