An Analysis of Chinese Censorship Bias in LLMs

When a large language model (LLM) is trained on text shaped by state censorship, those biases implicitly impact the outputs of the model. The authors define this phenomenon as censorship bias: a model trained on sanitized content is less likely to reflect prohibited views and more likely to reflect permitted ones, particularly when interacted with in a language predominantly used in a region with strong censorship laws. They introduce a methodology for identifying and measuring censorship bias and apply it to popular LLMs, including building CensorshipDetector, a Chinese-language classifier that distinguishes sanitized from non-sanitized text with 91% accuracy. The evaluation finds evidence of censorship bias across all models tested and discusses harms (notably the export of domestic information manipulation to diaspora populations) and mitigations.