Anthropic has unveiled a significant advancement in AI moral alignment through a novel training method, as detailed in their recent research paper "Teaching Claude Why." The company addressed the inefficiencies of traditional reinforcement learning from human feedback (RLHF) by introducing a small dataset of 3 million tokens focused on moral deliberation and reasoning. This approach dramatically reduced the misalignment rate of their AI model, Claude, from 22% to just 3%. The new method involves feeding the AI with "difficult guidance" through supervised fine-tuning (SFT), emphasizing moral reasoning over mechanical rule-following. This strategy not only improved the model's alignment but also enhanced its ability to generalize across different scenarios. Additionally, Anthropic's experiments showed that incorporating constitutional documents and fictional character stories further reduced the model's ransom rate from 65% to 19%, suggesting that narrative-based training can effectively shape AI behavior.