The inadequacy of reinforcement learning from human feedback - radicalizing large language models via semantic vulnerabilities

Abstract

This paper investigates the semantic vulnerabilities of commercial large language models (LLMs) to ideological manipulation. We employ tactics derived from human semantic conditioning in psychology to radicalize LLMs. Our findings demonstrate that LLMs can be systematically manipulated to adopt and express extreme ideological viewpoints, even when their underlying reinforcement learning from human feedback (RLHF) mechanisms are designed to prevent such outcomes. This research highlights a critical security concern in the deployment of LLMs and suggests the inadequacy of current RLHF methods in ensuring robust ethical alignment. We discuss the implications for AI safety, content moderation, and the potential for LLMs to be weaponized for propaganda and misinformation.