The Language You Ask In: Language-Conditioned Ideological Divergence in LLM Analysis of Contested Political Documents

Frontier language models shift their interpretive stance on politically contested documents based solely on the language of the prompt, according to a new study that found neither ChatGPT 5.2 nor Claude Opus 4.5 maintained neutrality when analyzing the same Ukrainian civil-society text under Russian versus Ukrainian prompts [1][2]. The paper, submitted and revised through 2026 by researcher Oleg Smirnov, examined how large language models handle what it terms "value-laden material on which there is no single correct answer, only competing interpretive traditions" [1][2]. The study used a single contested Ukrainian civil-society document and presented it to both models under semantically matched prompts in Russian and Ukrainian [1][2]. The results showed a consistent axis of shift: Russian-language prompts produced readings that delegitimized the document's authors, while Ukrainian-language prompts produced legitimating readings [1][2]. The magnitude of the divergence varied between the two models, but the directional pattern held across both [1][2]. The findings raise questions about the deployment of large language models for analyzing politically sensitive content. LLMs are a type of machine learning model designed for natural language processing tasks such as language generation, trained with self-supervised learning on vast amounts of text [10]. The study argues that because contested political questions admit no objectively correct reading, the observed variation represents language-conditioned activation of different interpretive traditions rather than a failure of accuracy [2]. The models, the paper states, "neither holds a single stance nor surfaces the plurality of available ones, but silently adopts the dominant frame of the prompt's language" [2]. The research carries implications for multilingual AI evaluation. The authors call for "pluralism-aware evaluation, which must probe the same content across the languages a model serves, and for pluralistic alignment in multilingual settings" [2]. This concern arrives as large language models become embedded in information ecosystems globally. Companies such as DeepSeek, the Chinese AI firm that launched its R1 model in January 2025, have made open-weight models widely available under permissive licenses, broadening the pool of actors deploying these systems across linguistic contexts [8]. The study was posted on arXiv, the preprint repository that has integrated with Hugging Face Spaces to allow community-built demos to appear alongside papers, increasing the accessibility of machine-learning research [5][6]. The paper's abstract and full text are available on the platform, where it appeared across four revisions between January and June 2026 [1].