Should you be polite to AI? Here’s what the research says

It’s not as simple as you might think. 

Decades ago, while serving in the US Navy, I taught a class on military communications. My classroom was a grassy field in the center of a temporary encampment and my students were first-year medical students serving on active duty. 

The purpose of my lesson was to teach future doctors how to communicate efficiently during high-stress situations. I started every class with the same spiel: 

“Hello. Repeat after me. Please, thank you, and you’re welcome” Once the students finished repeating what I’d said, I’d say “Good. That’s the last time I want to hear any of you say those words during this exercise.” 

Polite isn’t always right

I’m not against manners. In most circumstances, being polite speaks highly to a person’s character. It demonstrates empathy and a desire to ingratiate oneself to others. 

But when lives are on the line, and every second counts, words that don’t convey important information waste valuable time. Please and thank you aren’t worth dying or losing a patient over.

This is why you don’t see surgeons on TV saying “nurse, may I please have the scalpel. Thank you. Now, may I please have the sponge. Thank you, I appreciate it.” That kind of rhetoric may be important in most aspects of our daily lives, but it isn’t useful in the operating room or on the battlefield.

In the world of artificial intelligence, scientists have often pondered to what extent manners affect the outputs of large language models (LLMs). The current consensus seems to indicate that models tend to generate stronger outputs when prompts contain polite language. 

A preprint paper from researchers Waseda University and Riken AIP, published in 2024, found that “impolite prompts often result in poor performance, but overly polite language does not guarantee better outcomes.”

Microsoft’s Kurtis Beavers, a director on the design team for Microsoft Copilot, said that “using polite language [in prompts] sets a tone for the response.” According to them, the AI model is more likely to respond in kind when prompted politely. 

In the same blog post, Microsoft suggests that “rather than order your chatbot around,” we should all start our prompts with “please,” and we should “say thank you when it responds, and be sure to tell it you appreciate the help.” Because “doing so not only ensures you get the same graciousness in return, but it also improves the AI’s responsiveness and performance.”

That’s not necessarily a good thing

If the basic idea is to get the most utility out of chatbots, it seems counterintuitive to train them to waste tokens and time outputting polite language. 

According to Microsoft, when we use manners in our prompts we’re training the machines to respond in kind. This indicates that people with limited tokens are getting short-changed because subscribers with more tokens want AI models to say please and thank you.

The next time you get a response from a chatbot that says “you’ve run out of tokens, please wait a little while and try again,” you’ll know exactly who you should say “thank you” to. 

Food for thought: Is it better to spend the extra electricity, training time, and tokens to make AI models say “please” and “thank you,” or would you rather see those resources used to make models more capable at useful tasks? 

Read more: Experts are far more positive about AI than the public

Art by Nicole Greene