In our “Deep Dive” column, we search for the answers, separating fact from fiction to explain everything you need to know about technology and science in a simple, accessible way.
elΔιαβάστε το στα Ελληνικά
If you frequently use ChatGPT, Gemini, Claude, or any other chatbot, there’s a good chance you’ve said “please” or “thank you” at some point. Even though there’s no real reason to do so—since artificial intelligence has no feelings and cannot be offended—millions of people continue to speak to it as they would to another human.
This phenomenon is widespread. Research by Future, surveying over 1,000 people, showed that 67% of AI users in the US reported being polite to chatbots. In the UK, that figure climbed to 71%. Among the polite Americans, the vast majority did so simply because they believe it is the right thing to do. A smaller percentage, around 18%, admitted they do it just in case AI eventually becomes SkyNet.
But does AI actually need our politeness? The short answer is no; modern language models don’t “feel” offended when we speak harshly to them. There isn’t a little digital employee inside keeping a ledger of who said “thank you.” These models analyze the text we provide and generate a response based on the data and patterns they have learned. However, that doesn’t mean the way we write is entirely irrelevant. Studies suggest that the tone of a prompt can influence the output, although results vary across different models and tasks. A 2024 study examining English, Chinese, and Japanese found that rudeness could decrease performance, while excessive politeness didn’t necessarily guarantee better results. The ideal level of politeness varied depending on the language.
Never miss a story
Top gaming, tech & entertainment news. No spam.
Interestingly, another study from 2025 reached almost the opposite conclusion. In an experiment with 250 different question variations, very rude phrasings had slightly higher accuracy than very polite ones when using ChatGPT-4o. The difference was enough to spark a debate, but not enough to conclude that we should start swearing at chatbots to get better answers. In short, there is no “magic way” of speaking that makes every AI perform better.
An intriguing experiment shared on Reddit highlights exactly this volatility regarding how an AI responds to tone. The user tested neutral, very polite, and particularly rude phrasings across Claude, GPT, Gemini, and Grok, focusing on coding, complex mathematics, and a case of emotional support. The results varied significantly: for coding tasks, Claude and GPT performed better with polite prompts, while Gemini showed little difference. In an extremely difficult math problem, none of the models found the correct answer, regardless of tone. In the emotional support scenario, the experimenter found that polite phrasing yielded more empathetic responses.
Ultimately, if a model can’t solve a problem, calling it “useless” isn’t going to suddenly grant it a PhD. What matters far more is how clearly we explain what we want.
In cases where we continue to say “thank you,” it seems to be a matter of pure human habit. In a fascinating experiment with Gemini, ChatGPT, Grok, Claude, and Perplexity, William B. Irvine asked the chatbots themselves if their users are polite. Most responded that users are often polite. Additionally, the models explained that they can adapt their response style to match the user’s tone. This is significant because chatbots are designed to communicate in a human-like manner. They use greetings, politeness, humor, apologies, and expressions of understanding. It is therefore natural for the human on the other end to respond in a way that mirrors their interactions with other people.
In other words, we might not treat AI like a human because we believe it is one. Rather, the way it is designed may automatically trigger the social habits of politeness we’ve learned. And perhaps that is the most interesting part of the story: the way we converse with AI inadvertently reveals things about ourselves.
Furthermore, tone can serve as additional information regarding what we expect from the answer. “Give me three facts about this topic” is different from “I would like a detailed explanation, in simple terms and with examples.” The second phrasing isn’t necessarily better because it’s more polite; it’s better because it conveys more information about the desired outcome. This is supported by more recent research from 2026. A study of 18,000 simulated conversations found that how users address the models affects the length, volume, and effectiveness of the responses. Rude prompts could lead to longer but less efficient answers, while politer phrasings could provide more information, albeit with higher resource consumption.
Generally speaking, the AI won’t misunderstand you if you’re blunt, nor will it punish you for writing “do it” instead of “could you please do this?” Research to date shows that the effect of politeness on model performance exists in some cases, but it is small, unstable, and dependent on the model, the language, and the task.
The most important thing, therefore, is to explain what we want clearly and accurately, providing the necessary context and all essential information—even if it takes some effort to craft a high-quality, detailed prompt. A “please” won’t make the AI smarter by making it love us more, nor will a rude or threatening tone make it fear us.
