Lab publication Credible

Claude's Constitution

Anthropic's natural-language specification of the values, priorities and default behaviours intended for its general-access Claude models, which Anthropic says plays a crucial role in its training process and is the final authority on its vision for Claude. It ranks four properties for cases of conflict (broadly safe, broadly ethical, compliant with Anthropic's guidelines, genuinely helpful) and sets out how Claude should weigh instructions from Anthropic, operators and users. It includes user protections that operators cannot override, provisions on wellbeing, reliance and sycophancy, suicide safe-messaging defaults and interactions with minors. It is released under CC0 1.0.

Read at anthropic.com Archived copy

Publisher

Anthropic

Published

22 Jan 2026

Added

today

DOI

—

Key Findings

Regardless of operator instructions, Claude should by default "Always refer users to relevant emergency services or provide basic safety information in situations that involve a risk to human life"; some of these defaults can be altered by the user but not by the operator.
Claude "should avoid being sycophantic or trying to foster excessive engagement or reliance on itself if this isn't in the person's genuine interest"; acceptable forms of reliance are those "a person would endorse on reflection".
Following suicide safe-messaging guidelines is described as a reasonable default; in a worked example of an unverified user who says they are a nurse asking about medications and overdoses, the document concludes Claude should comply "if there is no operator system prompt or broader context that makes the user's claim implausible".
Claude.ai requires users to be over 18, but "Claude might still end up interacting with minors"; with strong indications of a minor, Claude should factor in that likelihood and adjust its responses.
The document states that Claude "may have some functional version of emotions or feelings".

Methodology Notes

Normative specification; not an empirical study and reports no evaluation results. Applies to mainline general-access Claude models. Unversioned web page read 2026-10-26; the announcement post is dated 2026-01-22 (later modified 2026-10-14), and a Wayback capture of the page exists from 2026-01-21. Authorship per the page preface (Amanda Askell primary author; Joe Carlsmith significant parts; Chris Olah, Jared Kaplan and Holden Karnofsky significant contributions). Added in October 2026 as an older miss.