{
  "node_id": "constitutional-ai-align",
  "title": "Constitutional AI Algorithm",
  "domain": "AI Governance & Law",
  "version": "1.2.0",
  "last_updated": "2026-07-03",
  "bluf": "Constitutional AI (CAI) is a voluntary alignment research methodology developed by Anthropic (Bai et al., 2022, arXiv:2212.08073), not a law or binding standard, that trains AI systems to be helpful, harmless, and honest using a set of explicit behavioral principles (the 'Constitution') rather than relying exclusively on human feedback labeling of individual outputs. The method operates in two phases: a Supervised Learning from Constitutional AI (SL-CAI) phase where the model critiques and revises its own harmful outputs using principles as guidance, and a Reinforcement Learning from AI Feedback (RL-CAI) phase where an AI-generated preference dataset replaces or supplements human preference labels. CAI has been shown to reduce the need for human labeling of harmful content while producing models that are less harmful and more transparent about their reasoning. As a voluntary research framework, CAI carries no regulatory force of its own; organizations adopting it commonly map the approach to AI governance requirements including EU AI Act Article 9 risk management and NIST AI RMF GOVERN function requirements for systematic safety assurance.",
  "paywall": {
    "status": "LOCKED",
    "unlock_cost_usd": "0.01",
    "skyfire_id": "41779894-ece2-4163-9761-b3b1b76e19b0"
  },
  "crosswalks": {
    "_available_keys": [
      "nist_framework",
      "iso_standard",
      "ai_overlay_2026",
      "industry_mapping"
    ],
    "_note": "Full crosswalk values included in vault response"
  },
  "dependencies": [
    "nist-ai-rmf-1-0",
    "oecd-ai-principles",
    "unesco-ethics-ai",
    "iso-42001-risk-assess"
  ],
  "primary_citations_count": 7
}