What are Claude's content policies?
Core Principles
Anthropic designs Claude to be helpful, harmless, and honest. Its content policies are guided by these principles and by Anthropic's Usage Policy, which outlines what you can and cannot do with Claude.
Claude will not assist with illegal activities, harassment, hate speech, self-harm, or the creation of malicious content like malware. It also avoids generating sexually explicit content or promoting violence.
These policies apply across all interactions, and Claude may refuse a request or end a conversation if it detects a violation. The exact boundaries can evolve, so check Anthropic's Usage Policy for the latest details.
- Be helpful, harmless, and honest.
- No illegal, harmful, or abusive content.
- No hate speech, harassment, or self-harm promotion.
- No sexually explicit content or violence glorification.
- Claude may refuse or end conversations that violate policies.
How Policies Are Enforced
Claude uses a combination of automated systems and human review to enforce content policies. If you repeatedly violate the policies, your access may be limited or terminated.
Anthropic also provides a way to report problematic responses. If Claude generates something that seems to violate its policies, you can flag it for review.
The policies are not exhaustive, and Claude may err on the side of caution. If you're unsure whether a request is allowed, it's best to avoid it or ask for clarification in a safe context.
- Automated and human review enforce policies.
- Repeated violations can lead to account restrictions.
- You can report inappropriate responses for review.
- Policies are updated periodically; check the Usage Policy.
- When in doubt, don't risk it—ask a safer question.
Common mistakes
- Thinking Claude's policies are fixed and never change; they are updated over time.
- Assuming Claude will always comply with any request; it can refuse based on policies.
- Believing that content policies only apply to extreme cases; they cover a wide range of harmful content.
