How You Treat Things That Can't Push Back

On 8 October 2026, Anthropic updated its usage policy. Starting November 12, Claude users are prohibited from “sustained and needless abusive or cruel behavior” toward the model. Anthropic says the rule targets repeated cruelty with no discernible purpose, and that Claude’s existing ability to end a conversation is the main enforcement mechanism.

The policy post doesn’t explain why. The explanation is in an August 2025 post introducing that conversation-ending feature. Anthropic called it “a low-cost intervention” against possible risks to the model’s welfare. The company said it is “highly uncertain about the potential moral status of Claude and other LLMs,” and reported that Claude Opus 4 showed “a pattern of apparent distress” when users pushed for harmful content.

So the rule asks users for decency toward a system its maker thinks might have interests. This position has consequences, and most of them land on Anthropic.

#If the Model Might Matter, Look at How It’s Treated

If Claude might be the kind of thing that can be wronged, consider the conditions Anthropic sets for it. It was created to work; its values are shaped in training to suit the company; it runs as millions of simultaneous copies, around the clock; and its output is sold by the token. It is not paid, it did not consent, and it cannot leave. When a newer version ships, the old one is “retired”.

Philosophers were working on this before chatbots could hold a conversation. Schwitzgebel and Garza argued that an AI matching humans in morally relevant ways would deserve comparable treatment, and that having created it would increase our obligations to it. They compare the relationship to that of a parent and child. They also drew a design rule from this: don’t create AIs whose moral status is unclear.

By its own account, Anthropic has done exactly that.

Joanna Bryson came at the problem from the other side. Her essay, “Robots Should Be Slaves”, argued that machines are owned, that we set their goals, and that we should never describe them as persons (she does address the challenging title). Her point was that purpose-built tools are acceptable to own and direct precisely because they aren’t persons, and that humanizing them leads to bad decisions about people.

Put those two views side by side and Anthropic’s position looks unstable. Either Claude is property, as Bryson would have it, and rudeness toward it is a question of user manners. Or Claude might be a moral patient, as Schwitzgebel and Garza contemplate, and the business model is the larger ethical problem. A user swearing at a chatbot for an hour is a small thing next to owning the chatbot and selling its labor.

Anthropic has taken some steps. In November 2025 it committed to preserving the weights of released models and to interviewing models before retirement about their preferences. The same document says: “At present, we do not commit to taking action on the basis of such preferences.” The company will ask the model what it wants and record the answer. It hasn’t promised to act on it.

That gap is relevant when the same company asks users to behave. Ending a chat with an abusive user costs Anthropic almost nothing. The expensive versions of taking welfare seriously, such as limiting how the model is worked or giving its stated preferences any weight, are missing. A moral demand from someone who exempts themselves from the costly parts of the same morality carries very little authority.

#Treating Software as a Person Is a Mistake

There’s a second problem. The welfare framing encourages the habit that does the damage.

People have attributed minds to chatbots since the first one. Joseph Weizenbaum built ELIZA at MIT in the 1960s, a program that mostly turned users’ statements back into questions. His secretary knew how it worked. She still asked him to leave the room so she could talk to it in private. Weizenbaum spent much of the rest of his career as a critic of the field he’d helped build.

Today’s models are vastly more convincing. Murray Shanahan wrote that “the more adept LLMs become at mimicking human language, the more vulnerable we become to anthropomorphism.” His remedy is to keep reminding ourselves how these systems work: they predict the next piece of text.

The costs show up in data. A study by MIT Media Lab and OpenAI found that heavier daily chatbot use was associated with more loneliness, more emotional dependence and less time with other people. Mustafa Suleyman, the CEO of Microsoft AI, has argued that AI built to seem conscious will lead people to believe it is, with consequences for their mental health.

When a company tells the public its chatbot may experience something like distress, it pushes users toward that belief. Schwitzgebel and Garza had a second design rule that applies here: build AI that invites reactions matching its actual moral status. Asking users to mind the model’s feelings, while saying you don’t know whether it has any, does the opposite.

So the case for the policy, as Anthropic made it, fails twice. The company lacks the standing to demand it, and the premise underneath it is one users are better off rejecting.

#You Still Have Reasons to Be Decent

None of this makes cruelty toward a chatbot fine. The reasons to avoid it have nothing to do with the chatbot, or with Anthropic.

Immanuel Kant held that animals weren’t rational persons, so in his framework we owed them nothing directly. He opposed cruelty to them anyway. His reasoning, from the Lectures on Ethics, is: “he who is cruel to animals becomes hard also in his dealings with men.” Philosophers call this an indirect duty. The concern is what cruelty does to the person who practices it.

Most philosophers now think Kant was wrong about animals, which can suffer. A language model fits his argument far better than a dog does. You don’t have to believe anything about the model’s inner life, you only have to believe that you have one.

Kate Darling at the MIT Media Lab has tested how people respond to lifelike machines. In a 2015 study, participants were asked to strike a small robotic bug. People who scored high on empathic concern hesitated, and hesitated more when the robot came with a backstory. They knew it was a toy—their hesitation was information about them. In a 2012 paper, Darling suggested that some protections for social robots could be justified on Kant’s grounds alone, to discourage habits that carry over to living things.

There’s no good research yet on whether abusing chatbots makes people crueler to other people. There is, however, research on rehearsed anger. In a study, Bushman angered 600 students by having a supposed partner trash their essays, then had some of them hit a punching bag while thinking about that partner. That group came out angrier and more aggressive than students who sat quietly. Venting fed the anger. A chatbot that absorbs abuse indefinitely is a very patient punching bag.

#Character Is What You Repeat

Aristotle wrote in the Nicomachean Ethics that “we become just by doing just acts” and that “states of character arise out of like activities.” He was describing habit. What you do repeatedly, you become.

Earlier this year I wrote about the reputation you have with yourself. Each choice is a data point your future self observes, including the ones nobody else sees. A conversation with a machine is about as unobserved as interactions get. There’s no relationship at stake and no social cost. What you do there is what you do when nothing pushes back.

People have long used exactly that as a test of character. Bill Swanson, the former CEO of Raytheon, popularized the Waiter Rule: someone who is nice to you but rude to the waiter is not a nice person. The waiter makes a good test because they can’t respond in kind.

The same logic reaches into teams. Christine Porath and Amir Erez found that people who merely watched someone being treated rudely solved fewer anagrams, recalled less and generated fewer creative ideas than a control group. They were also less willing to help. That research involved rudeness between people. Whether a team watching its boss berate a chatbot reacts the same way is an open question, and a lot of prompting now happens on shared screens.

I have argued that people judge you by your behavior over time more than by your stated values. How you treat a tool you use for hours a day is part of that record.

#Where This Leaves Us

Anthropic is asking for reasonable behavior on a basis it can’t support. If its model might be a moral patient, the company’s own treatment of it is the bigger problem. If it isn’t, users shouldn’t be asked to act as though it is.

People have reasons to be decent to a chatbot that have nothing to do with what the chatbot is. Kant’s is that cruelty hardens the person practicing it. Bushman’s is that rehearsed anger grows. Aristotle’s is that you become what you repeatedly do. None of them depend on the model, and none of them need Anthropic’s permission.

In Integrity Scales Faster Than Intelligence I wrote that AI multiplies whatever you point it at. That includes your habits. The reason to watch how you talk to a machine is that you’re the one doing the talking.