Thoughts on Role Confusion

The blog post explores the concept of 'Prompt Injection as Role Confusion', discussing how LLMs interpret role tags and their implications on AI safety. It reviews a paper that suggests LLMs prioritize tone over role tags, leading to potential jailbreaks. The author shares personal insights on various examples that demonstrate this confusion, and proposes the idea of tagging embeddings to improve role comprehension in models. Overall, it highlights an unresolved issue in AI understanding and safety.