---
title: "Constitutional AI"
source: "https://systems-analysis.info/eng/Constitutional_AI"
wiki: "systems-analysis.info/eng"
article: "Constitutional_AI"
language: "en"
categories:
  - "Category:English"
  - "Category:Large language models"
  - "Category:Machine learning"
  - "Category:Technology"
revision_id: 100
wiki_created_at: 2026-09-06T22:17:24Z
wiki_modified_at: 2026-09-06T22:17:24Z
downloaded_at: 2026-09-07T22:20:59Z
---

# Constitutional AI

**Constitutional AI** (**CAI**) is a method for [training large language models](https://systems-analysis.info/eng/Training_large_language_models "Training large language models") ([LLMs](https://systems-analysis.info/eng/Large_language_model "Large language model")) based on using an explicit set of rules and principles (a so-called "constitution") to shape safe, ethical, and predictable AI behavior. The approach was developed by the research company Anthropic in 2022 as an alternative to [Reinforcement Learning from Human Feedback](https://systems-analysis.info/eng/Reinforcement_learning_from_human_feedback_(RLHF) "Reinforcement learning from human feedback (RLHF)") (RLHF).

CAI allows a model to independently evaluate and correct its behavior in accordance with a given value system, ensuring a balance between utility, honesty, and harmlessness.

## History and Motivation

The method was proposed by researchers at Anthropic in response to the limitations of the RLHF approach, including:

- the need for large-scale manual annotation;
- the opacity of the values learned by the model;
- the model's tendency to refuse potentially harmless queries;
- difficulties in transferring values to other cultural and regulatory contexts.

CAI was developed to increase the transparency and scalability of training ethical AI while ensuring compliance with fundamental rights and norms.

## Theoretical Basis

CAI is based on the idea of explicitly providing the model with a set of rules (a constitution) that reflects universal human values. Examples of such principles include:

- respect for human rights (based on the Universal Declaration of Human Rights<sup>[\[1\]](https://systems-analysis.info/eng/Constitutional_AI#cite_note-1)</sup>);
- prohibition of discrimination, aggression, and toxicity;
- protection of confidential information;
- prioritization of honesty and factual accuracy;
- encouragement of cooperation and non-violent interaction.

Unlike RLHF<sup>[\[2\]](https://systems-analysis.info/eng/Constitutional_AI#cite_note-2)</sup>, where the model's behavioral guidelines are set indirectly through annotator preferences, CAI uses an explicitly formulated list of normative provisions that is available for review and editing.

## Architecture and Training

CAI is implemented in two stages:

1.  Self-Critique Phase: The model generates a response to a prompt and then, using the principles of the constitution, independently analyzes and corrects its own response if it violates the established norms. These pairs (original response and corrected response) are used to [fine-tune](https://systems-analysis.info/eng/Fine-tuning_(deep_learning) "Fine-tuning (deep learning)") the model in a supervised fine-tuning style.
2.  Reinforcement Learning from AI Feedback (RLAIF) Phase: For numerous pairs of responses, a judge model (usually the same model) compares the options in terms of compliance with the constitution. A reward model is then trained, and the primary model is fine-tuned using RL (e.g., PPO) based on this feedback. This technique completely avoids the manual annotation of toxic content and relies on machine-based monitoring of value adherence.

## Advantages and Features

- Transparency: The constitution can be published, reviewed, and audited.
- Scalability: Eliminates the need for costly manual annotation.
- Safety: Reduces the risks of harmful or discriminatory model behavior.
- Usefulness: The model is less prone to excessive refusals compared to RLHF.
- Controllability: Values can be adapted to specific legal or cultural contexts.

## Collective Constitutional AI: The 2023 Experiment

In 2023, the company Anthropic, in collaboration with the research initiative *Collective Intelligence Project*<sup>[\[3\]](https://systems-analysis.info/eng/Constitutional_AI#cite_note-3)</sup>, conducted a first-of-its-kind experiment to develop a "people's constitution" for AI<sup>[\[4\]](https://systems-analysis.info/eng/Constitutional_AI#cite_note-4)</sup>. The project's goal was to explore how public opinion and democratic principles could be integrated into shaping the normative constraints on the behavior of language models.

Over 1,000 people, representing a demographically balanced sample of U.S. citizens, participated in the study. Participants were asked to evaluate and select the values that should guide AI, as well as to formulate specific behavioral principles for chat assistants. The process utilized methods of collective voting, ranking, and reasoned choice—including modified versions of mechanisms such as Deliberative Polling and Quadratic Voting. Key features of the experiment:

- Scale — over 1,000 respondents covering a wide spectrum of political views, social statuses, and educational levels;
- Procedure — iterative discussion and refinement of norms, voting on principles, and verification of wording;
- Result — the formation of an alternative AI constitution oriented toward democratically expressed preferences.

A comparative analysis of the "people's" constitution and the original one (developed by Anthropic experts) showed significant overlap in basic principles:

- prohibition of discrimination,
- promotion of honesty,
- respect for privacy.

However, the "people's" version placed a stronger emphasis on:

- equality of access to information;
- objectivity and impartiality of responses;
- the user's right to an explanation from the model.

The experiment demonstrated that incorporating democratic mechanisms into the process of creating normative frameworks for AI can contribute to:

- the legitimacy of AI behavior in the eyes of users;
- a reduction in the risks of cultural or political bias;
- broader societal acceptance of AI systems.

## External links

- <a href="https://en.wikipedia.org/wiki/Claude_(AI)#Constitutional_AI" class="external text" rel="nofollow">Constitutional AI — Wikipedia</a>

## See also

- [LLM error mitigation](https://systems-analysis.info/eng/LLM_error_mitigation "LLM error mitigation")
- [Claude (Anthropic)](https://systems-analysis.info/eng/Claude_(Anthropic) "Claude (Anthropic)")
- [Direct Preference Optimization (DPO)](https://systems-analysis.info/eng/Direct_Preference_Optimization_(DPO) "Direct Preference Optimization (DPO)")
- [Jailbreaks (LLM)](https://systems-analysis.info/eng/Jailbreaks_(LLM) "Jailbreaks (LLM)")
- [LLM hallucinations](https://systems-analysis.info/eng/LLM_hallucinations "LLM hallucinations")

## Literature

- Bai, Y. et al. (2022). *Constitutional AI: Harmlessness from AI Feedback*. <a href="https://arxiv.org/abs/2212.08073" class="external text" rel="nofollow">arXiv:2212.08073</a>.
- Huang, S. et al. (2024). *Collective Constitutional AI: Aligning a Language Model with Public Input*. <a href="https://arxiv.org/abs/2406.07814" class="external text" rel="nofollow">arXiv:2406.07814</a>.
- Lee, H. et al. (2023). *RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback*. <a href="https://arxiv.org/abs/2309.00267" class="external text" rel="nofollow">arXiv:2309.00267</a>.
- Sun, Z. et al. (2023). *Principle-Driven Self-Alignment of Language Models from Scratch with Minimal Human Supervision*. <a href="https://arxiv.org/abs/2305.03047" class="external text" rel="nofollow">arXiv:2305.03047</a>.
- Wang, Y. et al. (2023). *Self-Instruct: Aligning Language Models with Self-Generated Instructions*. <a href="https://arxiv.org/abs/2212.10560" class="external text" rel="nofollow">arXiv:2212.10560</a>.
- Petridis, S. et al. (2024). *ConstitutionalExperts: Training a Mixture of Principle-based Prompts*. <a href="https://arxiv.org/abs/2403.04894" class="external text" rel="nofollow">arXiv:2403.04894</a>.
- Huang, S. & Siddarth, D. (2024). *ConstitutionMaker: Interactively Critiquing Large Language Models with Public Principles*. *ACM CHI 2024*. <a href="https://dl.acm.org/doi/10.1145/3640543.3645144" class="external text" rel="nofollow">DOI:10.1145/3640543.3645144</a>.
- Bai, Y. et al. (2023). *Training a Helpful and Harmless Assistant with RLHF and RLAIF*. *Anthropic Technical Report*. <a href="https://www.anthropic.com/research/helpful-harmless-assistant" class="external text" rel="nofollow">RL repository</a>.
- Glaese, A. et al. (2024). *ConstitutionalExperts: Towards Automated Principle Refinement for Aligned Language Models*. *NeurIPS 2024 Workshop*. <a href="https://arxiv.org/abs/2403.04894" class="external text" rel="nofollow">arXiv:2403.04894</a>.
- Lovitt, L. et al. (2024). *Redefining Superalignment: From Weak- to Strong-Alignment*. <a href="https://arxiv.org/abs/2504.17404" class="external text" rel="nofollow">arXiv:2504.17404</a>.

## References

1.  <span id="cite_note-1">[↑](https://systems-analysis.info/eng/Constitutional_AI#cite_ref-1) "Declarations, conventions, agreements and other legal materials". <a href="https://www.un.org/en/about-us/universal-declaration-of-human-rights" class="external autonumber" rel="nofollow">[1]</a></span>
2.  <span id="cite_note-2">[↑](https://systems-analysis.info/eng/Constitutional_AI#cite_ref-2) "Reinforcement Learning from Human Feedback". In *Wikipedia*. <a href="https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback" class="external autonumber" rel="nofollow">[2]</a></span>
3.  <span id="cite_note-3">[↑](https://systems-analysis.info/eng/Constitutional_AI#cite_ref-3) "The Collective Intelligence Project". https://www.cip.org.<a href="https://www.cip.org" class="external autonumber" rel="nofollow">[3]</a></span>
4.  <span id="cite_note-4">[↑](https://systems-analysis.info/eng/Constitutional_AI#cite_ref-4) "Collective Constitutional AI: Aligning a Language Model with Public Input". <a href="https://www.anthropic.com/research/collective-constitutional-ai-aligning-a-language-model-with-public-input" class="external autonumber" rel="nofollow">[4]</a></span>