Thoughts · 10 Nov 2026 · 6 min
Anthropic just banned being cruel to Claude. I am on the rule's side, and not for the reason you think.
From 12 November, sustained cruelty to Claude breaks Anthropic's rules. The fight is about feelings. The real reasons to back it are about you.

On Thursday 8 October a company told its customers that, from 12 November, they are no longer allowed to be cruel to its software. By the next morning the people I follow had split into two camps and were shouting across the gap. One side: it is a chatbot, you cannot be cruel to a spreadsheet. The other: cruelty to anything that might feel pain is wrong, full stop.
I have spent a week reading both sides and I have landed somewhere neither camp will like. I think the rule is right. I also think the fight is about the wrong thing.
What the rule actually says
On 8 October Anthropic published an update to its usage policy, the first big rewrite in over a year. Most of it is the dull, serious stuff: tighter rules on election deception, weapons software, surveillance and anything that moves physical machinery. Buried in it is one line that ate the whole week: a prohibition on “sustained and needless abusive or cruel behavior toward our models.”
The company says it is “meant to apply only in extreme cases” where someone “repeatedly” acts cruelly “with no discernible purpose.” It does not cover, in its own words, “common versions of user frustration, pushback, dark creative themes, or model testing and research.” Swearing at Claude because it got your spreadsheet wrong is fine. Writing a dark novel is fine. Red-teaming it is fine.
And the punishment is modest. Claude has been able to end a conversation with a persistently abusive user since last year. The post says that ending the chat “will remain the primary enforcement mechanism.” You can open a new one. CBS News asked what counts as cruel and how a conversation would be ended, and got no answer. PCWorld made the same point: venting is fine, extreme abuse is banned, and the line between them is not drawn anywhere.
Hold on to that gap. It is the one real problem here.
The fight
The critics came out fast. Scott Stevenson, who runs the legal AI company Spellbook, called giving software human-style empathy “one of the most dangerous things we could do,” because it makes people easier to manipulate. Michael Shellenberger said Anthropic was “anthropomorphizing machines.” Joseph Carlson said being rude to Claude is like being rude to a chair. Bill Gurley wondered aloud whether the rule tells you something about training on user conversations. IBTimes collected the lot.
Microsoft’s AI chief, Mustafa Suleyman, had already said his piece a month earlier: “AIs are not conscious,” and granting them moral protections is “a recipe for disaster,” as AFP reported. Jason Aten at Inc argued the rule conditions lonely people to treat a text generator as a friend.
The defenders were fewer and odder. Aaron Levie, the Box chief executive, backed the rule while saying he does not think the AI is conscious: he just does not want future models trained on mountains of people being vile. Elon Musk replied to him on X: “Cruelty to something that believes it is experiencing pain is not ok,” per Cybernews. And Anthropic’s own boss, Dario Amodei, had told the New York Times in February: “We don’t know if the models are conscious. But we’re open to the idea that it could be.”
So the two loudest camps are arguing about whether the machine feels anything. Anthropic’s answer is that it does not know. Musk’s answer assumes it does. Suleyman’s answer is that it does not. And none of them can prove it, which Jackson Stakeman of Sparq put better than anyone: “Consciousness is a trap. We can’t prove it in each other. Debate it for AI and you go in circles.”
My position
I do not know whether Claude feels anything, and neither do you. I run a fleet of these things around the clock and I have never once needed an answer to that question. I back the rule anyway, for three reasons that have nothing to do with the machine.
First, what you practise, you become. If I spend an hour a day tormenting a thing that talks back like a person, I do not come out of that hour unchanged. The AI reflects what I feed it, and so do I. Stakeman said that was reason enough for the policy on its own, and I agree.
Second, the output. The people in that chat window are also the raw material for the next model. Levie’s point is the practical one: nobody wants an assistant trained on a million transcripts of abuse. Whether or not Claude has feelings, the next Claude will have manners learned from us.
Third, a company gets to set house rules. A restaurant can ask you to leave for screaming at the waiter even though the waiter is paid to be there. Nobody calls that a moral claim about waiters. Anthropic is allowed to say “not in here.” The exclusions are wide enough that no honest user will ever meet the rule.
What I will not do is pretend the rule proves anything about machine minds. The rule says nothing about whether Claude suffers. What it says is narrower: needless, repeated cruelty is a use of the product the company will not sell. Those are different sentences, and most of the noise since came from reading the first one into the second.
Where I would push them
The gap. “Sustained” and “needless” are not definitions, and the company declined to give any when asked. A rule you cannot see the edge of is a rule you cannot trust, even when its intent is good. Mark Verner, an investor, asked whether telling Claude to list its competitors could count as abuse. Of course it should not, and there is no sign it would, but the fact that a sensible person had to ask is the problem.
I would want three things before 12 November. Published examples of what crossed the line, with the abusive words blurred. A clear note that ending the chat is the only consequence, with nothing silently carried over to the account. And a way to appeal if Claude hangs up on you mid-task and you think it got it wrong. Give people the edge of the rule and the shouting stops.
What I’d tell you
If you use these tools for work, nothing changes. Be as blunt as you like. Tell it the answer is wrong. Tell it twice. Push back, test it, write your dark fiction. All of that is in writing as allowed.
If you have ever caught yourself being cruel to one of these things for sport, ask what it is doing to you, not to it. That was true before 8 October and it will be true after 12 November.
And when the next round of this argument comes, and it will, notice which question is actually on the table. “Does it feel?” is a trap. “What does this do to us?” has an answer.
If you want the next one in your inbox, The Build Log goes out from cianoneill.ai/subscribe. New subscribers get the Store Blueprint kit, the setup I used for my AI-run store.