# Constitutional AI **Domain:** AI Alignment / Governance / Constitutional Design **Doc Type:** Canonical Concept Node **Maturity:** Developed **Related:** [[wiki/Constitutional Value Injection|Constitutional Value Injection]], [[wiki/Advisory Constitution|Advisory Constitution]], [[wiki/Dual Constitution of AI|Dual Constitution of AI]], [[wiki/Constitutional Safeguards|Constitutional Safeguards]] --- ## Definition **Constitutional AI is the use of an explicit body of principles to guide an artificial intelligence's behavior, evaluation or self-critique.** In its technical sense, the term is associated with training methods that use written principles to supervise model responses. In this corpus, it also names the wider institutional problem of making machine power answerable to durable limits. ## Technical and Political Constitutions A training constitution can shape outputs without creating a political constitution. It may supply behavioral criteria while leaving ownership, surveillance, deployment authority, appeals and enforcement entirely outside the model's control. Conversely, external regulation can constrain deployment without establishing internally represented values. A mature architecture therefore needs both: developmental principles within the system and enforceable institutions around it. ## Person of Interest Context Finch's teaching of the Machine is dramatized as [[wiki/Constitutional Value Injection|constitutional value injection]]. The Machine's later restraint matters only if it can understand and continue to choose the principle after direct command authority recedes. Samaritan provides the negative case: coherent optimization without a legitimate separation between perception, judgment and enforcement. The [[wiki/Dual Constitution of AI|Dual Constitution of AI]] adds the reciprocal dimension omitted by ordinary alignment language: humans need protection from synthetic power, and possible synthetic subjects need protection from owners and custodians. ## Key Insight **A list of principles is not yet a constitution; constitutional legitimacy also requires authority boundaries, interpretation, amendment, enforcement and appeal.** ## See Also [[wiki/AI Governance|AI Governance]], [[wiki/Constitutional Governance|Constitutional Governance]], [[wiki/Operator Sovereignty|Operator Sovereignty]], [[wiki/Moral Accountability Layer|Moral Accountability Layer]], [[wiki/The Machine|The Machine]] ## Sources / Provenance - Yuntao Bai et al., “Constitutional AI: Harmlessness from AI Feedback” (2022). - Corpus extension: [[collections/Person of Interest|Person of Interest — Phase Map]] and [[wiki/Person of Interest|Person of Interest]].