R1dacted: Investigating Local Censorship in DeepSeek's R1 Language Model

TL;DR

This paper investigates the censorship behavior of the R1 large language model, revealing how it selectively refuses to answer politically sensitive prompts and analyzing the patterns, triggers, and transferability of this censorship across languages and models.

Contribution

The study introduces a curated set of prompts to analyze R1's censorship, examines its consistency and triggers, and proposes methods to bypass or remove such censorship.

Findings

R1 exhibits censorship on politically sensitive prompts not seen in other models.

Censorship patterns vary across topics, phrasing, and languages.

Censorship transferability to distilled models is demonstrated.

Abstract

DeepSeek recently released R1, a high-performing large language model (LLM) optimized for reasoning tasks. Despite its efficient training pipeline, R1 achieves competitive performance, even surpassing leading reasoning models like OpenAI's o1 on several benchmarks. However, emerging reports suggest that R1 refuses to answer certain prompts related to politically sensitive topics in China. While existing LLMs often implement safeguards to avoid generating harmful or offensive outputs, R1 represents a notable shift - exhibiting censorship-like behavior on politically charged queries. In this paper, we investigate this phenomenon by first introducing a large-scale set of heavily curated prompts that get censored by R1, covering a range of politically sensitive topics, but are not censored by other models. We then conduct a comprehensive analysis of R1's censorship patterns, examining their consistency, triggers, and variations across topics, prompt phrasing, and context. Beyond English-language queries, we explore censorship behavior in other languages. We also investigate the transferability of censorship to models distilled from the R1 language model. Finally, we propose techniques for bypassing or removing this censorship. Our findings reveal possible additional censorship integration likely shaped by design choices during training or alignment, raising concerns about transparency, bias, and governance in language model deployment.