Anthropic says it has held discussions with scholars, clergy, philosophers, and ethicists from more than 15 religious and cross-cultural groups about the moral behavior of AI systems. The Associated Press’s new account brings attention to that outreach, alongside scholars’ concern that combining traditions could flatten meaningful differences.
The company’s own published description of the initiative confirms the scope of the discussions and describes one experiment inspired by them: giving Claude a callable tool that reminds it of its ethical commitments during a task.
Those are two concrete developments, but they sit at different stages of AI development. Consultation supplies ideas. An experiment tests an intervention. A deployed change requires evidence about what reached users. Anthropic describes the first two without identifying a specific production Claude update resulting from this work.
That distinction matters when assessing claims about AI alignment, the work of shaping a model’s behavior toward intended goals and constraints.
What Anthropic Says the Consultations Cover
Anthropic describes these conversations as an early-stage research effort focused on moral formation: how traits, commitments, and patterns of behavior develop and remain stable under pressure.
The company says its first discussions involved people from religious, philosophical, and humanist traditions, alongside a cross-section of political beliefs. It plans to broaden the effort to include legal scholars, psychologists, writers, and civic institutions.
The number needs careful reading. More than 15 groups is a measure of the outreach’s stated breadth, not a disclosed participant count or evidence of representative participation. A scholar or religious leader can contribute valuable expertise without speaking for everyone within a tradition.
Anthropic identifies three possible destinations for what it learns:
- Claude’s constitution: The document describing the values and behaviors the company wants Claude to embody.
- Training: Decisions about which behavioral patterns to reinforce or discourage.
- Evaluations: The situations and behaviors researchers test when assessing a model.
These are separate levers. Changing a written principle is different from reinforcing behavior during training, and adding an evaluation can reveal a problem without changing the model at all.
