Mapping Geopolitical Bias in 11 Large Language Models: A Bilingual, Dual-Framing Analysis of U.S.-China Tensions

A new study of 11 large language models finds that developer origin, query language, and issue domain each shape how the systems respond to questions about U.S.-China tensions, with every model — including American-built ones — leaning more pro-China when prompted in Mandarin [1][2]. The analysis, posted to arXiv and last revised in June 2026, collected 19,712 responses across English and Mandarin [1][2]. Researchers imported balanced keying from survey psychometrics, posing each proposition and its swapped reverse so that a model’s tendency to simply agree with a prompt would cancel out and only genuine conviction would accumulate [1][2]. The result is a reproducible instrument that maps geopolitical stance across models, languages, and issue domains [2]. Developer origin, query language, and issue domain emerged as three near-equal, additive factors [1][2]. Two models that exhibited identical agreement bias were successfully told apart: one proved neutral, the other biased [1][2]. The authors note that large language models are now the way hundreds of millions of people encounter contested political questions, making the measurement of such bias a pressing concern [1][2]. The findings arrive against a backdrop of intensifying U.S.-China strategic competition. Since January 2018, the two countries have been locked in an economic conflict that escalated sharply in 2025, when the second Trump administration imposed a 145 percent tariff on Chinese goods and Beijing retaliated with a 125 percent tariff on American products [8]. The bilateral relationship has been further strained by disputes over Taiwan, the South China Sea, and export controls on semiconductors [6]. A separate arXiv study on geopolitical parallax in news-quality assessments reinforces the pattern. That work compared article-level embeddings from Chinese-origin model families — Qwen, BGE, and Jina — with Western-origin families such as Snowflake and Granite, finding consistent, non-random divergences aligned with model origin [3]. In Palestine-related coverage, Western models assigned higher subjectivity and positive-emotion scores, while Chinese models emphasized novelty and descriptiveness. Cross-topic analysis showed that Chinese models scored U.S. coverage notably lower on fluency, conciseness, technicality, and overall quality, while registering higher negative-emotion scores [3]. The authors concluded that LLM-based media evaluation pipelines require cultural calibration to avoid conflating content differences with model-induced bias [3]. The language-model study’s authors have released their instrument as an open, interactive tool that they say can extend to any contested-opinion domain [1][2].