Does Being Polite to AI Get You Better Answers?
By Chatday Editorial Team ·
Most of us type “please” and “thank you” to a chatbot without thinking about it. It feels polite, and some people quietly suspect it earns them a better answer. So here is the honest version: being polite does not meaningfully improve what an AI gives you, and being rude does not reliably help either. The thing that actually changes the answer is how clear and specific you are.
That is not a guess. Researchers have now run the test, and the results are a little awkward for anyone who thought good manners were the secret.
What the politeness study actually found
In October 2025, two researchers at Penn State ran a neat little experiment. They wrote 50 multiple-choice questions across math, science, and history, then rewrote each one in five tones: Very Polite, Polite, Neutral, Rude, and Very Rude. That gave them 250 versions of the same questions, which they fed to ChatGPT’s GPT-4o model.
The surprise: the blunt prompts did slightly better. Very Polite phrasing landed the right answer 80.8% of the time. Very Rude phrasing hit 84.8%. The curt, no-niceties versions nudged ahead of the gushing ones.
Before you start barking orders at your chatbot, read the fine print. This was one model, 50 base questions, and a gap of about four percentage points. The authors themselves called it a starting point for more research, not advice to be mean. A result this narrow can flip with a different model or a different task.
And it does flip. An earlier cross-lingual study from 2024 found almost the opposite: impolite prompts often made models perform worse, while piling on extra flattery brought no clear benefit. That study also found the sweet spot shifts by language, since politeness norms in English, Chinese, and Japanese are not the same. Two studies, two different headlines. When the research disagrees this much, the sensible read is that tone is mostly noise.
So should you be rude to AI? No
Short answer, no, and not only because it is unpleasant to type.
The measured upside of rudeness is tiny and unreliable. The downside is real. Curt, aggressive prompts tend to strip out the context a model needs, because you are too busy being terse to explain what you actually want. You also train a habit. If you spend all day snapping at a chatbot, some of that tone leaks into how you write to actual people. Not worth four percentage points on a trivia quiz.
There is also a cost angle that went viral last year. When someone asked how much money all those extra “please” and “thank you” messages were burning in electricity, OpenAI’s Sam Altman replied that it was “tens of millions of dollars well spent.” It was a joke, mostly. The point stands though: politeness is not free for the companies running these models, and it is not buying you smarter answers either. It is just a habit, and a harmless one.
What actually changes the answer
If tone is noise, what is signal? Specificity. A vague question gets a vague answer no matter how sweetly you ask it. Here is the honest ranking of what moves the needle:
| What you tweak | Real effect on the answer |
|---|---|
| Saying please / thank you | Almost none |
| Being rude or demanding | Tiny, and not worth it |
| Adding details and context | Large |
| Stating the format you want | Large |
| Giving an example of a good answer | Large |
| Choosing the right model for the task | Large |
Look at the bottom four rows. They have nothing to do with manners and everything to do with being a clear communicator. This is the same reason the same question can return different answers depending on how you word it. Tone is the decoration. The request underneath is what the model works with.
A quick before-and-after makes it obvious. “Write me a nice email” is polite and useless. “Write a short, friendly email to my landlord asking to fix a leaking tap this week, keep it under 80 words” is specific and gets you something you can actually send. No “please” required, though you are welcome to add one.
The habit that beats good manners
Here is the part that turns this into something useful rather than trivia. Once you stop worrying about tone and start writing clear prompts, the next biggest lever is which model you send that prompt to. The same clear question can get a sharper answer from one model and a flatter one from another, depending on the task.
That is easy to see for yourself. Take one well-written prompt and run it through a few models side by side. For a careful reasoning task you might prefer one; for a quick, breezy reply you might prefer another. The tone you used barely registers. The model you picked, and how clearly you asked, decide almost everything.
You do not need five separate apps and five logins to do this. One place, every major model, one clear question.
So keep your manners if you like them. There is nothing wrong with a “thanks” at the end of a good answer. Just know that it is for you, not for the machine. The real gains come from asking clearly and sending that question to the right model.