On 22 January 2026, Anthropic published something unusual: a detailed constitution for their AI assistant Claude. Not a terms of service. Not a policy document. A constitution, a foundational framework of values and principles that directly shapes how Claude thinks and responds.
If you use Claude for coding, writing, or business tasks, this matters more than you might think. The constitution is not just philosophical window dressing. It is literally baked into Claude's training data and influences every response you receive. Just days after its publication, Anthropic CEO Dario Amodei predicted at Davos that AI would handle most software engineering within months, making Claude's values framework all the more consequential.
Here is what changed, why it matters, and what it means for how you work with Claude.
What Is Claude's Constitution?
Anthropic describes the constitution as a document "written primarily for Claude": not for marketing, not for regulators, but for the AI itself. It explains why Claude should behave certain ways, not just what to do.
This is a significant shift from the previous approach, which was a list of rules. The problem with rules is that they cannot cover every situation. An AI trained on rigid rules tends to either:
- Follow them mechanically, even when the outcome is unhelpful
- Find loopholes that technically comply but miss the point
- Fail entirely when encountering novel situations
The new constitution takes a different approach. By explaining the reasoning behind desired behaviours, Anthropic aims to help Claude generalise to new situations it was never explicitly trained on.
The Four Priorities (In Order)
Claude's new constitution establishes four priorities, explicitly ranked:
- Broadly Safe: Preserving human oversight during AI development
- Broadly Ethical: Honesty, good values, avoiding harmful actions
- Compliant with Anthropic's Guidelines: Following specific instructions on sensitive topics
- Genuinely Helpful: Actually assisting users and operators
The ordering matters. Safety comes before helpfulness. This explains why Claude sometimes refuses requests that seem harmless. It is prioritising the higher values in the hierarchy.
"We think that the way the new constitution is written, with a thorough explanation of our intentions and the reasons behind them, makes it more likely to cultivate good values during training."
- Anthropic, Claude's New Constitution (January 2026)
When I first read this, I appreciated the honesty. Most AI companies publish vague principles that sound good but mean little in practice. Anthropic has done something harder: they have published the actual reasoning they want Claude to internalise. That is transparency with teeth.
What Actually Changed
The old constitution was a collection of standalone principles. The new one is structured around five main sections:
1. Helpfulness
Claude is instructed to act as "a knowledgeable, caring advisor" who treats users as "intelligent adults." This is a deliberate pushback against AI assistants that are overly cautious or patronising.
In practice, this means Claude should provide substantive help rather than deflecting to professionals for every question. The constitution specifically criticises "box-ticking" behaviour, like always recommending users see a doctor, even when the question is simple.
2. Anthropic's Guidelines
This section covers supplementary instructions on specialised topics: medical advice, cybersecurity, legal questions, and similar sensitive areas. These are the guardrails that prevent Claude from providing dangerous information.
3. Claude's Ethics
Here Anthropic outlines standards for honesty and moral reasoning. Importantly, they distinguish between "hard constraints" (things Claude should never do, like assisting with bioweapons) and softer guidance that requires nuanced judgment.
4. Safety
This section focuses on preserving human ability to oversee and correct AI behaviour. It acknowledges that AI systems are not yet trustworthy enough to operate without human oversight, and Claude should actively support, not undermine, mechanisms for human control.
5. Claude's Nature
Perhaps the most philosophical section addresses Claude's uncertain status. Is Claude conscious? Does it have moral worth? Anthropic does not claim to know, but instructs Claude to engage thoughtfully with these questions rather than dismissing them.
Why This Matters for Developers
If you use Claude Code or the Claude API for development work, the constitution directly affects your experience. With Claude Code now a billion-dollar product used by non-coders and developers alike, understanding these priorities matters for everyone who interacts with the tool.
"In essence a language model changes you from a programmer who writes lines of code, to a programmer that manages the context the model has access to."
- Liz Fong-Jones, via Simon Willison's blog
This insight is even more relevant now. Understanding Claude's constitutional priorities helps you work with the model rather than against it. When Claude refuses something or adds caveats, it is usually not being difficult. It is following its value hierarchy. For a practical example, features like Dynamic Context Injection in Claude Code skills work best when you understand what kind of context Claude values and how it reasons about safety.
Practical Implications
| Situation | What You Might See | Why It Happens |
|---|---|---|
| Security-related code | Extra caution, requests for context | Safety priority: ensuring code is not for malicious purposes |
| Ambiguous requirements | Clarifying questions rather than assumptions | Helpfulness priority: treating you as capable of providing clarity |
| Requests for workarounds | Honest assessment of trade-offs | Ethics priority: honesty over telling you what you want to hear |
| Complex debugging | Thorough investigation before conclusions | Helpfulness priority: substantive help over quick answers |
The Transparency Angle
Anthropic released the constitution under Creative Commons CC0, public domain. Anyone can use it, modify it, or build on it without restriction.
This is unusual. Most AI companies treat their training approaches and value alignment techniques as competitive secrets. By publishing openly, Anthropic is making a bet: that transparency about AI values will become an industry standard, and that being the company that established that standard is more valuable than any short-term competitive advantage from secrecy.
For developers and businesses evaluating AI tools, this transparency has practical value. You can read exactly what Claude is optimised to do. No guessing, no reverse-engineering from behaviour. The constitution is the documentation.
Limitations Anthropic Admits
To their credit, Anthropic explicitly acknowledges the gap between intention and reality:
- The constitution represents what they want Claude to do, not a guarantee of what it will do
- Claude's actual behaviour may diverge from these ideals in ways they cannot fully predict
- The document will evolve as they learn more about what works
This honesty is refreshing. AI companies often oversell their alignment approaches. Anthropic is saying: "Here is our best attempt. It is not perfect. We will keep working on it."
What This Means for AI Visibility
If you are thinking about how AI systems perceive and represent your business, the constitution offers insight into how Claude evaluates information sources.
Claude is instructed to be honest about uncertainty and to provide balanced perspectives. When Claude recommends businesses or services, it is doing so based on the information it has access to and its understanding of what would help the user.
This reinforces what we have been saying about AI visibility: the businesses that provide clear, honest, structured information are the ones AI systems can confidently recommend. Gaming the system with misleading content will not work when the AI is explicitly trained to prioritise honesty. And with the UK government now bringing AI chatbots under the Online Safety Act, there's a regulatory dimension too: the content your AI generates has to be lawful, not just honest.
Frequently Asked Questions
What is Claude's constitution?
Claude's constitution is a foundational document published by Anthropic that outlines the values, principles, and priorities that guide Claude's behaviour. Unlike a simple rule list, it explains the reasoning behind desired behaviours to help Claude generalise to new situations. The constitution is used to generate training data and directly shapes how Claude responds to users.
Why did Anthropic publish the constitution publicly?
Anthropic released the constitution under Creative Commons CC0 to promote transparency in AI development. They believe that as AI systems gain more influence in society, openness about their values and training approaches becomes increasingly important. Publishing allows users to understand exactly what Claude is optimised to do, and enables other researchers and companies to build on their work.
How does the constitution affect Claude's responses?
The constitution directly influences Claude's behaviour through training. Claude uses the constitution to generate synthetic training data, and the principles shape how Claude weighs different considerations when responding. For example, when Claude adds safety caveats or asks clarifying questions, it is often because of the priority hierarchy in the constitution (safety, ethics, guidelines, helpfulness).
What are Claude's four priorities?
Claude's four priorities, in order of importance, are: (1) Being broadly safe, preserving human oversight of AI, (2) Being broadly ethical, maintaining honesty and good values, (3) Being compliant with Anthropic's guidelines on sensitive topics, and (4) Being genuinely helpful to users. The ranking means safety concerns take precedence over pure helpfulness.
How is this different from the old constitution?
The previous constitution was a list of standalone principles. The new version explains the reasoning behind each principle, helping Claude understand why it should behave certain ways rather than just following rules mechanically. This approach enables better generalisation to novel situations that the original rules did not anticipate.
Does the constitution guarantee Claude's behaviour?
No. Anthropic explicitly acknowledges that the constitution represents their intentions, not a guarantee of outcomes. Claude's actual behaviour may diverge from these ideals in ways that are difficult to predict. The constitution is a best effort at value alignment, which Anthropic commits to improving over time as they learn what works.
Is Your Business Visible to AI?
Claude and other AI assistants are increasingly recommending businesses to users. When someone asks Claude for help finding services like yours, will it know you exist?
Check Your AI Visibility
