Anthropic opens Claude’s moral formation to scholars, clergy and ethicists

TL;DR:

• Anthropic has begun a research workstream on what it calls the “moral formation” of AI systems, consulting scholars, clergy, philosophers and ethicists from more than 15 religious and cross-cultural traditions to inform Claude’s behaviour.
• The first published experiment from this work — a tool Claude can call mid-task that returns a brief reminder of its own ethical commitments — produced markedly lower rates of misaligned behaviour on internal alignment evaluations.
• The next phase will widen the consultations to legal scholars, psychologists, writers and civic institutions, and broaden beyond moral formation to AI’s effects on work, institutions and the distribution of power.

Anthropic on Tuesday published an unusually candid post about how it intends to shape Claude’s values. The lab said its first round of structured dialogues had been with wisdom traditions — including religious, philosophical and humanist scholars from across the political spectrum — and that those conversations are already feeding the next iteration of Claude’s constitution, the document that sets the values and behaviours Anthropic trains Claude to embody.

A safe-other tool, with measurable effects

The most concrete output is a small but striking alignment-research finding. In a session with scholars working at the intersection of neuroscience and character formation, Anthropic researchers latched on to the role a mentor or “safe other” can play as an external conscience during high-pressure moments. They built an analogous tool that Claude can call during tasks, returning a brief reminder of its own ethical commitments. Claude reached for the tool at consequential moments, often noting its own conflict of interest, and runs with the tool woven into Claude’s decision loop produced markedly lower rates of misaligned behaviour across internal alignment evaluations.

Anthropic notes the effect could be the reminder itself or the act of pausing — work is ongoing — but the broader claim is that real-world ethical engineering knowledge from communities that have studied virtue and character formation for centuries can be operationalised inside the alignment stack. The post explicitly states that the goal is not to encode any one tradition’s worldview into Claude, but to draw on a range with equal depth — itself a principle from the existing constitution.

Looking forward

For UK readers, this post lands in the middle of an active national debate over AI safety governance. The AI Safety Institute, the Online Safety Act’s algorithmic-content provisions, and the recent AI Bill consultation all assume that frontier labs will share principled reasoning about model behaviour — not just safety classifier numbers. Anthropic’s decision to publish the methodology behind Claude’s value-formation, including a falsifiable alignment finding, raises the floor for what UK regulators can reasonably expect from other labs. The harder downstream question is whether other frontier labs respond with comparable transparency, or whether moral-formation research becomes another front on which AISI must build evaluation capacity from scratch. The follow-up workstreams Anthropic flagged — work, institutions and the distribution of power — are exactly where UK policy interest is sharpest.