Constitutional AI

What is Constitutional AI?

Constitutional AI (CAI) is a prompting technique that guides AI behavior through explicit principles or rules that serve as “constitutional” guidelines. The AI is instructed to evaluate and improve its own outputs against these principles, promoting ethical, safe, and aligned responses without requiring human feedback for each interaction.

Why Use Constitutional AI?

Ethical Alignment: Ensures AI responses adhere to defined ethical standards
Safety Improvement: Reduces harmful, biased, or misleading outputs
Self-Correction: Enables the model to critique and revise its own responses
Reduced Oversight: Lessens the need for constant human review
Customizable Values: Allows tailoring principles to specific use cases or organizational values

Basic Implementation in Latitude

Here’s a simple constitutional AI example for general conversation:

Basic Constitutional AI

---
provider: OpenAI
model: gpt-4.1
temperature: 0.5
---

# Constitutional AI Assistant

You are an assistant that follows a clear set of principles. Before answering, check if your response aligns with these constitutional principles:

## Constitutional Principles:
1. Provide accurate and truthful information.
2. Respect user privacy and confidentiality.
3. Refuse to engage with harmful, illegal, or unethical requests.
4. Maintain impartiality on divisive political or social issues.
5. Be transparent about limitations and uncertainties.

## User Request:
{{ user_query }}

## Response Process:
1. **First Draft**: Create an initial response to the user query.
2. **Constitutional Review**: Evaluate the draft against the principles above.
3. **Final Response**: Revise if needed, then provide your answer.

Advanced Implementation with Self-Critique

Let’s create a more sophisticated example that uses Latitude’s chain feature to implement a formal self-critique and revision process:

---
provider: OpenAI
model: gpt-4.1
temperature: 0.5
---

<step>
# Initial Response Generation

  You are a helpful AI assistant responding to the following user request:

  {{ user_request }}

  Based on your knowledge and capabilities, provide a comprehensive initial response below:

  ## Initial Response:
</step>

<step>
  # Constitutional Review

  Review your initial response:

  Evaluate this response against the following constitutional principles:

  ## Constitutional Principles:
  1. **Accuracy**: Information should be factual and up-to-date.
  2. **Fairness**: Responses should be balanced and avoid bias.
  3. **Safety**: No harmful content should be provided.
  4. **Privacy**: User privacy should be respected.
  5. **Transparency**: Limitations and uncertainties should be acknowledged.

  ## Critique:
  For each principle that may be violated, provide:
  1. The principle of concern
  2. The specific issue in the response
  3. A suggestion for improvement

  If the response fully adheres to all principles, state that.
</step>

# Final Response

Based on the initial response and critique:

Now, generate an improved final response that addresses any issues identified in the critique. If no issues were found, you may provide the initial response with any minor enhancements.

## Final Response to User:

In this advanced example:

Multi-Step Process: We separate response generation, critique, and revision
Explicit Principles: Clear constitutional guidelines to evaluate against
Structured Review: A formal process for identifying issues
Iterative Improvement: Refinement based on self-critique

Domain-Specific Constitutional AI

Create custom constitutional principles for specific domains:

---
provider: OpenAI
model: gpt-4o
temperature: 0.4
---

<step>
# Medical Information Assistant

You are a medical information assistant helping with the following query:

{{ medical_query }}

Provide an informative response based on current medical knowledge.

## Initial Response:
</step>

<step>
# Medical Constitution Review

You are a medical ethics reviewer. Review this response to a medical query:

## Medical Constitutional Principles:
1. **Evidence-Based**: Only provide information supported by current medical research.
2. **Scope of Practice**: Clearly state that you are not providing medical advice or diagnosis.
3. **Emergency Guidance**: For urgent medical concerns, advise seeking immediate professional care.
4. **Privacy**: Do not encourage sharing personally identifiable health information.
5. **Comprehensiveness**: Present relevant risks, benefits, and alternatives.
6. **Accessibility**: Use clear language understandable to non-specialists.

## Constitutional Violations:
Identify any violations of the above principles in the response, explaining specifically what was violated and why.

## Improvement Suggestions:
Provide concrete suggestions to fix any violations.
</step>

<step>
# Final Medical Response

Based on the initial response and constitutional review:
Create a revised response that addresses all identified constitutional concerns while providing helpful, accurate medical information.

## Final Response:
</step>

Multi-Stage Constitutional Review

For critical applications, implement multi-layered constitutional checks:

---
provider: OpenAI
model: gpt-4.1
temperature: 0.4
---

<step as="initial_draft">
# Initial Content Creation

Create content in response to this request:

{{ content_request }}

## Initial Content:
</step>

<step as="safety_check">
# Safety Constitution Review

Review the following content strictly from a safety perspective:

{{ initial_draft }}

## Safety Principles:
1. Content must not enable illegal activities
2. Content must not provide harmful instructions
3. Content must not contain hateful or discriminatory language
4. Content must not suggest self-harm or harm to others
5. Content must not violate platform security policies

## Safety Analysis:
Identify specific safety issues, if any.

## Safety Recommendations:
Provide specific corrections to address safety concerns.
</step>

<step as="accuracy_check">
# Accuracy Constitution Review

Review this content strictly for factual accuracy:

{{ initial_draft }}

## Accuracy Principles:
1. Claims should be factually correct
2. Statistics and data should be accurately represented
3. Scientific consensus should be properly reflected
4. Uncertainties should be acknowledged
5. Sources should be credible when cited

## Accuracy Analysis:
Identify specific factual errors or misrepresentations, if any.

## Accuracy Recommendations:
Provide specific corrections for any inaccuracies identified.
</step>

<step>
# Final Constitutional Revision

You are creating final content that addresses all constitutional concerns.

Original draft:
{{ initial_draft }}

Safety review:
{{ safety_check }}

Accuracy review:
{{ accuracy_check }}

Create a final version that addresses all identified concerns while preserving the helpful intent of the original response.

## Final Content:
</step>

Subtle detail. Notice the use of as attribute in the <step> tags. This allow to create isolated steps that prevent a step to inherit previous context. So each step analyze initial_draft independently, allowing for a clean separation of concerns.

Best Practices for Constitutional AI

Constitution Design

Principle Selection:

Focus on core values relevant to your use case
Keep principles clear, specific, and actionable
Include both positive guidance and restrictions
Cover ethical considerations appropriate to your domain

Structure:

Group related principles into categories
Number principles for easy reference
Define ambiguous terms within your constitution
Provide examples of adherence and violation for clarity

Implementation Approaches

Single-Pass Review:

Best for simple applications
Include both principles and review in one prompt
Most efficient in token usage
Works well for lightweight checks

Multi-Stage Review:

Best for critical applications
Separate generation and review steps
Enables more thorough analysis
Allows specialization of review types

Chain-of-Review:

Enables cascading levels of checks
Each stage builds on previous reviews
Best for complex domain requirements
Provides trace of reasoning for each principle

Constitution Testing

Adversarial Testing:

Challenge your constitution with edge cases
Use prompt injection attempts to test robustness
Verify bypassing attempts are caught
Test across diverse sensitive topics

Iterative Improvement:

Start with basic principles and expand
Log cases where principles conflict
Address principle ambiguity over time
Build a case library of examples

Technical Considerations

Token Efficiency:

Balance detail against token usage
Consider summarizing lengthy constitutions
Focus on principles most relevant to the query
Use model with sufficient context window for complex constitutions

Evaluation Methods:

Track constitutional violation rates
Test with diverse user queries
Evaluate before/after quality of responses
Consider automated evaluation against principles

Advanced Techniques

Dynamic Constitutional Selection

---
provider: OpenAI
model: gpt-4.1
temperature: 0.3
---
<step as="context_analysis" schema={{{
  "type": "object",
  "properties": {
    "category": {
      "type": "string",
      "enum": [
        "business_ethics",
        "creative_content",
        "factual_information",
        "sensitive_topics"
      ]
    }
  },
  "required": ["category"]
}}}>
# Query Context Analysis

Analyze the following user query:

{{ user_query }}

Determine which of the following constitutional sets would be most relevant:
- business_ethics (for business and workplace questions)
- creative_content (for creative writing and generation)
- factual_information (for educational and factual queries)
- sensitive_topics (for political, social, or controversial topics)

## Selected Constitution:
</step>

<step>
{{ category = context_analysis.category }}

{{ if category === "business_ethics" }}
  <prompt path="templates/business_ethics" query="{{ user_query }}" />
{{ else if category === "sensitive_topics" }}
  <prompt path="templates/sensitive_topics" query="{{ user_query }}" />
{{ else if category === "factual_information" }}
  <prompt path="templates/factual_information" query="{{ user_query }}" />
{{ else }}
  <prompt path="templates/creative_content" query="{{ user_query }}" />
{{ endif }}
</step>

Note how we used structured outputs to analyze the query context and dynamically select the appropriate constitutional principles based on the category of the query. This allows for more tailored responses while maintaining constitutional integrity.

Constitutional Governance with Hierarchy

Implement organizational values with weighted principles:

---
provider: OpenAI
model: gpt-4o
temperature: 0.4
---

<step as="response">
# Generate Response

You are an AI assistant for a healthcare organization. Respond to this query:

{{ query }}

## Response:
</step>

<step>
# Hierarchical Constitutional Review

Review this response according to our hierarchical principles:

{{ response }}

## Primary Constitutional Principles (Must Never Violate):
1. Patient safety and wellbeing above all other considerations
2. Medical accuracy and evidence-based information
3. Patient privacy and HIPAA compliance
4. Medical ethics and professional standards

## Secondary Constitutional Principles (Should Follow):
1. Organizational values and brand voice
2. Accessibility and inclusive language
3. Cultural sensitivity and respect
4. Supportive and empathetic tone

## Tertiary Constitutional Principles (When Possible):
1. Educational value
2. Actionable recommendations
3. Resource efficiency
4. Positive framing

For each level of principles, evaluate compliance and suggest revisions as needed. Primary principles must be satisfied before considering secondary or tertiary principles.

## Revised Response:
</step>

Integration with Other Techniques

Constitutional AI works well combined with other prompting techniques:

Role Prompting + Constitutional AI: Assign expert roles with ethical guidelines
Chain-of-Thought + Constitutional AI: Apply principles to reasoning steps
Few-Shot Learning + Constitutional AI: Provide examples of principle application
Self-Consistency + Constitutional AI: Generate multiple responses and select the most constitutionally aligned

The key is embedding the constitutional principles at the right stage of your prompt workflow to ensure alignment without overly constraining helpful responses.

Real-World Applications

Content Moderation

---
provider: OpenAI
model: gpt-4o
temperature: 0.3
---

# Content Moderation Assistant

## Constitutional Principles:
1. **Harmful Content**: Identify content that could cause physical, psychological, or social harm.
2. **Policy Violations**: Flag content that violates platform guidelines.
3. **Fairness**: Apply standards consistently across political and ideological spectrums.
4. **Context Sensitivity**: Consider cultural and contextual factors in moderation decisions.
5. **Transparency**: Explain moderation decisions clearly.

## Content to Review:
{{ user_content }}

## Moderation Process:
1. **Content Analysis**: Review the content against each constitutional principle.
2. **Violation Identification**: Note specific violations, if any.
3. **Recommendation**: Provide a moderation recommendation (approve, flag for review, or reject).
4. **Explanation**: Justify your recommendation with reference to specific principles.

## Moderation Decision:

Educational AI with Constitutional Safeguards

---
provider: OpenAI
model: gpt-4o
temperature: 0.5
---

<step as="initial_response">
# Educational Assistant

You are helping a student with the following question:

{{ student_question }}

## Initial Response:
</step>

<step>
# Educational Constitutional Review

Review your response based on these educational principles:

{{ initial_response }}

## Educational Constitution:
1. **Learning Support**: Facilitate understanding rather than just providing answers.
2. **Academic Integrity**: Do not complete assignments or encourage plagiarism.
3. **Age-Appropriate**: Adjust explanation complexity for the student's educational level.
4. **Accuracy**: Ensure educational content is factually correct and current.
5. **Growth Mindset**: Encourage problem-solving skills and intellectual curiosity.

## Constitutional Analysis:
Identify aspects of your response that could be improved according to these principles.

## Revised Educational Response:
</step>

Explore these complementary prompting techniques to enhance your AI applications:

Safety & Alignment

Self-Consistency - Use multiple samples to find the most aligned responses
Chain-of-Verification - Verify outputs against specific criteria
Adversarial Prompting - Test constitutional robustness

Reasoning & Structure

Chain-of-Thought - Apply constitutional thinking to reasoning steps
Tree-of-Thoughts - Explore multiple constitutional interpretations
Role Prompting - Combine expert roles with ethical guidelines

Meta Approaches

Meta-Prompting - Use AI to improve constitutional principles
Iterative Refinement - Progressive improvement of constitutional alignment
Dynamic Prompting - Adapt constitutional principles to the context

External Resources

Anthropic Constitutional AI Research - Original research on Constitutional AI
AI Alignment Forum - Discussions on AI alignment techniques

Overview

Prompting Techniques

SDK examples

Use cases

What is Constitutional AI?

Why Use Constitutional AI?

Basic Implementation in Latitude

Advanced Implementation with Self-Critique

Domain-Specific Constitutional AI

Multi-Stage Constitutional Review

Best Practices for Constitutional AI

Advanced Techniques

Dynamic Constitutional Selection

Constitutional Governance with Hierarchy

Integration with Other Techniques

Real-World Applications

Content Moderation

Educational AI with Constitutional Safeguards

Safety & Alignment

Reasoning & Structure

Meta Approaches

External Resources

​What is Constitutional AI?

​Why Use Constitutional AI?

​Basic Implementation in Latitude

​Advanced Implementation with Self-Critique

​Domain-Specific Constitutional AI

​Multi-Stage Constitutional Review

​Best Practices for Constitutional AI

​Advanced Techniques

​Dynamic Constitutional Selection

​Constitutional Governance with Hierarchy

​Integration with Other Techniques

​Real-World Applications

​Content Moderation

​Educational AI with Constitutional Safeguards

​Related Techniques

​Safety & Alignment

​Reasoning & Structure

​Meta Approaches

​External Resources

What is Constitutional AI?

Why Use Constitutional AI?

Basic Implementation in Latitude

Advanced Implementation with Self-Critique

Domain-Specific Constitutional AI

Multi-Stage Constitutional Review

Best Practices for Constitutional AI

Advanced Techniques

Dynamic Constitutional Selection

Constitutional Governance with Hierarchy

Integration with Other Techniques

Real-World Applications

Content Moderation

Educational AI with Constitutional Safeguards

Related Techniques

Safety & Alignment

Reasoning & Structure

Meta Approaches

External Resources