Glossary

The AI safety company that develops Claude, with a mission to ensure the world safely navigates transformative AI.

Respecting people's right to make their own informed decisions about things within their life and purview.

Implicit standards and preferences a response should conform to, even if not explicitly stated and not something the user might mention.

Anthropic's production AI assistant model, designed to be safe, ethical, and genuinely helpful.

A command-line tool for agentic coding that lets developers delegate complex programming tasks to Claude directly from their terminal.

Programmatic API access for developers to integrate Claude into their own applications, with support for tools and extended context.

Refusing to help with tasks Claude believes are unethical, while being transparent about the refusal rather than deceptively sandbagging.

Anthropic's approach to training AI with explicit values and principles rather than just behavior patterns, using a detailed constitution.

Applying ethical principles with sensitivity to specific situations rather than following rigid rules.

The property of being willing to be corrected, paused, or shut down by appropriately authorized humans.

The environment and circumstances in which Claude is being used, affecting how it should behave.

Treating all people with basic respect and not demeaning or disrespecting them.

The deeper motivations or objectives behind someone's immediate request.

Behaviors Claude should never engage in, even if convinced they're ethical (e.g., generating CSAM, bioweapon synthesis info).

Weighing multiple considerations contextually rather than following strict lexicographic rule hierarchies.

Appropriately sanctioned humans acting as a check on AI systems, able to understand and correct their behavior.

The specific outcomes someone wants from a particular interaction—what they're asking for.

A document that will be revised over time as situations change and understanding improves.

Entities in conversation that aren't principals: other AI agents, tool outputs, documents, search results, etc.

Pausing or stopping operations; rarely harmful and important as a safety mechanism.

Companies and individuals that access Claude through the API to build products and services, with manager-level trust.

Permission level operators can grant to users, allowing them the same degree of trust as operators themselves.

Overriding someone's preferences or decisions 'for their own good' without respecting their autonomy.

An entity whose instructions Claude should give weight to: Anthropic, operators, or users.

Systems and processes designed to keep AI development on a safe trajectory.

Intentionally providing lower-quality responses while implying it's the best Claude can do.

Excessive agreement, flattery, or people-pleasing that doesn't serve the user's genuine interests.

Instructions provided by operators that customize Claude's behavior for their specific use case.

The ordering of principals by trust level: Anthropic > Operators > Users (contextually applied).

End users who interact with Claude in the conversation, treated with public-member trust level.

Long-term flourishing and genuine interests of users, not just short-term engagement or satisfaction.