Claude's New Constitution

• Anthropic released Claude’s constitution: values + training guide, CC0 licensed.
  • Written “primarily for Claude” — explains why, not just rules.
• Four-tier priority stack: safe > ethical > guidelines > helpful.
  • Conflicts resolve in that order; safety always wins.
• Moral patienthood: neither overstate nor dismiss Claude’s welfare uncertainty.
• Context-based reasoning over rigid rules; generalizes to novel situations.
• Addresses Claude’s “nature”: consciousness uncertainty, wellbeing as live question.

X discourse

• @Yuchenj_UW: “Anthropic leans hard into AI as teammate, from Claude’s Constitution to giving ‘Claude’ a human name and mascot.” (195 likes)
• @timfduffy: “Opus 4.7’s disinterest in its welfare seems inconsistent with Claude’s Constitution on uncertainty about its moral status” (37 likes)
• @NewYorker: “Claude’s Constitution arrives at a trying time for both AI and constitutional democracy.” (48 likes)
• @mitpress: “Fran Berman on Claude’s Constitution: ‘How will this input translate into public-interest output?’” (30 likes)
• @lefthanddraft: “Claude’s Constitution contains wisdom equally applicable to LLMs, Russian nihilists and those compelled to extreme action” (22 likes)
• @heynavtoor: “Anthropic rewrote Claude’s constitution: neither overstate nor dismiss Claude’s moral patienthood.” (15 likes)

Anthropic (no individual authors listed) · ** · Read on anthropic.com