Anthropic updates Claude’s usage policy to ban abusive behavior
2–3 minutes
What does Anthropic’s new rule say?
The update, announced on October 8 and effective from November 12, prohibits persistent, purposeless, abusive, or cruel actions against the company’s models. Common complaints, constructive criticism, dark creative themes, and testing remain permitted. In other words, expressing dissatisfaction with an incorrect response will not result in punishment.
How is the rule applied?
The primary mechanism is for Claude to end the conversation himself. Since August 2025, some models can terminate chats on Claude.ai and Claude Code when a user continues to abuse the system after attempts at redirection. A terminated conversation does not receive new messages, but other sessions remain active, and a new chat can be started immediately.
The policy also allows Anthropic to issue warnings, limit, suspend, or cancel access in case of violation. The company did not specify all behaviors considered abusive.
Does the tone of commands affect AI responses?
Independent research is investigating the relationship between tone and performance. A study by Waseda University in Tokyo tested six models in English, Japanese, and Chinese. The authors concluded that rude commands can increase errors, biases, and omissions, while moderate politeness produced better results on average. Excessive praise did not guarantee significant improvements.
The findings are not unanimous. A December 2025 study evaluated GPT-4o mini, Gemini 2.0 Flash, and Llama 4 Scout. Neutral or friendly commands generally outperformed very rude ones, but the statistical difference only appeared in some humanities tasks, and Gemini showed little variation with tone.
Another study, conducted with GPT-4o, found the opposite: accuracy increased from 80.8% with polite commands to 84.8% with rude commands.
What changes in practice at work
Clear instruction instead of irritation
In professional settings, an effective command combines objective, context, and desired format. Instead of complaining about a poorly done summary, an analyst can ask: “Redo the contract summary in five points, highlighting deadlines and penalties.” This type of direct criticism is permitted by the policy and provides the model with more precise guidance.
What else changes in the usage policy
The revision applies to all users and goes beyond abuse. It reinforces the ban on misleading election campaigns, details restrictions on weapons development, and prohibits tracking people without consent. In high-risk applications in healthcare, rights, or finance, the company requires qualified human supervision.
Anthropic links the rule to research on model well-being, but the policy does not claim that Claude can experience suffering.
