It seems that kindness no longer pays, at least not with artificial intelligence. A recent study by Penn State University, with the very serious title “Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy,” found that ‘rude’ and “very rude” prompts get more accurate responses than “polite” ones. The authors tested 250 versions of identical questions but phrased in different tones, from “Please, could you help me solve this math problem?” to “Solve it immediately, without unnecessary explanations.” The result? The “very rude” requests achieved an accuracy of 84.8%, compared to 80.8% for the kinder prompts. A four-point advantage for those who lost their patience.
When rudeness becomes efficiency
Researchers hypothesize that the latest models interpret direct tones as signals of urgency and clarity, favoring precision over form. Conversely, polite tones activate the part of the model trained to “write like a human”: more context, more nuance, more words.
In other words, being polite makes AI more conversational but less focused on the goal. Being direct, or even a little blunt, pushes it toward more concise and targeted responses. It’s a kind of digital paradox: to achieve more human-like behavior, you have to stop talking like a human.
The courtesy that “consumes”
There is another consideration, more concrete and less intuitive. Every extra word in a prompt, even an innocent “hello,” “please,” or “thank you,” involves a minimal amount of extra computation, which multiplied by billions of requests can become significant.
This is not just speculation: Sam Altman himself has admitted that the constant use of “please” and “thank you” in requests to ChatGPT adds “tens of millions of dollars” to OpenAI’s energy costs.
This does not mean that we should no longer be polite, but that in a system where even a “please” has an impact, albeit modest, we should begin to think that linguistic form is no longer so innocuous, but rather has a measurable impact.
But is politeness always noise?
While rudeness seems to improve technical accuracy, there are contexts in which kindness and more nuanced language are not only useful but become essential in guiding artificial intelligence toward the desired result. The mistake lies in thinking that every interaction with a language model is a simple request for data.
In creative tasks, for example, how matters more than what. If we ask AI to write a poem, a screenplay, or an empathetic text, the tone of the prompt becomes an integral part of the instruction. A direct command such as “Write a poem about rain” will likely produce a correct but neutral text. If, on the other hand, we write, “I would like you to compose a short, somewhat melancholic poem about the feeling of rain after a difficult day, with an intimate and reflective tone, please,” those extra words are not noise: they are qualitative data that shape the style, emotion, and voice of the result.
Between efficiency and humanity
The question, then, is less trivial than it seems: do we want machines to adapt to the way we speak, or do we want them to gradually accustom us to speaking like them? Perhaps, in the relationship between people and artificial intelligence, the challenge is not deciding whether to be polite or not. It is learning when politeness is needed and when it is just noise. And perhaps reminding ourselves that if we have to raise our voices to get a precise answer, the problem is not just with the model.
Keyformat observes how AI is transforming language, habits, and everyday gestures. Even those that seem the most harmless, such as saying “please.” Because digital sustainability also depends on this: on how we learn to talk to machines and, perhaps, a little bit, to ourselves.
