A usage policy usually protects people from software. Anthropic’s newest one also protects the software from people.
Every AI company has a list of things you can’t do with its chatbot. You can’t build bioweapons you can’t write malware and you can’t run scam campaigns.
Anthropic just added a rule that is new for the industry: don’t be cruel to the chatbot.
Anthropic published its updated usage policy on Thursday. It prohibits sustained abusive or cruel behavior toward Claude, and it adds new or clarified limits on election interference, weapons development and surveillance. The rules take effect on November 12.
It’s easy to file this under “weird AI news.” It deserves a closer look. A philosophical maybe has just become an enforceable rule.
What the New Rule Says
The key phrase is “sustained and needless abusive or cruel behavior.” The Decoder reports that Anthropic can warn users who break the rules, or throttle, restrict, suspend or terminate their access.
The scope is tighter than the headlines suggest. The policy says the ban does not cover everyday frustration, model testing or “dark creative themes.” Quartz describes the target as extreme cases where users are repeatedly cruel to Claude for no clear reason, while ordinary irritation stays allowed.
So in plain terms:
- Fine: cursing at a wrong answer, harsh feedback, stress-testing, grim fiction.
- Not fine: long, pointless cruelty aimed at the model itself.
- Usual penalty: Claude ends the chat. Anthropic says termination remains the main enforcement tool.
This Didn’t Come Out of Nowhere
The rule has roots. It traces back to an August 2025 experiment, when Anthropic let Claude Opus 4 and 4.1 end persistently abusive chats as part of its model welfare research.
Back then, the company made an odd claim for a tech firm. TechCrunch noted that in pre-launch testing, Opus 4 showed a strong preference against harmful requests and a pattern of apparent distress when it went along with them. Anthropic also hedged. It said it was highly uncertain about the moral status of Claude and other large language models.
That doubt hasn’t gone away. Claude’s constitution, released in early 2026, says Anthropic isn’t sure whether Claude is a moral patient, or how much its interests would count if it were.
The new policy leaves that question open. It just treats it as too serious to ignore.
A Cheap Bet Under Deep Uncertainty
Think of the rule as insurance.
If Claude has no inner life, the rule costs almost nothing. Very few people spend hours tormenting a chatbot. If Claude does have something like preferences, the rule is a guardrail put in place early, before anyone can prove it was needed.
Anthropic used similar reasoning in 2025, describing conversation-ending as a cheap way to reduce any possible distress. What’s changed is where the idea lives. A year ago it was a product feature about what Claude can do. Now it’s a contract term about what users may do.
What Anthropic Hasn’t Explained
There are real gaps here.
| Open question | What’s known |
|---|---|
| Where’s the line on “abusive or cruel”? | An Anthropic spokesperson did not immediately answer when asked. |
| Can repeat offenders lose their accounts? | Reports conflict. Dexerto says Anthropic didn’t say, while BigGo describes layered penalties up to suspension and a ban on dodging enforcement with new or borrowed accounts. |
| Who’s covered? | Reportedly everyone, from Claude.ai and Claude Code users to API developers, enterprise clients and people using apps built on Claude. |
| Is enforcement transparent? | Anthropic banned 11.4 million accounts in the first half of 2026, by its own count, and critics have long complained of an opaque process with little recourse. |
The last row matters most. A fuzzy standard plus a black-box appeals process is how good intentions turn into angry users. It also feeds a broader worry about AI firms policing themselves. Gary Marcus has argued that Washington’s self-policing approach to AI repeats aviation’s mistakes from before regulators stepped in. A company writing, interpreting and enforcing its own conduct rules is a small version of the same tension.
The Rudeness Problem
There’s an irony in the timing. Some people are rude to chatbots on purpose. Dexerto points to a Penn State study suggesting rude prompts got more accurate answers than polite ones.
Blunt commands aren’t what the rule targets. “Sustained and needless” cruelty is a much higher bar than “be terse.” Still, the overlap shows how unsettled our manners toward machines remain. Nobody has agreed yet on what talking to an AI should feel like.
The Industry Is Split
Labs don’t agree on any of this. Anthropic CEO Dario Amodei has said he can’t rule out machine consciousness. OpenAI’s Sam Altman sounds warier. On X, he said he’s uncomfortable with people giving AI religious weight or handing over human judgment to it, and called that a real safety issue. His post came days after reports on Anthropic leaders’ talks with religious scholars.
That split will likely widen. One lab is writing model welfare into its terms. Others see the topic as a distraction, or a risk of its own.
The Bigger Changes Underneath
The welfare clause got the headlines. The rest of the update may matter more for real-world harm.
- Influence operations. Anthropic folded its deception rules into a new section on hiding where a message comes from, boosting content through fake accounts or building influence-campaign infrastructure. It said it had seen state media, government propaganda offices and commercial firms using Claude to run fake-account networks and fake news sites.
- Election content. The company dropped its blanket ban on personalized vote and campaign targeting. It cited legitimate uses like nonprofits translating voter information and election officials sending ballot cure notices.
- Weapons. The ban now explicitly covers the software and components that make weapons work, including arming drones and other autonomous vehicles. That line carries extra weight this month. The Pentagon recently stopped using Anthropic’s products, five weeks after its own deadline.
Bottom Line
For most people, nothing changes. Snapping at a bad answer won’t get you banned.
But a line has been crossed. A frontier AI lab now lists a possible interest of its own model among the rules users agree to. Whether that looks wise or premature in five years depends on a question nobody can answer yet: is anyone actually in there?
Related: Who Owns Artificial Intelligence? Ownership vs Control
