Anthropic bans abuse of Claude in new Usage Policy

Anthropic updates this policy every year, and it says most of the changes this time are about clarifying rules that already existed. The backdrop is that Claude has taken on longer, more autonomous work over the past year, so the new version adds examples of how the rules apply to newer capabilities. It also folds in misuse patterns Anthropic reported in a threat intelligence report published in September, covering influence operations, weapons development, and surveillance.

The cruelty clause is tied to a feature already live in Claude.ai and Claude Code, where Claude can end a conversation with a user who keeps abusing it. Anthropic says that ending conversations will remain the main way the rule is enforced. The deceptive-activity rules were previously scattered across sections on elections, fraud, privacy, and disinformation; the election section is now narrowed to banning deception of voters and interference with elections. On weapons and surveillance, Anthropic says the clarifications — including software and parts that make weapons work, arming drones, tracking people without consent, and recommending investigation or arrest targets — do not change actual enforcement.

High-risk uses that affect health, legal rights, or finance still require a qualified person authorized to review and correct the AI's suggestions, plus telling affected people that AI is being used. Because many users asked whether their use counted, Anthropic now lists covered and uncovered activities across 11 fields, including legal, medical, financial, and employment. It also adds requirements for autonomously operating equipment that could cause injury, based on the Model Hardware Standard it announced in August — a qualified operator must be able to monitor and stop the device, and it must stay safe if the connection to Claude drops.