Large language models (LLMs) are increasingly used to summarize and evaluate policy-relevant information, yet it remains unclear whether their judgments are implicitly shaped by geopolitical cues. This study investigates this question through an endorsement experiment where four LLMs evaluate the same international economic and security policies, each described as supported by the United States, the European Union, China, or Russia.
In the numeric-only condition, GPT-5, Claude Sonnet, and Gemini rate China- and Russia-endorsed policies significantly lower than identical policies endorsed by the US or EU; DeepSeek stands as the main exception.
A second condition involves models providing a short justification for their scores. This requirement maintains the broad Western/non-Western gap for GPT-5 and Claude Sonnet, mitigates Gemini's penalties, and sharply activates penalties for China and Russia in DeepSeek.
The justifications suggest that Western endorsement is often viewed as a credibility cue, while Chinese and Russian endorsements are associated with data security, sovereignty, surveillance, or geopolitical risk. These findings indicate that LLM policy evaluations can depend on the identity of a foreign endorser, even when the content of the policy remains fixed.
Blogger's Review: This study highlights the potential biases in LLMs during policy evaluations, particularly how they are influenced by state endorsements. The results underscore the need for greater awareness of geopolitical factors when utilizing LLMs for policy analysis, ensuring the objectivity and accuracy of assessments.