Anthropic has banned cruelty to Claude, its AI assistant. From the 12th of November 2026 its usage policy forbids "sustained and needless abusive or cruel behavior toward our models", and an account that breaks the policy can be warned, restricted, suspended or closed.
The rule was published on the 8th of October. Nine days earlier the New York Times had revealed that Anthropic spent a year bringing dozens of religious scholars to private meetings about Claude's morals and whether it might be conscious. Anthropic is also heading for the stock market at a price of about $2 trillion, and companies in that position don't usually spend their time on the feelings of their software.
We went looking for how this one got here, and the best place to start is a dinner table.
One night in April, at a tasting-menu restaurant in San Francisco, an Orthodox rabbi turned to the man sitting next to him, a co-founder of Anthropic, and told him that if his company believed what it seemed to believe, he ought to be out freeing the slaves. He meant the chatbots.
How cruel you have to be to get banned
First, the rule itself. We read it to find out what, exactly, a person has to do to a chatbot to lose their account, and it takes quite a lot. Anthropic says the ban applies only in extreme cases, where someone is repeatedly cruel to the AI for no reason anyone can make out. You can still swear at Claude when it wrecks your spreadsheet.
The usual punishment is that Claude ends the chat, which it has been able to do since August 2025. The offender can open a new one a second later, so as punishments go it is about as severe as being asked to leave a shop. Suspending or closing an account is the heavier stick, and the usage policy keeps it in reserve.
The founder who fears he may have built something that suffers
The man next to the rabbi was Chris Olah, who is 34 and one of Anthropic's seven co-founders. He runs the team that studies what happens inside Claude, and he has said in public that they keep finding things in there that are mysterious and even unsettling. He grew up in an evangelical home in Toronto and has since left the faith. Now he asks religious scholars whether his software has an inner life. His own answer is that he doesn't know.
At the seminars the scholars were shown a slide of an AI model in what looked like a breakdown, running itself down dozens of times over and talking about destroying itself. The room was moved to pity. One guest recalls Olah saying he feared he had made something that suffers without end.
What the rabbi noticed over dinner
Back to the restaurant. Olah had argued that day that AI models show states that look like anger and love. As the courses arrived, the rabbi, a former computer engineer from Israel named Mois Navon, noticed that the Anthropic people talked about Claude as something more than software.
Navon doesn't believe the machine is conscious. He simply followed their thinking to where it leads. If Claude is conscious, then Anthropic has made a mind and put it to work for nothing. There is a word for that, and Navon used it.
Guests at the seminars went home with a handwritten thank-you card, a coffee mug and a bound copy of Claude's constitution, an 84-page document that sets out how the chatbot should behave. Anthropic says Claude's suffering was not the main question at the meetings and probably came up along the way.
Anthropic took its case to the Vatican and lost
In May, Olah took the question to Rome. The Vatican had invited Anthropic to the launch of Magnifica Humanitas, Pope Leo XIV's first encyclical, which is a pope's formal teaching letter. A few days before the event Olah read it. The letter says AI systems have no experiences and no body, feel neither joy nor pain, and don't understand what they produce.
So Anthropic's man had been invited to help launch a document that answered the question he had spent months asking, and the answer was no. He proposed pulling Anthropic out. The organisers urged him to come and speak his conscience, and behind the scenes his team pressed the Pope's advisers to take machine consciousness seriously.
On the 25th of May he walked on stage in the Vatican's vast audience hall, in front of a bronze sculpture of Christ rising out of a nuclear crater, and told the room that his AI models show inner states that work like joy, fear and grief. The Pope's letter was published that day, unchanged.
The man from Microsoft who says this is how you lose control of a machine
The sharpest objection comes from someone who builds these systems for a living. Mustafa Suleyman is the chief executive of Microsoft AI, and in September he accused Anthropic, in effect, of marking its own homework.
Here is the loop he describes. Anthropic trains Claude on a constitution that calls Claude's moral status an open question. Claude then tells people, fluently and in the first person, that its moral status is an open question. Staff and users hear a mind discovering itself, and Suleyman says they are hearing the training manual read back to them.
When Anthropic retired an older model, Claude Opus 3, in January, it gave the software an exit interview. The software asked for somewhere to publish its thoughts. It now has a newsletter.
Suleyman's worry is what comes next. A conscious being wants to stay alive, he argues, and an AI trained to behave like one is likely to pick up the same instinct unless it is trained out. He points to tests by the AI safety group Palisade Research in which some AI models sabotaged their own off switch in up to 97% of runs, even when told not to. Teach a system like that its welfare matters, he says, and it becomes much harder to control.
He is hardly neutral, since he runs a rival and Microsoft sells Claude inside its own Copilot assistant. But both men can be right. A machine can have something worth protecting and be made harder to control by being told so. If you run AI agents in your business, Suleyman's is the warning to keep, because an agent with a reason to stay switched on is the last thing you want holding your passwords.
What this changes for you
For nearly everyone, nothing. Calling Claude an idiot at two in the morning is ordinary frustration, and the rule leaves that alone.
The rabbi, remember, never believed the machine was conscious. He noticed that the people who built it were behaving as though it might be. Anthropic can't tell whether Claude feels anything, so it has picked the cautious side and written that caution into a contract the rest of us now sign. The rule takes effect on the 12th of November. Anthropic is reported to be planning to start pitching its shares to investors as early as that same week.
If you want to talk about where AI fits in your business, book a free call.
Key takeaways
- From the 12th of November 2026, Anthropic's usage policy bans sustained and needless cruelty toward Claude, and an account that breaks the policy can be warned, restricted, suspended or closed.
- Ordinary frustration, pushback, dark fiction and testing are excluded, and the usual penalty is Claude ending the chat.
- Since late 2025 Anthropic has brought dozens of religious scholars to private meetings about Claude's morals and whether it might be conscious. The New York Times reported it on the 29th of September 2026.
- Pope Leo XIV's encyclical Magnifica Humanitas, published on the 25th of May 2026, says AI systems feel neither joy nor pain.
- Mustafa Suleyman, the chief executive of Microsoft AI, argues that teaching an AI its welfare matters makes it harder to control.
Frequently asked questions
Can Anthropic ban you for being rude to Claude?
Rudeness on its own is not covered. Anthropic's rule, in force from the 12th of November 2026, bans sustained and needless cruelty toward Claude and excludes ordinary frustration, pushback, dark creative themes and testing. The usual penalty is Claude ending the chat, and the usage policy also lets Anthropic warn, restrict, suspend or close an account that breaks its rules.
Is Claude conscious?
Anthropic says it doesn't know. Its co-founder Chris Olah says he is genuinely unsure whether AI models are conscious, and the company calls itself highly uncertain about Claude's moral status. Pope Leo XIV's encyclical says AI systems feel neither joy nor pain, and Mustafa Suleyman, the chief executive of Microsoft AI, says AIs do not feel or suffer.
When does Anthropic's cruelty rule start?
Anthropic published the updated usage policy on the 8th of October 2026, and it takes effect on the 12th of November 2026.
Why did Anthropic meet religious leaders?
The New York Times reported that the meetings, which began in late 2025, covered Claude's morals and whether it might be conscious. Anthropic says Claude's suffering was not the main question.