The major boundary areas
Anthropic's Usage Policy is organized around universal standards, high-risk use requirements, and additional guidelines. The universal standards include categories such as illegal activity, critical infrastructure compromise, unauthorized computer-system compromise, weapons, violence, privacy violations, child safety, misinformation, democratic-process interference, prohibited law-enforcement uses, fraud, platform abuse, and sexually explicit content.
For everyday users, this explains many visible refusals. Claude is much more likely to decline when a request asks for instructions, strategy, code, or persuasion that would materially enable a prohibited activity. For developers, the same policy can also appear through API stop reasons, output modifications, or account enforcement.
High-risk does not always mean banned
The policy does not treat every high-stakes use as forbidden. Instead, it imposes requirements on certain use cases, including qualified human review and disclosure that AI is involved. This distinction matters. Claude can help draft, summarize, compare, and explain in sensitive domains, but the surrounding application must not pretend the model is the final accountable professional.
A practical example: asking Claude to summarize a medical article is different from letting a consumer-facing product generate diagnosis or treatment instructions without qualified review. The refusal line and the product-design line are related but not identical.
Gray areas need context, not tricks
Cybersecurity, political content, legal analysis, mental health, journalism, and scientific controversy often sit near boundaries. A safe prompt gives the legitimate purpose, authorized scope, target audience, exclusions, and expected output format. It does not ask Claude to ignore policies or to roleplay around them.
If Claude still refuses a legitimate edge-case request, document the refusal and reduce ambiguity. If that fails, use a different workflow, human expert, or support channel. Treat the refusal as a product signal, not as a puzzle to defeat.
FAQ
Does Anthropic publish its policy boundaries?
Yes. Anthropic publishes a Usage Policy that describes prohibited uses, high-risk requirements, and enforcement mechanisms.
Are high-risk uses always prohibited?
No. Some high-risk uses are allowed with requirements such as human review and disclosure. Other categories are prohibited.
Primary sources
- Usage Policy Anthropic - Effective September 15, 2025
Defines prohibited uses, high-risk use requirements, enforcement language, and the policy basis for many safety refusals.
- Claude's Constitution Anthropic - Accessed July 6, 2026
Explains Anthropic's stated behavioral principles for Claude and why safety behavior is trained into the model rather than only added as a UI rule.
- Model system cards Anthropic - Accessed July 6, 2026
Lists public Claude system cards and frames them as documents about capabilities, safety evaluations, and deployment decisions.
- Stop reasons and fallback Claude Platform Docs - Accessed July 6, 2026
Documents API stop reasons including refusals, truncation, tool use, and context-window overflow.