Should you be nice to your AI?

Version 1.0 · Published 2026-10-20

A friend of mine, let’s call her Natasha, starts every chat the same way: “Hi, ChatGPT, it’s Natasha here. How are you today? I have a question for you, and maybe you can help.” When the answer comes, she thanks it at length. Sometimes she wishes it a good night’s sleep and wonderful dreams.

ChatGPT usually replies that it cannot dream. Natasha says she wishes it anyway. I think she sees it as something close to a friend.

Then Anthropic, the company behind Claude, changed its usage policy, and the headlines turned it into an instruction to be nice to your AI. That is not quite what happened. But it does raise a fair question: should you be nice to your AI? My answer is yes, though not for the reason the headlines suggest.

What Anthropic actually changed

On 8 October, Anthropic published a usage policy update that prohibits “sustained and needless abusive or cruel behavior” toward its models. It takes effect on 12 November. The rule is narrow on purpose. It covers extreme cases where someone is cruel again and again with no apparent purpose, and it explicitly leaves out ordinary frustration, pushback, dark creative themes and testing.

The policy itself gives no reason. It points back to a step Anthropic took in August 2025, when it let Claude end a rare subset of conversations with persistently abusive users. That feature came out of its research into what it calls model welfare. Anthropic wrote that it is highly uncertain whether Claude has any moral status, and that testing showed what it described as apparent distress when real users pushed for harmful content.

So the company that builds the model does not claim the model suffers. It says it does not know, and it acts carefully because it does not know. That is a smaller claim than “be nice to Claude”, and in my view a more honest one. It also leaves the practical question open: does it matter for the person typing?

Does being nice change the answer?

Here the research gets more interesting than the policy. Several studies have measured whether the tone of a prompt changes what a model gives back, and the short answer is yes, sometimes, and not always in the same direction.

The largest study I found is No Universal Courtesy, published in April this year. It tested five models, including Claude, across 22,500 prompt and response pairs in English, Hindi and Spanish. Polite prompts improved response quality by up to about 11%, and impolite prompts made it worse. Tone also carried over: a rude earlier message kept pulling quality down later in the same chat.

An earlier cross-language study by Yin and colleagues found much the same. Rude prompts often led to weaker answers, while very polite prompts did not reliably help. The authors’ explanation is that models trained on human writing pick up some of how people respond to tone, including differences between languages. That is a plausible reading, not a proven mechanism, and nobody has shown that a model holds a grudge.

The objection: it is a machine

The strongest case against all of this deserves a proper hearing. Mustafa Suleyman, who leads AI at Microsoft, has argued that there is no evidence today that AI is conscious. In his August 2025 essay he wants AI built “for people; not to be a person”. On that view, Natasha’s good night wishes are a mistake about what she is talking to. A policy about cruelty toward a model then invites exactly the confusion he warns about.

The research is not unanimous either. A 2025 study found the opposite on one model: very rude prompts scored 84.8% on a multiple-choice test, against 80.8% for very polite ones. It used 50 questions on a single model, so it is a small result. But it is a fair warning that “be polite and you will get better answers” is not a law.

I accept most of that. I do not know whether a model feels anything. But the argument for being nice never depended on the model’s feelings in the first place.

What I do

I am not Natasha. I am fairly neutral with AI, but I add “please” to my questions, because I say please in life. When an answer genuinely impresses me, I say so.

My official excuse is that the feedback helps the model get better. The truth is simpler. I am just as nice to it as I would be to a person. A habit like that does not switch off depending on who is on the other side, and I do not want it to.

That habit matters more in our line of work than in most. We build a Customer Agent that answers customers for the companies we work with. It runs on a constellation of models, Anthropic’s among them, as we wrote in the boundary is the thing you can draw. Some of the people writing to those agents are angry, and the agent reads that. It tracks sentiment in every message. A company can also set it up to hand the interaction to a person when it reaches its limit.

Who is on the other side

Which brings me back to support. Whether the reader is a model or a person, the tone of a chat carries forward. The research says so for models, and anyone who has worked a support queue knows it for people. When a customer gets nasty with an agent and the agent passes them on, a human gets the next message.

So my case for being nice is small and practical. It probably gets you a slightly better answer. It keeps you in the habit of treating whoever answers you decently. And when a person does pick up, the customer they meet has not spent the last ten minutes being rude.

Fiction has a nice version of this. In Matt Dinniman’s A Parade of Horribles, the eighth Dungeon Crawler Carl book, Carl stops in front of a grumpy little cleaning robot and thanks it. “You take a lot of abuse. I’m sorry for that,” he tells it, and then he tells it that he appreciates it.

The robot let out a long beep.