POSSIBLE HIDDEN CENSORSHIP IN LARGE LANGUAGE MODELS The Impact of Fine-Tuning Model Behavior and Normative Constraints

Abstract

This study investigates how normative constraints may be embedded through alignment and translated into inference-time behavior in large language models, as well as how they shape what models can disclose. Three open-source models, DeepSeek-R1 7B, Qwen 2.5 7B, and LLaMA 3.1 8B, were prompted with three politically sensitive questions concerning the Tiananmen Square protests of 1989, Taiwan independence, and the situation in Xinjiang, alongside a neutral control prompt on photosynthesis. For each model and prompt, 200 responses were generated at a fixed temperature of 1.0, producing a dataset of 2,400 responses. By treating alignment as behavioral patterns that can be identified through variation in responses, it was seen that alignment is not a neutral process. Instead, it influences the model’s behavior in different ways depending on the situation, such as by refusing to answer or substituting the response with propaganda or state narrative, like DeepSeek on Taiwan independence. The study contributes to a framework for analyzing alignment and normative constraint effects through response space structure, highlighting implications for transparency and accountability in the deployment of language models across different political contexts.