Constitutional AI is not a constitution (hadleylab.org)

Dario Amodei, co-founder of Anthropic, recently faced significant challenges after the Pentagon designated his company a supply chain risk due to its refusal to engage in military contracts concerning autonomous weapons. A federal judge intervened, allowing Anthropic to maintain its integrity and focus on its distinctive AI approach, known as "Constitutional AI." This training method empowers AI systems to evaluate their outputs against a set of principles, but crucially, it does not incorporate formal governance mechanisms, raising concerns about accountability and enforceability. The system aims to make AI less harmful by using AI-generated feedback instead of human input, but critics highlight that it lacks a structure for verification and real-time compliance. The implications of this distinction are significant for the AI/ML community. While Constitutional AI has shown promise in creating safer models, it contrasts sharply with structured governance systems like CANONIC, which introduces formal contracts that are auditable and enforceable. CANONIC allows for a more rigorous governance framework, ensuring that model behavior is aligned with institutional commitments and providing a clear audit trail. This divergence illustrates the ongoing challenge of establishing robust frameworks for AI governance, emphasizing the need for both effective training techniques and enforceable governance contracts to build trustworthy AI systems.