Collective Constitutional AI: Aligning a Language Model with Public Input

This report describes Anthropic and the Collective Intelligence Project’s experiment engaging ~1,000 U.S. participants via the Polis platform to draft a public constitution for AI, how the public-sourced principles were translated into Constitutional AI training, the training and evaluation of two Claude Instant-sized models (Public vs Standard), key findings (comparable helpfulness/harmlessness, modestly reduced bias in the Public model), and lessons learned about participant selection, moderation, principle mapping, prompt database alignment, and evaluation challenges.