New Anthropic Claude policy tightens rules on political misuse

The update, announced Oct. 8, gathers scattered older restrictions into clearer sections and adds new ones in response to misuse Anthropic says it has seen. The main changes:

• Abuse of Claude: repeated, pointless cruelty toward the model is now prohibited.
• Deceptive campaigns: a new ban covers commercial or political influence campaigns that hide who is behind a message or boost it with fake accounts, fake posts or made-up news outlets.
• Elections: a section called “Do Not Undermine Democratic Processes” bars misleading people about candidates or how to vote, impersonating candidates or election officials, and trying to suppress turnout.
• Weapons: the existing weapons ban now explicitly covers software and components that make weapons work, plus tasks such as arming drones and other autonomous vehicles.
• Surveillance: tracking people without consent is banned, whether live or by analyzing data collected earlier.

The policy also clarifies rules for high-risk decisions that affect people’s lives, along with uses in health and finance.

Anthropic says the abuse rule is meant only for extreme cases, where someone keeps acting cruelly toward the model with no apparent purpose. It says the rule doesn’t cover everyday frustration, pushing back on a wrong answer, dark themes in creative writing, or testing and research on the model.

In practice, that means you can still tell Claude it’s wrong, argue with it, or write a horror story with it. Security researchers can keep stress-testing it. What Anthropic hasn’t published is a detailed list of examples showing exactly where the line sits, so the standard is broad rather than a precise checklist.

Enforcement is light for now. Since August 2025, Claude has been able to end conversations with users who are persistently harmful or abusive on Claude.ai and in Claude Code. Anthropic says ending the conversation remains the “primary enforcement mechanism.” An ended chat doesn’t stop you from opening a new one, and Anthropic didn’t say whether repeat offenders could face account bans.

https://t.co/8XJVIXDpQz

— Mustafa Suleyman (@mustafasuleyman) September 14, 2026

The rule grows out of Anthropic’s research into “model welfare,” a program it launched in April 2025 to explore whether advanced AI systems could one day have experiences that deserve moral consideration. Anthropic does not claim Claude can suffer. It says there is no scientific agreement on whether AI can be conscious, and CEO Dario Amodei said earlier this year that the company doesn’t know whether its models are conscious.

Not everyone in the industry agrees with the approach. Mustafa Suleyman, who leads Microsoft AI and has argued against the idea of model welfare, responded to the announcement with a melting-face emoji. A draft Microsoft AI code of conduct he shared in September said model welfare was the wrong idea and that AI should not have rights or legal personhood. Barry Scannell, a technology lawyer at the Irish firm William Fry, argued on LinkedIn that people can’t be cruel to “numbers and maths.”

The surveillance changes follow Anthropic’s own threat reports, which found people increasingly using Claude-powered tools to identify and track political dissidents. The new policy also says police and prosecutors can’t use Claude to pick or suggest which people to investigate, arrest or put on trial, and can’t be used to build or improve surveillance tools.

There is an exception. Anthropic says it may adjust these rules for some government customers if, in its judgment, contract terms and safeguards are strong enough to limit the harm. The company has worked with the US military before.

A brand-new rule covers AI hooked up to hardware. If an Anthropic model controls equipment that can act on its own and could injure someone, a qualified operator must be able to watch it and stop it. The policy doesn’t say whether that operator has to be a person. It hints at where the industry is heading as AI labs move into robotics.

For most people, very little. If you use Claude for writing, coding or research and occasionally get annoyed at it, nothing in the abuse rule applies to you.

Businesses and developers have more homework. Anyone using Claude in sensitive decisions about people, in political or marketing messaging, in security tools or connected to physical equipment should read the full policy before Nov. 12 and check that their setup complies, including human oversight for anything that moves.