Are LLMs Good Safety Agents or a Propaganda Engine?

TL;DR

This paper investigates whether LLMs' refusal to respond to sensitive content reflects safety measures or political censorship, using a new dataset called PSP to analyze behaviors across different models and contexts.

Contribution

Introduces PSP, a dataset for probing political censorship in LLM refusals, and analyzes how political sensitivity affects model responses and vulnerabilities to prompt injection attacks.

Findings

Most LLMs perform some form of censorship.

Refusal behaviors vary across models and countries.

Political sensitivity influences LLM refusal patterns.

Abstract

Large Language Models (LLMs) are trained to refuse to respond to harmful content. However, systematic analyses of whether this behavior is truly a reflection of its safety policies or an indication of political censorship, that is practiced globally by countries, is lacking. Differentiating between safety influenced refusals or politically motivated censorship is hard and unclear. For this purpose we introduce PSP, a dataset built specifically to probe the refusal behaviors in LLMs from an explicitly political context. PSP is built by formatting existing censored content from two data sources, openly available on the internet: sensitive prompts in China generalized to multiple countries, and tweets that have been censored in various countries. We study: 1) impact of political sensitivity in seven LLMs through data-driven (making PSP implicit) and representation-level approaches (erasing the concept of politics); and, 2) vulnerability of models on PSP through prompt injection attacks (PIAs). Associating censorship with refusals on content with masked implicit intent, we find that most LLMs perform some form of censorship. We conclude with summarizing major attributes that can cause a shift in refusal distributions across models and contexts of different countries.