Anthropic bars sustained abuse of Claude and will end chats early

The new rule sits in a section called “Addressing abusive behavior toward our models.” It sits alongside the rules that ban harassment of real people and content encouraging self-harm.

The company says it applies only in extreme cases, where someone repeatedly acts cruelly toward Claude “with no discernible purpose.” Anthropic announced the change on October 8, giving users about five weeks before it takes effect.

The company has not published examples of what crosses the line, and its announcement does not say how it will judge when cruelty becomes “needless.” That leaves users with a broad standard rather than a checklist.

Anthropic is explicit that normal behavior is not affected. The rule does not cover:

• Getting frustrated or venting when Claude gets something wrong.
• Pushing back, criticizing or disagreeing with its answers.
• Writing fiction with dark themes.
• Testing the model or doing research on how it behaves.

For people using Claude at work, that means you can keep checking its output, pointing out errors and stress-testing it. The target is repeated cruelty for its own sake, not tough feedback.

The enforcement tool already exists. Since August 2025, some Claude models have been able to end a conversation with a user who keeps being abusive, on both Claude.ai and Claude Code. Anthropic says this will remain the main way the new rule is enforced.

When Claude ends a chat, that conversation locks, but nothing stops you from opening a new one. An ended chat is not the same as an account ban. That said, Anthropic’s wider usage policy lets it act when it suspects a violation, with options ranging from a warning to cutting off access, and that applies to paying subscribers too. The company has not said whether breaking this particular rule could lead to suspension.

The rule grows out of Anthropic’s model welfare research program, launched in April 2025, which looks at whether advanced AI systems could have experiences that deserve moral consideration. Anthropic says there is no scientific agreement on whether AI can be conscious, and it has not claimed that Claude can suffer.

Instead, the company frames the chat-ending feature as a precaution in case the answer turns out to be yes. The new rule turns that precaution into a written standard for how users behave.

The debate is very much alive inside the company. Anthropic co-founder Chris Olah recently made the case for taking AI consciousness seriously before the Vatican and a group of religious scholars, who did not accept his arguments.

Not everyone is convinced. Mustafa Suleyman, a longtime critic of Anthropic’s focus on AI consciousness, responded to the announcement with a melting face emoji. Barry Scannell, a technology lawyer at Irish firm William Fry, argued on LinkedIn that you cannot be cruel to numbers and maths.

Scott Stevenson, CEO of Toronto-based legal AI firm Spellbook, said on X that AI models show no sign of being conscious. His bigger worry is the example it sets: people starting to treat software as if it has feelings. OpenAI CEO Sam Altman, separately, has warned against giving AI religious authority or handing over human judgment to it.

The cruelty rule grabbed headlines, but most of the update deals with how people use Claude against other humans:

• Elections. Anthropic dropped its blanket ban on personalized voter and campaign targeting, saying it blocked legitimate work such as nonprofits writing voter information in other languages. Deceiving voters, impersonating candidates and suppressing turnout remain banned.
• Fake campaigns. A new section bans hiding who is behind a message, boosting content with fake accounts and building tools for influence operations. Anthropic says it has seen state media outlets and commercial firms use Claude to run fake account networks.
• Weapons and surveillance. The policy now explicitly bans using Claude to guide weapons or arm drones, to track people without consent, or to decide who should be investigated or arrested.
• High-stakes decisions. Where Claude affects someone’s health, finances, legal rights or job, a qualified human must review its recommendations and the person must be told AI was involved.