Stratified regulation: unpacking value alignment of Chinese LLMs

While previous studies examine how biases in training data and models shape LLMs’ values, less attention has been paid to how political orientations vary across different deployment modes. This paper proposes a stratified regulation framework, arguing that the Chinese government balances technological innovation with information regulation. Cloud-based models, accessible to the public, are strictly regulated to ensure ideological security, whereas local deployments face relaxed restrictions to facilitate innovation. Using 100 ideological prompts and stance analysis, the paper examines how information regulation affects DeepSeek R1 in local versus cloud contexts. The results reveal that locally deployed models display more diverse perspectives, while the cloud versions exhibit a pronounced ideologically positive orientation, reflecting real-time alignment and official-tone templates. English prompts elicit more critical responses, whereas Chinese prompts produce more political rightness content. These findings demonstrate a stratified regulatory approach: broad training-phase filtering across all deployments and real-time alignment for cloud models. This study is the first to document the value differences across different deployment contexts, providing methods and data to quantify LLM regulation and advance the understanding of technological governance in China.