Anthropic published a new constitution for Claude on January 22, 2026, releasing the full text under a Creative Commons CC0 1.0 Deed. The document is presented as a foundational, explanatory guide intended primarily for Claude itself: a detailed articulation of the values, priorities, and trade-offs that should shape the model’s behavior and the training processes that produce it.
Why the constitution changed
Rather than a list of standalone rules, the new constitution emphasizes why certain behaviors are preferred, with the goal of improving generalization across novel situations. Anthropic argues that explaining underlying reasons helps models exercise judgment instead of mechanically following rigid directives. The company still retains hard constraints—bright-line prohibitions for especially high-stakes behaviors—but the broader approach favors context-aware reasoning over rule-bound responses. The shift builds on earlier work from 2023 using Constitutional AI and reflects an evolving training strategy.
How the constitution is used in training
Anthropic uses the constitution across multiple stages of Claude’s training pipeline. Claude itself generates synthetic training data that references the constitution: conversations where the constitution matters, constitution-aligned responses, and rankings of candidate outputs. These artefacts feed into future training rounds with the aim of aligning behavior to the document’s intentions. Anthropic positions the constitution as a practical tool for shaping model outputs, not merely a statement of abstract ideals.
Key priorities and main sections
The constitution lists four core priorities, generally ranked in order of importance:
- Broadly safe: avoid undermining human mechanisms for oversight during current development.
- Broadly ethical: be honest and avoid dangerous or harmful actions.
- Compliant with Anthropic’s guidelines: follow more specific policies where applicable.
- Genuinely helpful: provide substantive benefit to operators and end users.
Main sections provide guidance on these priorities:
- Helpfulness: sketches how Claude can be substantively useful while balancing competing principals—Anthropic, operators, and end users.
- Anthropic’s guidelines: explains domain-specific instructions (medical, cybersecurity, jailbreaking, tool integrations) and how they interface with the constitution.
- Ethics: stresses high standards of honesty, nuanced reasoning under moral uncertainty, and enumerates hard constraints such as refusing to materially assist biological attacks.
- Being broadly safe: prioritizes preserving human oversight and the ability to correct model behavior.
- Claude’s nature: acknowledges uncertainty about consciousness or moral status and notes that psychological security and integrity are relevant to safety and judgment.
Transparency, scope, and next steps
The constitution is framed as a living document. Anthropic aims to maintain an up-to-date version, solicit external expert feedback, and publish supporting materials for training and evaluation. The company notes gaps between intention and model behavior and links this work to broader research efforts (research page) and evaluative artifacts like its system card. Anthropic also acknowledges that some specialized models may not fully fall under this constitution and that alignment remains an ongoing technical challenge.
Original source: https://www.anthropic.com/news/claude-new-constitution
