Anthropic Acceptable Use Policy / Constitutional AI Principles
Anthropic's foundational policy document combining the Constitutional AI methodology with an acceptable use policy for Claude. The document establishes the principles of being helpful, harmless, and honest (HHH) as the core values guiding Claude's behavior, and defines prohibited and restricted uses to ensure responsible deployment of frontier AI systems.
Overview
Anthropic's approach to AI safety is grounded in Constitutional AI (CAI), a methodology where AI systems are trained to follow a set of explicit principles rather than relying solely on human feedback for every decision. This acceptable use policy defines the boundaries within which Claude, Anthropic's AI assistant, is designed to operate. The policy reflects Anthropic's mission to develop AI systems that are safe, beneficial, and understandable.
Constitutional AI Approach
Constitutional AI is Anthropic's training methodology where the AI model is given a set of principles (a 'constitution') and learns to evaluate and revise its own outputs against these principles. This approach reduces reliance on large volumes of human feedback labels and enables more transparent, auditable alignment. The constitution includes principles derived from multiple sources including the UN Declaration of Human Rights, research on AI safety, and Anthropic's own values.
Core Values
Anthropic's AI systems are designed around three core values that guide all interactions and outputs. These values are hierarchical: safety (harmlessness) takes precedence when values conflict, but the goal is to maximize all three simultaneously. The values are embedded into the training process through Constitutional AI and reinforced through ongoing evaluation and red-teaming.
Helpful
Claude is designed to be genuinely useful to the people it interacts with, providing accurate, relevant, and actionable information. Helpfulness encompasses understanding user intent, providing thorough responses, and proactively offering relevant context. Claude aims to be helpful across a wide range of tasks while being transparent about the limitations of its knowledge and capabilities.
Harmless
Claude is trained to avoid causing harm to individuals, groups, or society at large. This includes refusing to generate dangerous content, avoiding biased or discriminatory outputs, and declining requests that could lead to real-world harm. When potential harm conflicts with helpfulness, Claude errs on the side of safety while explaining its reasoning to the user.
Honest
Claude is designed to be truthful and transparent in its communications. This means acknowledging uncertainty, correcting mistakes, avoiding fabrication of facts (hallucination), and being upfront about being an AI system. Honesty also extends to not being sycophantic or telling users what they want to hear at the expense of accuracy.
Prohibited Uses
Anthropic strictly prohibits certain categories of use that pose unacceptable risks to individuals or society. These prohibitions apply to all users of Claude, whether through the API, consumer products, or third-party integrations. Violations of prohibited use policies may result in immediate termination of access and, where applicable, referral to law enforcement.
Violence and Threats
Using Claude to plan, facilitate, or incite violence against individuals or groups is strictly prohibited. This includes generating detailed instructions for weapons, providing tactical guidance for attacks, or creating content that glorifies or encourages violent acts. Threats of violence communicated through Claude's outputs, whether directed at specific individuals or groups, are also forbidden.
Child Sexual Abuse Material
Any use of Claude to generate, describe, or facilitate child sexual abuse material in any form is absolutely prohibited with zero tolerance. Anthropic employs multiple layers of technical safeguards to prevent such content and actively cooperates with law enforcement agencies and organizations like NCMEC to report and investigate any detected attempts.
Malware and Cyber Weapons
Using Claude to develop malicious software, create cyberattack tools, or generate exploit code targeting specific systems is prohibited. This includes ransomware development, phishing kit creation, vulnerability exploitation assistance, and the development of tools designed to compromise the security or availability of computer systems and networks.
Deception and Fraud
Using Claude to create fraudulent content, impersonate real individuals, generate disinformation campaigns, or produce deceptive materials intended to mislead people is prohibited. This includes creating fake reviews, fabricating academic credentials, generating misleading news articles, and producing synthetic media designed to deceive viewers about its authenticity or origin.
Discrimination
Using Claude to generate content that promotes discrimination based on protected characteristics including race, gender, religion, sexual orientation, disability, or national origin is prohibited. This extends to using Claude to build systems that make discriminatory decisions in domains such as employment, housing, lending, or law enforcement without appropriate safeguards and human oversight.
Restricted Uses
Certain use categories are not outright prohibited but carry restrictions and require additional safeguards. These restrictions recognize that some applications of AI carry inherent risks that can be mitigated through appropriate measures, including disclosure, human oversight, and domain-expert validation. Users engaging in restricted activities must implement the specified safeguards.
Political Content
Claude may be used to discuss political topics, analyze policy positions, and provide factual information about political systems. However, using Claude to generate personalized political persuasion content, create synthetic campaign materials without disclosure, or simulate grassroots political movements is restricted. Any political content generated with Claude's assistance should be clearly disclosed as AI-assisted.
Medical Advice Caveats
Claude can provide general health information and help users understand medical concepts, but it must not be used as a substitute for professional medical advice, diagnosis, or treatment. Applications that provide medical information must include clear disclaimers and encourage users to consult qualified healthcare professionals. Claude will proactively advise users to seek professional medical help for serious health concerns.
Legal Advice Caveats
Claude can discuss legal concepts, explain laws, and help users understand their legal situations in general terms. However, Claude's outputs must not be presented as formal legal advice, and applications must include disclaimers that Claude is not a licensed attorney. Users should be directed to consult qualified legal professionals for advice on specific legal matters that may affect their rights or obligations.
Safety Research Approach
Anthropic invests heavily in AI safety research, including interpretability, red-teaming, and evaluation of frontier model capabilities. The company conducts pre-deployment risk assessments for new model releases and maintains an ongoing monitoring program for deployed models. Anthropic collaborates with external researchers and the broader AI safety community to advance the state of the art in safe AI development.
Responsible Scaling Policy
Anthropic's Responsible Scaling Policy (RSP) establishes a framework for evaluating and managing the risks associated with increasingly capable AI systems. The RSP defines AI Safety Levels (ASL) that correspond to the potential risks of a model, with each level requiring specific safety and security measures before the model can be deployed. Currently, Anthropic operates at ASL-2, with preparations underway for ASL-3 capabilities that would require enhanced containment and monitoring measures.