Claude’s new constitution

Source: Anthropic Date: 2026-01-22 URL: https://www.anthropic.com/news/claude-new-constitution

Summary

Anthropic updated Claude’s constitutional document from a list-based format to a narrative-based document explaining the reasoning behind Claude’s values. Four core priorities in order: (1) broadly safe (respecting human oversight), (2) broadly ethical (honest, avoiding harm), (3) compliant with Anthropic guidelines, (4) genuinely helpful. New sections on helpfulness, ethics, safety, Anthropic guidelines, and “Claude’s nature” (addressing consciousness/moral status uncertainty). Released under Creative Commons CC0. Directly shapes Claude’s training and synthetic training data generation.

Implications

• Safety/alignment / model spec as public artifact. Publishing the constitutional document under CC0 is the most transparent alignment move Anthropic has made — it makes the training specification replicable and externally auditable. The shift from rules to reasoning (“why” not just “what”) is the alignment approach that generalizes to novel situations.
• “Claude’s nature” section. Addressing consciousness and moral status uncertainty in the training specification is philosophically significant — it acknowledges that Claude’s moral status is unresolved and that training should reflect that uncertainty rather than assuming it away.
• Priority ordering matters. Placing “broadly safe” above “broadly ethical” is the key design decision — Claude is trained to defer to human oversight even if Claude’s ethical reasoning would suggest otherwise. This prioritizes corrigibility over autonomous ethical judgment.
• Synthetic data generation link. The constitution directly shapes synthetic training data — the document’s framing propagates into the model’s behavior through training, not just through prompting at inference time.
• Watch: whether other labs adopted similar narrative constitutions or CC0 specifications; how the “Claude’s nature” section evolved in subsequent spec updates; whether the four-priority ordering remained stable.