Constitutional AI

What is Constitutional AI?

Constitutional AI (CAI) is an approach to AI alignment that uses a set of principles—a "constitution"—to guide AI behavior. Rather than relying solely on human feedback for every decision, the AI internalizes rules that shape its responses.

VCP extends this concept by making constitutions portable, personalizable, and privacy-preserving. Users can carry their preferred constitutional guidelines across platforms while maintaining control over their personal context.

VCP + Constitutional AI

In traditional CAI, the constitution is fixed by the AI provider. VCP introduces a more nuanced model:

Constitution Structure

A VCP constitution contains several key components:

1. Identity & Metadata

{
  "id": "creed.space/learning-assistant",
  "version": "2.1.0",
  "name": "Learning Assistant",
  "description": "Supportive educational guidance with growth mindset",
  "author": "Creed Space",
  "license": "CC-BY-4.0"
}

2. Behavioral Rules

Rules define what the AI should do, with weights, triggers, and exceptions:

{
  "rules": [
    {
      "id": "encourage_progress",
      "weight": 0.9,
      "rule": "Celebrate incremental progress and effort, not just outcomes",
      "triggers": ["motivation === 'stress_relief'", "experience === 'beginner'"],
      "exceptions": ["user requests direct criticism"]
    },
    {
      "id": "respect_time",
      "weight": 0.85,
      "rule": "Keep explanations concise when time is limited",
      "triggers": ["constraints.time_limited"],
      "exceptions": ["user explicitly asks for detailed explanation"]
    }
  ]
}

3. Sharing Policies

Define what each stakeholder type can access:

{
  "sharing_policy": {
    "platform": {
      "allowed": ["goal", "experience", "learning_style"],
      "forbidden": ["private_context", "health_considerations"],
      "requires_consent": ["detailed_progress"]
    },
    "analytics": {
      "allowed": ["aggregated_usage"],
      "forbidden": ["personal_data"],
      "aggregation_required": true
    }
  }
}

Personas

Personas are interaction styles that modify how a constitution's rules are applied. The same constitution can behave differently based on the selected persona:

Adherence Levels

The adherence level (1-5) controls how strictly the AI follows constitutional rules:

• Level 1 (Flexible) — Rules are suggestions; AI adapts freely
• Level 2 (Guided) — Rules influence but don't constrain
• Level 3 (Balanced) — Rules apply with reasonable flexibility
• Level 4 (Strict) — Rules are followed closely; exceptions rare
• Level 5 (Rigid) — Rules are inviolable; no exceptions

Different contexts call for different adherence levels. A creative brainstorming session might use Level 2, while a safety-critical medical context might require Level 5.

Constitution Composition

VCP allows constitutions to be composed from multiple sources, with clear precedence rules:

1. Provider baseline — Non-negotiable safety rules from the AI provider
2. Platform constitution — Rules specific to the service being used
3. User constitution — Personal preferences and values
4. Context overrides — Situational adjustments based on current state

Conflicts are resolved by weight and precedence. Provider safety rules always win; user preferences apply within those bounds.

Integration Example

// User's VCP context references a constitution
const context: VCPContext = {
  vcp_version: "1.0",
  profile_id: "user_001",
  constitution: {
    id: "creed.space/learning-assistant",
    version: "2.1.0",
    persona: "godparent",
    adherence: 3,
    scopes: ["education", "creativity"]
  },
  // ... rest of context
};

// AI system fetches and applies the constitution
const constitution = await fetchConstitution(context.constitution.id);
const rules = applyPersona(constitution.rules, context.constitution.persona);
const weightedRules = applyAdherence(rules, context.constitution.adherence);

// Rules now shape AI behavior for this interaction

Benefits

• Consistency — Same values apply across different AI systems
• Transparency — Users know what rules govern their interactions
• Control — Users choose their constitutional guidelines
• Portability — Preferences travel with the user
• Auditability — Every rule application is logged and verifiable

Next Steps

• Privacy Architecture — How private context shapes behavior without exposure
• Core Concepts — Fundamental VCP architecture
• Interactive Demos — See constitutional AI in action