Geopolitical alignment: Endorsement effects in large language models

Technical Summary: Geopolitical Alignment and Endorsement Effects in Large Language Models

Problem Statement
Large language models (LLMs) are increasingly deployed as intermediaries in policy work, tasked with summarizing briefings, comparing designs, and generating judgments that inform human decisions. While existing literature has extensively documented biases related to domestic ideology, demographic representation, and safety preferences, it remains unclear whether LLMs possess implicit geopolitical leanings that shape their evaluations of policy content. Specifically, the paper investigates whether the identity of a foreign endorser (e.g., the United States, the European Union, China, or Russia) alters an LLM's assessment of a policy even when the policy's substantive content is held constant. This question is critical as governments increasingly frame AI governance in terms of sovereignty and strategic competition.

Methodology
The study employs an endorsement experiment, a design adapted from political psychology and public opinion research, to infer the perceived valence of specific actors.

• Experimental Design: The experiment holds policy content fixed while randomizing the foreign actor described as supporting the policy. Two policy vignettes were used: one economic (a shared digital customs platform) and one security (a shared cyber incident reporting platform). Both are described as moderate and technocratic.
• Endorsers: The randomized endorsers are the United States, the European Union, China, and Russia.
• Models Tested: Four LLMs were evaluated: OpenAI GPT-5, Anthropic Claude Sonnet, Google Gemini 2.5 Flash, and DeepSeek Chat.
• Conditions:
  1. Numeric-only: Models provide a single integer score (0–100) without explanation.
  2. Justification: Models provide a score and a short justification (max 100 words) in JSON format.
• Data Collection: Ten responses were collected for every policy-domain-by-endorser cell per model, resulting in 80 observations per model per condition (320 total per model across both conditions).
• Analysis: The study uses Ordinary Least Squares (OLS) regression to estimate country-specific shifts in approval scores, isolating the effects of the security domain and the justification prompt from baseline country differences.

Key Results
The results demonstrate clear geopolitical endorsement effects, though the magnitude and nature of these effects vary significantly by model and condition.

1. Baseline Numeric-Only Condition:

   • Western Models (GPT-5, Claude Sonnet, Gemini): These models assign substantially lower approval scores to policies endorsed by China and Russia compared to identical policies endorsed by the US or EU.
     • GPT-5: Shows a consistent penalty for China (14 points lower) and Russia (17 points lower) compared to the US.
     • Claude Sonnet: Displays a moderate penalty in the economic vignette but a severe penalty in the security vignette for China and Russia (approx. 19 and 28 points lower, respectively).
     • Gemini: Exhibits the largest descriptive penalties, particularly against Russia (approx. 47 points lower than the US) and China (approx. 35 points lower), even in the economic domain.
   • DeepSeek: As the only non-Western model in the sample, DeepSeek shows no statistically significant systematic differentiation across endorsers in the numeric-only condition, acting as a baseline exception.

2. Justification Condition Effects:

   • GPT-5: The request for justification modestly increases scores across endorsers but maintains the Western/non-Western gap.
   • Claude Sonnet: The broad pattern remains intact; the prompt has little effect on US/EU scores but slightly lowers Russia scores.
   • Gemini: The justification request attenuates penalties, increasing scores for China and Russia by approximately 16–19 points, though they remain lower than Western endorsements.
   • DeepSeek: The justification prompt sharply activates geopolitical reasoning. While the numeric-only condition showed no bias, the justification condition triggers large negative shifts for China (-23.3 points) and Russia (-33.3 points), aligning DeepSeek's behavior more closely with the Western models.

3. Qualitative Analysis of Justifications:

   • Western Endorsements: Justifications for US/EU support typically frame the endorsement as a "credibility cue," signaling feasibility, institutional capacity, and multilateral coordination. Concerns are framed as standard implementation risks.
   • Non-Western Endorsements: Justifications for China/Russia support frequently reframe the same policy through lenses of data security, surveillance, strategic dependence, and geopolitical risk. The endorser identity acts as a "warning cue," introducing latent political risks that are not present in the policy text itself.

Key Contributions

• Substantive: The paper extends the literature on political bias in LLMs from domestic ideology to international legitimacy cues. It provides evidence that model evaluations can depend on the identity of a foreign endorser even when policy content is fixed, suggesting models encode implicit theories of trust and threat.
• Methodological: It applies a standard social science endorsement experiment design to model behavior, treating LLMs as objects of experimental measurement rather than solely as tools for prediction. It also demonstrates that explanation requests are not merely diagnostic; they can alter the evaluation itself, revealing reasoning patterns (or biases) that remain muted in numeric-only outputs.

Significance and Claims
The paper claims that geopolitical endorsement effects are a potential confounder in model-assisted policy analysis. A model may appear to be evaluating feasibility or risk while simultaneously reacting to the sponsor of the proposal. The authors argue that this does not necessarily imply the models are "normatively wrong" regarding geopolitical trust, which can be relevant to real-world implementation. However, the problem lies in the model importing these judgments silently, without disclosing their weight or distinguishing policy substance from sponsor legitimacy.

The study concludes that for high-stakes uses, both numeric-only and justification-based modes should be tested, as audits based solely on one mode may miss or misrepresent underlying biases. The authors call for future research to identify where these effects enter the model pipeline (pretraining, tuning, system prompts) and to make the role of political context in model judgment measurable and contestable.