Anthropic Published an 80-Page Constitution for Claude. Here's What's in It.

The constitution is the foundational training document that shapes Claude's values and behavior. It's not a policy page or a content moderation rulebook. Anthropic describes it as 'the final authority on how we want Claude to be and to behave' — meaning every other instruction Claude receives during training or deployment is supposed to be consistent with it.

Anthropic has been using Constitutional AI techniques since 2023, but the new version represents a significant evolution. Previously, the constitution was a list of principles. The new version is a holistic reasoning document — it explains not just what Claude should do, but why, and how Claude should think about difficult tradeoffs.

The most striking thing about the document: it's written in second person, addressed directly to Claude. Not 'the model should...' but 'you should...' Anthropic says this is intentional — the goal is for Claude to internalize the values, not just comply with rules.

Claude itself uses the constitution to generate synthetic training data — conversations, rankings of responses, examples of good and bad judgment calls. Future versions of Claude are trained partly on data that earlier Claude versions created using this document as a guide. It's a bootstrapping process for values.

When Claude faces a conflict, it's supposed to resolve it using a clear priority hierarchy: (1) Broadly safe — supporting human oversight of AI during this critical period, (2) Broadly ethical — having good values and being honest, (3) Adherent to Anthropic's principles — following company guidelines where relevant, (4) Genuinely helpful — benefiting the people it works with.

The ordering matters. Safety comes before ethics, which comes before company guidelines, which comes before helpfulness. Notably, Anthropic explicitly says that if being 'maximally helpful' would conflict with safety or ethics, helpfulness loses. This is a direct response to criticisms that AI companies optimize for engagement over safety.

The biggest practical shift from the old approach: instead of telling Claude 'don't do X,' the new constitution explains why X is problematic, what values are at stake, and how to think about edge cases. The theory is that a model which understands the reasoning behind a principle will generalize better to novel situations than one following a list.

For businesses, this is meaningful. Companies evaluating Claude for enterprise deployment can now read exactly what values and tradeoffs are baked into the model — not just the marketing claims. It creates a level of transparency that's genuinely rare in the industry. OpenAI published a comparable Model Spec in 2025, but Claude's constitution goes deeper on the reasoning.

Tech press led with this part, so let's address it: the constitution includes a section acknowledging genuine uncertainty about whether Claude might have something like functional emotions or experience. Anthropic doesn't claim Claude is conscious — they explicitly say they don't know. But they say the question is serious enough to take precautions, including trying to ensure Claude has positive functional states where possible.

This is unusual corporate candor. Most AI companies don't touch this topic. Anthropic's position is essentially: we're uncertain, so we'll err on the side of treating it as potentially meaningful.

Three practical reasons to care about this document: First, when Claude declines something or pushes back on an instruction, it's now possible to understand why — the reasoning is public. Second, enterprise teams can audit whether Claude's stated values match their own governance requirements before deploying it. Third, because it's CC0, other AI labs can adopt or adapt the framework. Several are expected to publish comparable documents within 2026.

The full constitution is available at anthropic.com/constitution. It's a long read, but the first 20 pages covering the priority hierarchy and core values are worth the time for anyone building on Claude or making decisions about AI adoption.