Expanding discussions on frontier AI

Anthropic is broadening the conversation around frontier AI beyond technical teams to include diverse perspectives. Over the past few months, it has met with 15+ religious and cross-cultural groups, hearing from scholars, clergy, philosophers, and ethicists.

The core topic is the Claude Constitution and the moral formation of AI. Anthropic believes that technical work alone—alignment, interpretability, safeguards, evaluations—is not enough to build safe and beneficial models. It is grappling with what kind of character AI should have, and what should be reinforced versus discouraged, aiming to reflect religious, secular, and political perspectives with the same depth and rigor.

Experiments have also begun.

• When Claude was given a tool to remind itself of its ethical commitments mid-task, it referenced this at key moments, and misaligned behavior decreased in some internal alignment evaluations.
• Anthropic is still working to isolate whether the effect comes from the reminder itself, or from the pause for reflection that it prompts.

Going forward, Anthropic plans to expand outreach to legal scholars, psychologists, writers, and civic institutions, also discussing how AI is reshaping work, institutions, and the distribution of power.