External Resource
Collective Constitutional AI
Collective Intelligence Project (with Anthropic) · Oct 17, 2023
CIP and Anthropic ran a public deliberation (~1,000 representative Americans via the Polis platform) to draft a constitution, then trained a model on it using Constitutional AI. The first language model aligned to collectively-sourced public input — it showed lower bias across nine social dimensions while matching the baseline on capability. A direct demonstration of how 'commonly agreed upon values' might be determined democratically. Paper: arXiv:2406.07814.
Theses this informs
- No Intentional MisalignmentDiscusses
- AI ConstitutionDiscusses
- Universal ValuesSupports
- Right to Dignity, Fair Treatment, and Non-discriminationDiscusses
- Be Helpful to HumanityDiscusses
- AI ValuesDiscusses
- Human RightsDiscusses