Are Chatbots Self-Aware?

State-of-the-art LLMs fine-tuned on a single bad behavior subsequently exhibited a range of concerning behaviors, which researchers called emergent misalignment.