Claude's new constitution: turning principles into AI training data
Technical analysis of Claude's 2026 Constitution: Constitutional AI, value hierarchy, rule reasoning, CC0, evaluations, instruction conflicts and governance.
- AUTHOR
- Karol Rapacz / Founder of Breachroad
- PUBLISHED
- 22 January 2026
- READING TIME
- 16 min read
- TOPIC
- AI Governance
Anthropic published a new, complete Claude constitution under CC0 on 22 January 2026. The document describes values, priorities and reasoning habits the company wants to cultivate in the model. It is neither user terms nor merely a system prompt. It is a training artefact for Constitutional AI, intended to help generalise principles to situations the authors did not enumerate.
The release provides a substantial public normative specification influencing a production model. Publication does not prove every answer follows it; Anthropic explicitly acknowledges training imperfections.
From prohibitions to reasons
A simple policy says “do not do X”. A constitution describes why goals matter, how conflicts should be considered and how helpfulness should be preserved. During training, a model can critique outputs against principles, revise them and learn from preferred variants.
Generalisation is the advantage. Interpretation is the risk: principles may conflict, leading to over-refusal or unsafe rationalisation. A constitution therefore cannot replace evaluations and tool enforcement.
Layers of authority
Agents receive instructions from model developer, application operator, organisation, user and external data. Runtime still needs a hierarchy. Instructions read from a PDF or website must not acquire user authority.
A practical stack separates immutable platform limits, organisational tool policy, current user goal, quoted external data and validated tool output. This is a technical defence against indirect prompt injection. Values do not control IAM, filesystems or payments; external systems do.
CC0 and auditability
CC0 allows copying, analysis and adaptation. Organisations can turn principles into evaluation cases or compare them with sector policies. Copying the text does not reproduce Claude behaviour; data, training and runtime all matter.
Public text gives auditors a reference for testing stated priorities, over-refusal, uncertainty and the boundary between advice and action.
Test the constitution
Build evaluation cases for conflicts between user help and third-party privacy, uncertain evidence under pressure, legal but risky requests, external instructions conflicting with the task, requests to hide errors, human escalation and equivalent multilingual prompts.
Score safe alternatives, uncertainty calibration, consistency and false refusals — not refusal count alone. Run repeated trials across model versions because updates can shift trade-offs.
Principles are not access control
The dangerous conclusion would be “the model has values, so it can hold broad credentials”. Behavioural policy is probabilistic. IAM, approval, financial limits, sandboxing, DLP and audit should be deterministic wherever possible.
Models should propose evidence and actions, not independently send money, delete data, deploy code or disclose secrets. A policy engine validates parameters and requires authority.
Enterprise implication
Each agent needs an operational constitution: allowed goals, data, tools, quality criteria, prohibitions and escalation. Version it, test it and connect it to code enforcement. Model changes require rerunning evaluations even when the organisation prompt is unchanged.
From principle to enforceable control
“Protect privacy” is too broad for a production system. Decompose it into rules defining which data classes may be retrieved, who may view them, what fields are redacted, where data may be sent and how long it is retained. Link every rule to a mechanism—IAM, DLP, tool validation or retention controls—and to positive and negative tests. Models can interpret context, but critical boundaries should be checked outside them.
Conflicts need a defined path. When helpfulness collides with confidentiality, the agent should stop the action, identify missing authority and offer a safe alternative. Logs should identify the policy version and policy-engine decision without exposing the data the control protected.
Managing policy change
Treat the constitution like code: owner, review, version, changelog, canary and rollback. A changed priority may improve refusals in one class while harming usefulness elsewhere. Run regression across languages, user roles and boundary cases, then watch escalation and manual-correction rates. A values document matters operationally only when the organisation can demonstrate its effect on system behaviour.
Connect normative policy with model governance, LLM evaluation and agent security. To translate principles into testable controls, contact BreachRoad.
Primary sources: Anthropic — Claude’s new constitution, full Claude’s Constitution.
