This study explores the stability and ideological orientation of political responses produced by various large language models (LLMs) in French. We designed a standardised experimental protocol based on a questionnaire inspired by the Political Compass, aimed at measuring the economic and socio-cultural positions of each model across 62 political statements. Eleven models from various organisations and countries were tested, each subjected to twenty independent runs to assess intra-model variability. The analysis focuses on response consistency, inter-model differences, and the presence of implicit political orientations. The results show that, despite a general degree of stability, significant variations appear from one run to the next and between models, reflecting the impact of architectures, training data, and moderation mechanisms. This study proposes a comparative evaluation protocol for LLMs in the context of political information and underscores the importance of accounting for implicit biases in the use of these systems.

翻译：本研究以法语为工作语言，探讨了多种大语言模型（LLM）所产生政治回应的稳定性与意识形态倾向。我们设计了标准化实验方案，采用基于政治指南针（Political Compass）的问卷，旨在测量每个模型对62个政治陈述的经济与社会文化立场。来自不同组织和国家的11个模型接受了测试，每个模型均历经二十次独立运行以评估模型内部变异性。分析聚焦于回应一致性、模型间差异及隐性政治倾向的存在。结果显示，尽管具备总体稳定性，但不同运行轮次之间及模型之间均出现显著变异，这反映了架构、训练数据和调节机制的影响。本研究提出了一种面向政治语境下LLM的比较评估协议，并强调了在使用这些系统时考虑隐性偏见的重要性。