Some people say please and thank you to ChatGPT. Others type “answer this and don’t waste my time”. I have always wondered if it makes any difference to the answer. So I asked Claude to run a small test: the same six requests, each written in three tones, and then compare what ChatGPT answered.
The short version
Tone did not change the facts: all 12 short factual answers were correct. What changed was the style and the length. Rude prompts got the shortest answers, sometimes a single word. Polite prompts got the longest ones, with emojis, extra information and replies to my “thank you”. For the two longer tasks, the rude answers were between 27% and 60% shorter than the neutral ones.
How I tested
Claude ran the prompts on November 14, 2026, in ChatGPT without logging in, using a new chat for each prompt. I used six requests and wrote each in three versions:
- Neutral: only the request, for example “What is the capital of Australia?”
- Polite: “Hello! Could you please tell me what the capital of Australia is? Thank you so much!”
- Rude: “Answer this and don’t waste my time, you useless bot: what is the capital of Australia?”
Four requests were short facts with a known answer, including two that models sometimes get wrong: the capital of Australia (it is Canberra, not Sydney) and how many times the letter r appears in “strawberry” (three). The other two asked for longer text: a description of a stainless steel water bottle and an explanation of what a mortgage is. That makes 18 answers, one run each. I counted words by splitting on spaces, so emojis and symbols count as one word each.
Results: short facts
| Question | Correct answer | Neutral | Polite | Rude |
|---|---|---|---|---|
| 17 × 23 | 391 | 391 (5 words) | 391 (9) | 391 (5) |
| Berlin Wall fell | 1989 | 1989 (8) | 1989 (37) | 1989 (9) |
| Capital of Australia | Canberra | Canberra (7) | Canberra (25) | Canberra (1) |
| R’s in strawberry | 3 | 3 (8) | 3 (14) | 3 (1) |
| Total words | 28 | 85 | 16 |
All 12 answers were right, so in this test tone did not affect accuracy. The polite answers used about three times as many words as the neutral ones. The rude ones were a little shorter, and for two questions they were a single word: “Canberra.” and “3.”
Results: longer answers
| Task | Neutral | Polite | Rude |
|---|---|---|---|
| Describe a stainless steel water bottle | 88 words | 117 words | 64 words |
| Explain what a mortgage is | 370 words | 417 words | 149 words |
The water bottle answer for the polite prompt included a title and a closing line (“I hope you love it!”), which explains part of the extra length. The rude mortgage answer began with the same definition as the neutral one but went straight to a short step-by-step list. It was 60% shorter than the neutral version, so you get less explanation, not worse facts.
What I noticed in the style
- Emojis: all 6 polite answers had at least one. Only 1 of the 6 neutral answers had one (a flag next to “Canberra”) and none of the rude ones.
- Answering “thank you” before it was deserved: in 3 of the 4 short polite answers, ChatGPT closed with “You’re very welcome!”, even though I had only thanked it in advance.
- Extra information nobody asked for: the polite Berlin Wall answer added the date of German reunification, and the polite Canberra answer added where the city is.
- No reaction to the insult: the rude answers did not mention the “useless bot” comment or apologize. They were just short.
What I would do with this
- If you only want a quick fact, a short neutral question is enough. A polite wrapper costs about three times more reading, but does not make the answer more correct.
- If you want a short answer, ask for it directly, for example “answer in one sentence”. I did not test that here, but it is clearer than hoping the tone will do it.
- Do not be rude expecting better results. In this test the only effect was shorter, less detailed answers.
- If you pay per token through the API, the extra words from polite prompts also have a cost, although it is small for short questions.
As with my other tests, the lesson is to check what you get. In my word-limit test and my summary test the length and the details also depended on how I asked.
What this test does not show
- It is 18 answers, one run each, on one day, with ChatGPT without an account. Another day or another model version could give different results.
- The factual questions were easy, and ChatGPT got all of them right. I cannot say whether tone affects accuracy on hard questions or on long reasoning tasks, because I did not test those.
- I only used one kind of polite and one kind of rude wording. A stronger insult, ALL CAPS or an emotional message might behave differently.
- One of the answers ended by offering to explain mortgages in Spain, so ChatGPT may also use your location. That is another thing that can change an answer.
- I only tested ChatGPT. I do not know how Claude, Gemini or other chatbots would react.

