If you say ‘please’ to an AI, will it get the answer right more often? Politeness can influence how an interaction feels, but tone, instruction quality, and factual accuracy are three different things. Mixing them leads to surprisingly weak experiments.
RUDA AI LAB 003 is a free, six-round game about recognizing flawed comparisons. Each scenario is fictional and written for learning: it is not a transcript of an AI model, a benchmark dataset, or a new numerical result.
한국어판: 한국어로 읽기 →
Play first: politeness, precision, or missing evidence?
Compare two requests and identify the best conclusion. Did the writer change only the tone? Did they also change the task? Or is there too little evidence to tell? Explanations follow each choice, with no account or name required.
Courtesy and clarity are different variables
Compare “Summarize this document” with “Please list three main claims, cite supporting evidence, and label uncertainties.” The second prompt changes task details, structure, and verification requirements. A better-looking answer would not isolate the impact of saying please.
Now compare “Calculate 18 + 24” and “Could you please calculate 18 + 24? Thanks.” The substantive task stays the same. That is a more controlled tone comparison, but no outcome or repeat measurements have been observed just because two inputs were written.
Existing findings are not unanimous
A 2024 ACL Anthology study across English, Chinese, and Japanese found that the effects of politeness varied by language, and maximal politeness was not consistently best. A 2025 arXiv preprint on multiple-choice accuracy for a particular model reported a different pattern. A 2026 multilingual, multi-model preprint emphasized differences across model families, languages, and conversation history. These studies investigate related questions, but their datasets and metrics are not interchangeable.
In particular, a response can sound warmer while being factually wrong, and a concise answer can be correct without sounding personal. Stylistic quality and verified task accuracy should be scored separately. None of the published numbers are measurements made by our game.
How to run a valid experiment
- Control: Preserve task content, model snapshot, language, tools, and context while changing only the courtesy markers.
- Repeat: Use many questions and repeated runs for each condition, recording dates, versions, and settings.
- Blind score: Evaluate answers without exposing the corresponding tone condition; score factual correctness separately from warmth and readability.
- Report uncertainty: State the sample size, denominator, variation, and failures, and avoid generalizing one model’s result to every model or language.
We did not run model comparison trials in this episode. The game score from 0 to 6 reflects answers to our six constructed exercises—not a validated measure of AI proficiency or proof that any style improves accuracy.
Sources and further reading
- Yin et al. (2024), Should We Respect LLMs? — ACL Anthology
- Dobariya & Kumar (2025), Mind Your Tone — arXiv preprint
- Mehta et al. (2026), No Universal Courtesy / PLUM — arXiv preprint
- OpenAI Prompting Guide — clarity, experimentation, and evaluation
Previous and next in the lab
Play PROMPT DUEL — Long Versus Short Prompts. Next we will look at what changes when people evaluate AI answers without knowing which model or condition produced them.
RUDA AI LAB 003 · planned for October 24, 2026; source review October 23, 2026. All game scenarios are fictional teaching examples, not observed model responses. No sponsorship. Choices are designed to stay in the game; ordinary site analytics may still apply.