This research reveals systematic patterns in how large language models selectively filter or refuse to provide information on political topics.

• Identifies both hard censorship (outright refusals) and soft censorship (selective omission) in LLM responses
• Documents significant variations in censorship practices across different LLM systems
• Demonstrates how political topics receive inconsistent treatment compared to non-political queries

For security professionals, this work highlights critical transparency issues in AI information systems that may influence public discourse and decision-making without users' awareness.

What Large Language Models Do Not Talk About: An Empirical Study of Moderation and Censorship Practices