Most people I train on AI tools are too polite with them. Saying please and thank you does no harm. The problem is what we leave out to stay polite: clear direction, honest feedback and the full picture of what we need.

We spend our whole lives learning to soften requests, avoid bluntness and accept other people's effort gracefully. Those habits make us good colleagues. With AI, the same habits lower the quality of what comes back.

What the research says about tone

The words of courtesy themselves make only a small difference. A 2024 study from Waseda University and RIKEN tested prompt politeness in English, Chinese and Japanese. Rude prompts often performed worse, and extra politeness did not reliably improve results. The best level of politeness also varied by language.

A 2025 study at Penn State tested GPT-4o on 50 questions, each rewritten in five tones. "Very rude" prompts scored 84.8% and "very polite" prompts scored 80.8%. The authors did not recommend rudeness, and neither do I.

Here is what I take from both studies. Tone moves results a little. What you tell the AI moves them a lot.

Five polite habits that weaken the output

  • Hedged requests. "Could you maybe take a look at this when you get a chance?" is kind to a colleague. To an AI model it gives almost nothing to work with. State the task, the audience, the length and what a good result looks like.
  • Holding back context. Between people, over-explaining can feel rude, as if you doubt the other person. An AI model starts from zero every time. Give it the background you would normally think was too obvious to mention.
  • Accepting the first draft. When a colleague hands us something, we thank them and move on. AI has no feelings to protect. The second and third passes are where most of the quality comes from, so ask for them.
  • Soft feedback. "This is great, maybe one small tweak" tells the model that little needs to change, and it will change little. Name the problem plainly: "The opening is too long. Cut it to two sentences and lead with the cost to the customer."
  • Never asking for disagreement. Language models tend to agree with the person using them. Anthropic researchers documented this pattern, called sycophancy, in 2023. When both sides are being agreeable, you get agreeable, average work. Ask the model to find the three weakest points in your plan before you accept it.

Why this is hard to change

Social norms run deep, and we learned them for good reasons. I lived and worked in Japan for 10 years. That time taught me how much meaning lives in indirect, courteous language between people who share the same context.

That is the catch. Indirect language works because the listener fills in the gaps. An AI model shares none of your context unless you write it down. The gaps stay empty, and the model fills them with safe, generic guesses.

A before and after

Before: "Hi, could you possibly help me write something about our new service? No rush, whatever you think is best. Thanks so much."

After: "Write a 150-word LinkedIn post announcing our new backup service to IT managers at mid-sized firms. Lead with the problem it solves. Use plain language and no hype. Give me two versions with different openings. Thanks."

The second prompt is still courteous. It keeps the thank you. It also tells the model who the reader is, what matters, how long to write and how to give you a choice. That is the difference between a draft you rewrite and a draft you can use.

Be direct, then be as kind as you like

This is the same idea behind directing AI instead of asking it questions. A good film director is courteous to the crew. The same director still says "cut, let's go again, faster this time" as often as the scene needs. Kindness and clear direction work together.

It is also why the AI systems I build start from a knowledge base of 86 interlinked files, co-built with a colleague. The context is written down once, so nobody has to remember to spell it out or feel awkward about over-explaining.

Keep your please and thank you if they feel natural. Drop the hedges, the vague requests and the polite acceptance of work that is not ready yet.

Try this today

Open the last prompt you wrote. Delete every hedge. Add one piece of context you skipped because it felt obvious. Before you accept the result, ask the model what is weakest about it.

You can do this in five minutes, and the difference usually shows in the first reply. I would love to hear how it goes.

Sources: Yin, Wang, Horio, Kawahara and Sekine (2024), Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance. Dobariya and Kumar (2025), Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy. Sharma et al. (2023), Towards Understanding Sycophancy in Language Models.