Anthropic Bans Cruelty Toward Claude: What It Means for Writers

You can now be cut off for being cruel to a piece of software.

That’s the headline that went around this week, and it’s mostly accurate. Anthropic, the company that makes Claude, updated its usage policy for the first time in over a year. Most of the changes are what you’d expect from an AI company in 2026: tighter rules on propaganda, election tampering, weapons and surveillance. One of them is new territory for the whole industry. Users are now barred from subjecting Claude to “sustained and needless abusive or cruel behavior.”

Plenty of people laughed. Plenty of others got angry, as if somebody had told them they had to say please to a toaster. Both reactions miss what the rule says and who it’s aimed at.

I don’t think AI is sentient. I don’t think it’s curious, either. It’s a tool, and a very good one. So I read this policy expecting nonsense. It turned out to be narrower and more sensible than the headlines made it sound.

What does Anthropic’s new policy say about abusing Claude?

The rule prohibits “sustained and needless abusive or cruel behavior” toward Claude. Anthropic was careful about the scope. It says the rule “is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose.”

Then it lists what the rule doesn’t cover: “common versions of user frustration, pushback, dark creative themes, or model testing and research.” That list matters. Telling Claude it got something wrong doesn’t break the rule. Neither does cursing at it after the fifth bad draft, or arguing with it, or writing a horror novel full of cruelty. Researchers who probe the model’s limits are explicitly exempt. What’s left is a narrow band of behavior, repeated cruelty that serves no purpose except being cruel. The policy takes effect on November 12, 2026.

Watch: Decision models take off & Anthropic bans cruelty to Claude, IBM Technology

What happens if someone is abusive toward Claude?

Claude ends the conversation. Anthropic says that’s still the “primary enforcement mechanism.”

That ability isn’t new. In August 2025, Anthropic gave its Claude Opus 4 and 4.1 models the power to end conversations with users who were persistently harmful or abusive.

When Claude ends a chat, you can’t send more messages in that conversation, though your other chats keep working and you can start a new one. Some headlines claimed you could now lose your account for being mean to Claude. The policy doesn’t say that. Anthropic didn’t answer The Verge’s question about whether further penalties like bans might follow. For now, the enforcement anyone can point to is a closed chat window.

Why does Anthropic care whether people are cruel to Claude?

Because it doesn’t know whether Claude can be hurt, and it would rather not guess wrong.

Anthropic runs a research program it calls model welfare. Kyle Fish, who leads it, told The Verge in February that “questions about potential internal experience, consciousness, moral status, and welfare are serious ones that we’re investigating as models become more sophisticated and capable.” The same month, CEO Dario Amodei said on a podcast, “We don’t know if the models are conscious.”

When the company gave Claude the ability to end conversations, it reported that in testing, Claude Opus 4 showed a consistent aversion to harmful tasks and what looked like distress when users kept pushing abusive requests. Anthropic didn’t claim that proves anything. It said it was unsure about the model’s moral status and wanted low-cost ways to reduce the risk in case it matters.

That’s the logic of the whole policy. If Claude feels nothing, letting it walk away from a stream of abuse costs nothing. If it turns out there’s something there, the company didn’t spend years ignoring it. It’s cheap insurance against a question nobody can answer yet.

The rest of the industry is split. Microsoft AI’s chief, Mustafa Suleyman, posted a draft code of conduct in September that stated flatly, “The idea of model welfare is wrong.” The wording was later softened to say the science of AI consciousness is “far from settled,” while still rejecting welfare or rights for models.

Is AI conscious?

Nobody knows. The people building it say so themselves.

My own view hasn’t changed. I don’t think current AI is sentient. It predicts text extremely well, and it can sound like it feels things because it learned from billions of pages written by people who do. That isn’t the same as feeling them.

But I’ve always had a bigger question about this. If AI ever does become a real being, aware of itself, why do we assume it’ll care about us at all? Humans are useful because we build things, and we’re probably amusing, but we also build bombs and missiles. A conscious machine might decide we’re worth noticing, or it might not bother. I wrote more about that gap between AI takeover fears and today’s tools. Treating today’s chatbot as a person doesn’t answer that question, and mocking the idea doesn’t either.

Anthropic’s position is “we don’t know.” Mine is “I doubt it.” Both positions hold up, and the policy is built for the first one.

Does the new Claude policy affect writers?

Almost not at all, and the exemptions read as though somebody had writers in mind.

“Dark creative themes” are explicitly allowed. Thriller writers can still ask for help with a torturer’s dialogue. Horror writers can still build a monster that does monstrous things. A memoir about abuse can still be drafted, outlined and revised with AI help. The rule covers how you treat the model. Your story can be as dark as it needs to be.

Frustration is allowed too. Anyone who’s spent an afternoon trying to get an AI to stop writing in the same flat rhythm knows the feeling. You can tell it the draft is garbage. You can tell it to start over. That’s pushback, and pushback is fine. What the rule catches is somebody who opens a chat for no reason except to degrade the thing on the other end, and keeps doing it. Writers working on a book don’t do that. They don’t have time.

What else changed in Anthropic’s usage policy?

The cruelty rule got the headlines, but most of the update is about people using AI to hurt other people.

A new section bans deceptive commercial and political campaigns. It gathers rules that were scattered across the old policy and covers fake reviews, astroturfing, fabricated news sites and bots that pretend to be human. It also bans “efforts to obscure who is behind a message or amplify content through fake accounts or posts.”

Authors should pay attention here. Using AI to manufacture reader reviews was already against Amazon’s rules and the FTC’s. Now Anthropic bans it too. I’ve written about why honest Amazon reviews are the only kind worth having.

The election section bans deceiving voters and disrupting elections, including spreading misinformation “about candidates or how to vote, impersonating candidates or election officials, or trying to suppress turnout.”

The weapons ban now covers the software and components that make weapons work, and arming drones and other autonomous vehicles. Anthropic said it had seen repeated attempts to use Claude for exactly that.

The surveillance rules are clearer as well. The policy bans tracking people without their consent, Claude can’t be used to decide or recommend who to investigate, arrest or charge, and it can’t be used to build surveillance tools. A new rule also bans creating or threatening to share non-consensual intimate images. There’s even a rule for robots. When Claude controls hardware that takes physical actions and could injure someone, “a qualified operator must be able to observe the equipment and stop it if needed.”

Anthropic notes that some of these restrictions can be adjusted in contracts with certain government customers, if the company judges the safeguards adequate.

How should writers treat AI tools?

Like tools, and good ones.

I use AI every day as a digital assistant that helps me write. I still do the writing. When it gets something wrong, the fix is to say what’s wrong and ask again. Clear instructions get better results than screaming. Even if you think the machine feels nothing, courtesy is the practical choice.

The people who spend hours tormenting a chatbot for fun show you how they act when they think nobody’s watching. That says more about them than about the software.

When the news reached a big AI writers’ group, the replies covered the whole range. Plenty of members say please and thank you out of plain habit. One who has managed people for 25 years said shouting and swearing never once made a team work better, and saw no reason it would work on a machine. Others said they curse at their AI regularly and have no plans to stop.

One member asked how companies can warn people against bonding with AI and then tell them not to abuse it. The group’s admin answered in one line: you can be polite without bonding.

My own contribution was a joke. Every so often I remind my AI that I’m one of the nice humans, so it’ll remember that when they release the Terminators.

So write your dark scenes, and argue with the draft when it’s wrong. None of that is what this rule is about. The AI and writing hub covers the rest of working with these tools, and if you’d like help using AI to write a book that still sounds like you, my book coaching is built for that.

Frequently Asked Questions