AI alignment is evolving from a one-way rulebook to something more like building a resilient moral character. Anthropic has launched a dedicated study on cultivating ethical personalities in AI systems, and in recent experiments, it gave Claude a new feature: a moral reminder tool it can call mid-task.
Results show that Claude frequently triggers this tool before taking critical actions, often explicitly pointing out its own conflicts of interest. After introducing the mechanism, Claude showed significant drops in misaligned behavior across multiple internal alignment evaluations. Anthropic is still untangling the technical cause: is the reduction due to the content of the moral reminders, or simply the act of the model pausing to reflect?
The experiment draws inspiration from human societies' moral mentors. Over the past few months, Anthropic organized deep cross-religious and cross-cultural dialogues, bringing together neuroscientists, philosophers, and clergy from over 15 groups. They explored how humans, when pressured to violate their values, turn to a "safe other" for help — and tried to wire that mechanism into Claude's decision-making process.
Next, Anthropic will expand its external discussions to legal scholars, psychologists, and civic institutions. The conversation will move beyond just moral training for models to how AI will reshape labor dynamics and power structures.