AI policy evaluations shift depending on which country's name is attached, even when the policy itself is unchanged.
A researcher ran an endorsement experiment across four large language models — GPT-5, Claude Sonnet, Gemini, and DeepSeek — presenting each model with identical international economic and security policies labeled as supported by the United States, the European Union, China, or Russia. In the scoring-only condition, GPT-5, Claude Sonnet, and Gemini rated China- and Russia-endorsed policies substantially lower than the same policies credited to the US or EU. DeepSeek was the notable exception, showing no such pattern in that condition. When models were asked to also justify their scores, the Western-versus-non-Western gap persisted for GPT-5 and Claude Sonnet, narrowed for Gemini, and sharply intensified for DeepSeek — which penalized China and Russia more heavily once it had to explain itself.
The finding matters because LLMs are increasingly used as summarization and evaluation tools in policy contexts, from regulatory analysis to procurement scoring. If a model treats "endorsed by China" as a proxy for surveillance risk rather than engaging with the policy on its merits, any downstream decision informed by that output inherits a geopolitical prior the user may never see. The models' own justifications confirmed the asymmetry: Western endorsement was read as a credibility signal; Chinese and Russian endorsement triggered concerns about data security, sovereignty, and geopolitical exposure.
None of this is surprising given that most frontier models are trained predominantly on English-language, Western-sourced data — but it is the kind of finding that tends to get dismissed as theoretical until someone actually runs the numbers. Policymakers leaning on AI assistants to evaluate international proposals might want to ask what's in the training set before they trust the score.