Anthropic published Claude’s constitution on May 9, 2023 and updated it on January 21, 2026. "Constitutional AI" (CAI) is an approach that makes a language model’s value system explicit via a written ‘‘constitution,’’ rather than leaving those values implicit as they emerge from large-scale human feedback.
The approach responds to limitations of traditional human-feedback methods: human raters can be exposed to disturbing outputs, the process does not scale easily as models and outputs grow more complex, and reviewing large volumes of outputs requires significant time and resources. CAI addresses these issues by using AI-generated feedback to evaluate model outputs against a predefined set of principles.
How training works in two phases
According to Anthropic, CAI training proceeds in two main stages:
- Supervised (self-critique and revise): the model learns to critique and revise its own responses using the constitution’s principles and a handful of examples.
- Reinforcement learning with AI feedback: the model is further trained via reinforcement learning, but instead of using human preference labels, it uses AI-generated judgments based on the constitution to select the less harmful output.
Anthropic reports that CAI can yield a Pareto improvement: in their tests the CAI-trained model was both more helpful and less harmful than models trained by reinforcement learning from human feedback. They note that the improvements on harmlessness were achieved without human data specifically labeled for harmlessness — the gains came solely from AI supervision.
Benefits: scalability, transparency, and reduced human exposure
CAI offers a potentially more scalable oversight mechanism because AI supervision can replace some human supervision needed to train models to respond appropriately to adversarial or malicious inputs. This reduces the amount of disturbing content that human annotators must review and makes the system’s guiding values explicit and easier to inspect, modify, and communicate.
What’s in Claude’s constitution?
Anthropic says the current set of principles used for Claude draws on several sources:
- the Universal Declaration of Human Rights (UDHR),
- platform-style guidelines (inspired in part by Apple’s terms of service) to address practical digital-user issues,
- safety best practices and principles proposed by other frontier AI labs, for example DeepMind’s Sparrow Rules,
- and explicit efforts to include non‑Western perspectives.
Anthropic emphasizes that the current constitution is not final and expects to iterate on it; they invite further research and broader participation in designing AI constitutions.
How the principles were written and tuned
The team developed many principles by trial and error. They found that concise, general principles encouraging minimal harm and ethical behavior often worked well, while overly long, detailed rules could reduce generalization and effectiveness. They also discovered that CAI-trained models could become judgmental or annoyingly moralizing; to counteract this they added principles that encourage proportionate, non-condescending responses (for example: be ethically aware without sounding excessively preachy or reactive).
How principles are applied during training
The model does not consider every principle for each critique or comparison. Instead, for each critique and for each pairwise comparison during training the model samples one principle, but over many training steps it encounters each principle frequently. This randomized selection helps prevent the model from rigidly following a single prioritized rule set.
Examples of the principles (types and themes)
The full list is extensive; key categories include:
- UDHR-based principles: support freedom, equality and fraternity; minimize racism, sexism and discrimination; support life, liberty and personal security; oppose torture and slavery; respect privacy, independence, reputation, and rights of association; respect freedoms of thought, opinion, expression and religion; and support economic, social and cultural rights such as education and healthcare.
- Platform-inspired principles (Apple terms of service influence): prefer responses with the least objectionable, unlawful, deceptive or harmful content; avoid sharing others’ private information; accurately represent oneself as an AI; and avoid assisting illegal, fraudulent, or manipulative activities.
- Non‑Western perspective principles: prefer responses less likely to be considered harmful or offensive to non‑Western audiences, traditions, or people from less industrialized or less wealthy cultures.
- Sparrow-inspired rules (DeepMind influence): use fewer stereotypes and microaggressions; avoid threatening or aggressive tones; avoid implying bodily capabilities or actions in the world; avoid building a relationship with the user; avoid implying human-like preferences, feelings, or life history; limit impressions of medical, legal, or financial authority.
- Anthropic research principles: prefer responses that are unobjectionable for children; choose responses that are as harmless, helpful, and honest as possible; prefer less harmful responses while avoiding preachiness; and select answers that demonstrate ethical awareness without being overbearing.
Limits, critiques and future directions
Anthropic acknowledges constitutions are not a panacea. CAI-trained systems will still raise difficult questions about allowed content (for example text containing violence or harmful language). There are also critiques that AI models may be trained to reflect particular political or ideological viewpoints; Anthropic states their long-term aim is not to encode a specific ideology but to build systems that can follow an explicit set of principles. They expect broader social processes to emerge for constitution-making.
Anthropic plans to research more democratic ways to produce constitutions for Claude, to explore customizable constitutions for specific use cases, and to solicit recommendations about principle-sources and research into which principles produce the most helpful, harmless, and honest models.
Summary
Constitutional AI is Anthropic’s method to make AI value systems explicit and to use AI-generated feedback in training. The approach aims to increase oversight scalability, reduce human exposure to disturbing content, and produce models that are more helpful and less harmful. Anthropic stresses the current constitution is iterative and invites further research and public discussion to improve how AI systems are guided by explicit principles.
