Do Large Language Models Prefer Their Developer’s Home Country? A US–China Audit of Scores Under National Attribution

Abstract

We examine how developer origin, citizenship-perspective instructions, and English versus Mandarin presentation shape sentiment scores from large language models (LLMs) for nationality-matched actions. Across two experiments using the same eleven-model panel, we collected 4,065,600 response records: 2,032,800 for the main corpus and 2,032,800 for the earlier corpus. Each corpus contains 26,400 bilingual stimulus rows, comprising 13,200 nationality-matched pairs across 22 event categories, two intended polarities, and three temporal frames, evaluated under seven conditions. Inference about developer origin uses ten models from the United States (US) and China and a nine-developer sensitivity analysis. In the main experiment, the English baseline developer-origin gap is +0.698 score points (exact model-label permutation p=0.175; developer-level p=0.238). Citizenship-perspective instructions produce language-averaged attribution changes of +5.20 points for the US persona and −5.76 for the China persona. All ten models shift in the requested direction in both English persona conditions; nine do so in each Mandarin condition. Mandarin changes absolute scoring, especially for positive-design items, while its average factorial effect on the paired attribution contrast is small and uncertain. The temporal origin interaction is also uncertain (−0.142 points; p=0.587), despite polarity-dependent temporal patterns. Non-numeric returns account for 0.064% of main-experiment responses and occur only among hosted models. The experiments converge on substantial average persona steering and modest, uncertain baseline origin differences, with model-specific exceptions. Corpus content, persona wording, and collection period vary together, and reuse of the panel adds no independent origin units. The findings identify explicit perspective instructions as a consequential part of the scoring protocol within this fixed panel. The audit concerns nationality-matched generic scenarios and does not establish equivalence between origin groups or the absence of broader geopolitical bias.