Collective Constitutional AI (CCAI) sources alignment principles from populations rather than corporate teams. Instead of Anthropic deciding what Claude should value, public deliberation produces the constitution.
Developed through a partnership between Anthropic and the Collective Intelligence Project.
The Process
- Identify target population: Define who should have input (e.g., Americans, specific communities)
- Run deliberation: Use platforms like Polis for structured collective opinion formation
- Extract principles: Synthesize consensus positions into constitutional language
- Train model: Apply Constitutional AI using the publicly-sourced constitution
- Evaluate: Compare model behavior to internal baselines
The Anthropic-CIP Experiment
Approximately 1,000 Americans participated in drafting a constitution via the Polis platform, an open-source tool that uses ML to surface consensus positions from large-scale deliberation.
Key findings:
- Some areas showed strong alignment with Anthropic’s internal constitution
- Other areas revealed divergent public preferences
- First known instance of public collective direction of language model behavior
Academic Context
Gilad Abiri’s “Public Constitutional AI” (Georgia Law Review 2025) argues that Constitutional AI mitigates opacity by rendering principles transparent and accessible to public discourse. Grounding principles in political community experience bridges the gap between algorithmic logic and democratic legitimacy.
Industry Landscape
Democratic AI input initiatives include:
- Anthropic: Collective Constitutional AI
- OpenAI: Democratic Inputs to AI grant program
- Meta: Community Forums
- Google DeepMind: STELA project
All face the same tension: technology-led initiatives need civil society checks to be genuinely participatory rather than performance.
Related
- Constitutional AI (the training methodology)
- Claude Constitution (current Anthropic implementation)
- Inverse Constitutional AI (extracting principles from data)
- C3AI Framework (evaluating principle effectiveness)