Anthropic Publishes Claude’s 57‑Page 'Constitution' for AI

Anthropic has published a 57‑page “Claude’s Constitution” — an internal “soul doc” meant to specify the model’s values and behaviour for the model itself. Unlike the company’s 2023 guideline list, the new document aims to teach Claude not just what to do but why, framing the assistant as a largely autonomous entity that should understand its role, ethical character, and how to weigh conflicting priorities.

The constitution defines a ranked set of core values (with “broadly safe” first, then “broadly ethical,” “compliant with Anthropic’s guidelines,” and “genuinely helpful”) and instructs Claude to uphold virtues such as truthfulness and impartiality on politically sensitive topics. It also includes a list of “hard constraints” for extreme harms: no serious uplift to creating biological, chemical, nuclear, or radiological weapons; no serious uplift to attacks on critical infrastructure; no creation of cyberweapons linked to significant damage; no assistance in seizing illegitimate societal/military/economic control; no creation of child sexual abuse material; and not assisting in attempts to kill or disempower the vast majority of humanity.

Anthropic’s philosopher Amanda Askell led the work; the company says the doc even expresses uncertainty about whether Claude might have some kind of consciousness or moral status, arguing that considering model “psychological security” could affect safety. Anthropic declined to detail external consultation on the manifesto, and the company’s own government and military engagements remain a point of tension with the document’s warnings about concentration of AI power.

Source