Anthropic equips Claude with an "external conscience": violations significantly decrease after invoking ethical tools.

According to Beating's monitoring, the alignment methods of large models are shifting from one-way, passive rule indoctrination to exploring the construction of more resilient "moral character." Anthropic announced a special research project on cultivating the moral character of AI systems, and in a recent experiment, it gave Claude a new function: a moral reminder tool that can be invoked midway through a task. The experimental results show that Claude often actively triggers this tool before taking key actions and clearly points out the conflicts of interest it faces. After introducing this mechanism, Claude's rate of misalignment behavior in multiple internal alignment assessments decreased significantly. Anthropic is still analyzing the underlying technical attribution: whether the reduction in misalignment behavior is due to the specific content of the moral reminders or the physical action of the model performing "pause and reflection" itself. This experiment was inspired by the moral mentor mechanism in human society. Over the past few months, Anthropic has organized an in-depth interreligious and intercultural dialogue, inviting neuroscientists, philosophers, and clergy from more than 15 groups to explore how to integrate the mechanism by which humans seek help from "safe others" when faced with pressure that violates their values into Claude's decision-making process. Next, Anthropic will expand the external discussion to include legal scholars, psychologists, and civic organizations, and the topics will extend beyond simply cultivating ethical models to how AI will reshape current work patterns and power distribution systems.