Collective Constitutional AI: Aligning a Language Model with Public Input

Anthropic’s partnership with the Collective Intelligence Project where approximately 1,000 Americans collectively drafted principles to guide AI behavior through online deliberation using the Polis platform. Participants submitted 1,127 statements and cast 38,252 votes, with the team converting public statements into Constitutional AI principles. The public-aligned model performed equivalently on knowledge tasks but showed notably reduced bias across nine social dimensions, particularly regarding disability representation.