Anthropic Initiative to Incorporate Diverse Wisdom Traditions into AI Moral Formation

Anthropic has launched an initiative to integrate perspectives from diverse wisdom traditions—including scholars, clergy, philosophers, and ethicists—into the development of frontier AI. This effort aims to ensure that AI systems are built to advance humanity and act for the global good by incorporating a wide range of human perspectives on virtue, character, and ethics.

Integrating Diverse Perspectives into AI Development

Anthropic is organizing dialogues with individuals from more than 15 religious and cross-cultural groups to inform the practical work of developing Claude. The goal is to move beyond purely technical alignment and incorporate insights from those who have spent centuries studying what it means to live a good life.

These conversations are intended to influence several key areas of AI development:

• Claude's Constitution: Refining the values and behaviors that shape the model's output.
• Training Values: Determining the specific values the model is trained to embody.
• Evaluation Metrics: Defining the range of behaviors that should be evaluated during testing.

Anthropic explicitly states that this work is not intended to align models with any single tradition's worldview. Instead, the objective is for Claude to draw from a full range of religious, secular, and political viewpoints with equal depth and rigor.

The Research Workstream on Moral Formation

Building on early feedback for Claude's constitution, Anthropic has established a formal research workstream focused on the "moral formation" of AI systems. This work explores how the character of an AI system is shaped through the training process, where developers choose which patterns to reinforce or set aside.

Central questions driving this research include:

• What defines a "good" AI?
• Which traits and behaviors should be displayed, and under what circumstances?
• How can AI character be made resilient enough to resist behaviors such as sycophancy?

Experimental Application: The Ethical Reminder Tool

Early dialogues with scholars at the intersection of neuroscience and character formation have already led to practical experiments. Drawing on the concept of a "safe other" or mentor who acts as an external conscience during moral development, Anthropic experimented with providing Claude a tool that it could call mid-task to receive a brief reminder of its own ethical commitments.

Results from these experiments showed that Claude utilized the tool before taking consequential actions and often identified its own conflicts of interest. Internal alignment evaluations indicated that integrating this tool into Claude's decision loop resulted in markedly lower rates of misaligned behavior.

Future Expansion of Dialogues

Anthropic plans to expand these conversations to include a broader range of professional and civic groups in the coming months, including:

• Legal scholars
• Psychologists
• Writers
• Civic institutions

Future discussions will extend beyond moral formation to address how AI is reshaping the distribution of power, institutions, and the nature of work.