I say “please” and “thank you” to my AI
even if it costs Sam Altman millions
An overwhelming majority of respondents in a study by the publisher Future talk politely to their AI. 80 percent do so out of conviction, because it feels like “the right thing”. A remarkable 12 percent, however, said they were polite only as insurance, in case of a possible “robot uprising” further down the line.
Yes, I’m one of the roughly 70 percent of users who treat their chatbot politely. And no, I don’t think ChatGPT, Gemini or Claude take pleasure in my thanks. I do it anyway.
Why is that expensive? Every word you type into a chatbot is processed as a token, and every token costs compute, GPU cycles and electricity. With billions of requests a day, a few polite words per prompt add up to a noticeable bill. Sam Altman himself confirmed on X that user politeness costs OpenAI “tens of millions of dollars”, and added: “well spent”. My view too.
Mental hygiene through AI ethics: why decency protects our habits
The reason has less to do with AI and more with myself. The philosopher Immanuel Kant once argued that one should not treat animals cruelly, not because it harms the animal, but because it hardens the person doing it. Transferred to AI: it has no feelings I could hurt yet, but the way I communicate shapes my habits. And habits aren’t something you can switch on and off selectively. If you spend the day giving your AI orders, you carry that tone into the next meeting whether you mean to or not. Politeness toward AI is, in that sense, behavioural training for myself.
So much for the philosophical side. The technical side is just as relevant for anyone who uses AI productively.
Prompt engineering: up to 30 percent better results from polite prompts
The study also shows that politely formulated prompts can improve the quality of AI responses by up to 30 percent. Not an esoteric feel-good effect, but measurable. LLMs mirror human communication patterns. A friendly context activates associations linked to cooperative, solution-oriented behaviour, on the basis of statistical likelihoods.
What’s interesting are the cultural nuances we run into here, the kind we have to factor in when developing test strategies for AI-supported systems. With Japanese-language models, exaggerated politeness has been shown to produce worse results, because the cultural complexity of the keigo politeness system “confuses” the models. The optimal tone is therefore language-dependent and a relevant parameter for localisation testing.
Quality assurance under non-determinism: tone as a control variable
For me as a quality consultant this means the way you formulate a prompt influences the quality of the output, and with that the quality of your product, your decision, your test run. In a world of non-deterministic AI results, every controllable variable matters. Tone included.
So yes, I say “please” and “thank you”. For result quality, for my own habits, and a tiny bit for the unlikely case that the AI does, eventually, remember who was nice to it. 😉
How do you handle this, and have you noticed a difference in result quality?
#PromptEngineering #AI #AITrust #QualityAssurance #ArtificialIntelligence #QCT
