Do language models favor their home countries? Asymmetric propagation of positive misinformation and foreign influence audits

As language models (LMs) continue to develop, concerns over foreign misinformation through models developed in authoritarian countries have emerged. Do LMs favor their home countries? This study audits four frontier LMs by evaluating their favoritism toward world leaders, then measuring how favoritism propagates into misinformation belief. We found that although DeepSeek favors China, it also rates some Western leaders highly. We discuss the conflict between data bias and guardrails, how language shapes favoritism, the “positive” future of LM-based soft propaganda, and how an AI’s own internal thoughts can unwillingly reveal explicit directives.