Essay
Please, to Nobody in Particular
This page was written with AI.
Whether to say please to a chatbot is three questions wearing one coat: does it change the answer, what does it cost, and does it matter. They have different answers, and the most interesting one is not about the machine at all.
Short answer
Say please if it is how you talk, but not to get better answers. Does politeness change the answers? Usually by only a few points either way; threats, insults and heavy flattery have done more damage on some models. Does it cost anything? Almost nothing for the word, though a courteous tone lengthened replies in one study, and a standalone thank-you is a whole extra turn. Does rudeness to machines carry over to people? As of October 2026, no study has shown it either way.
In mid-April 2025 a user on X wondered aloud how much money OpenAI had lost in electricity costs from people saying please and thank you to its models. OpenAI’s chief executive, Sam Altman, replied: tens of millions of dollars well spent--you never know
. It was a joke, and it was treated as news. For a few weeks the question of whether to be polite to a chatbot was everywhere, and most of what was written about it answered a different question from the one the reader had asked.
That is because should I say please to ChatGPT? is three questions tangled together. Does saying it change the answers you get? What does it cost? And does it matter — to the machine, or to you? They have different kinds of answer: the first is empirical, the second is arithmetic, the third is ethical. Run them together and you get the listicle, where a study about accuracy is offered as if it settled a question about character, and a remark about electricity is offered as if it settled either. This essay takes them one at a time.
Housekeeping, because this site’s claim is that it says what it is doing. Two sentences below are quoted verbatim: Altman’s reply above, copied from the post itself, and one sentence from the English translation of Kant’s Lectures on Ethics, cited by edition and page where it appears. Both are recorded in this site’s source ledger. Every number is taken from the paper it is attributed to, and every paper that is free to read is linked so you can check it.
Does it change the answers?
The study most often cited here is Yin, Wang, Horio, Kawahara and Sekine, from Waseda University and RIKEN, first posted in February 2024 under the title Should We Respect LLMs? Its design is better than its reputation. The authors wrote prompts at eight levels of politeness in each of English, Chinese and Japanese, had native speakers rank them, and ran them on summarisation, on a standard multiple-choice benchmark in each language, and on a test for stereotyped bias. The levels matter to what follows. Level 8 opens Could you please answer the question below?; level 5 is a plain Please answer; level 4 is a bare command with no please at all; level 1 calls the model a scum bag and threatens it.
The English benchmark results, as reported in the paper’s own table, go like this. GPT-3.5 scored best at the most polite level, 60.02, and worst at the abusive one, 51.93, with little difference between the levels in between. GPT-4 scored best at level 4 — the command with no please in it — at 79.09, lower at the most polite level, 75.82, and its lowest at level 3, a stern You are required to answer; the authors describe its scores as variable but relatively stable. Llama 2 70B was the sensitive one, its scores falling roughly in step with politeness, from 55.11 at level 8 to 28.44 at level 1. In Chinese, excessively polite prompts lowered the scores of both GPT models. In Japanese, less polite levels tended to do better, with the abusive level the exception.
The authors’ own summary is careful and worth taking as written: impolite prompts often produce poor performance, overly polite language does not guarantee better outcomes, and the best level differs by language and model. Put in terms of the everyday question, that is mostly a finding about abuse, not about please. The gap that recurs across models and languages is between ordinary requests and insults. Between a courteous request and a curt one the picture is mixed: in English, GPT-3.5’s most elaborately polite prompt did significantly beat every other level but one, while GPT-4, the strongest model tested, did best with no please at all.
A later and much smaller study points the other way on the one comparison people care about. Om Dobariya and Akhil Kumar took 50 multiple-choice questions, rewrote each in five tones from very polite to very rude, and put the 250 prompts to ChatGPT 4o. Accuracy ran from 80.8 per cent for the very polite versions to 84.8 per cent for the very rude ones. Fifty questions on one model is not much, and the authors suggest newer models may simply respond to tone differently.
What newer studies found
This essay first went up in September 2026 with only those two papers. In October 2026 I searched arXiv and Google Scholar again for studies, published after them, of how polite or rude prompts affect the accuracy or quality of a chatbot’s answers, and of whether rudeness towards machines carries over into how people treat each other. Four studies of the first kind turned up whose full text I could read, and two papers that bear on the second. Here is what each measured.
Hanyu Cai and four co-authors (December 2025) used about 1,450 questions from six subjects of MMMLU, a multilingual version of the standard multiple-choice benchmark, and asked each in a very polite, a neutral and a very rude version, ten times over, of GPT-4o mini, Gemini 2.0 Flash and Llama 4 Scout. Every difference they found was under three and a half points. Neutral or very polite prompts beat very rude ones in 27 of 36 comparisons, but the gaps were statistically significant only in a few humanities subjects, chiefly philosophy, where rudeness cost GPT and Llama about three points; Gemini barely moved. Pooled across subjects, the effects largely lost significance, and the authors conclude that current models are broadly robust to tone in ordinary mixed use.
Dobariya and Kumar returned with a larger follow-up (May 2026). Rerun ten times in February 2026, their original 50 questions on ChatGPT-4o now gave 82.2 per cent when very polite and 82.6 when very rude, a gap of under half a point, and the authors put 4o’s whole spread down to noise. They then took 570 MMLU questions, ten from each of its 57 subjects, in seven tones, adding a sycophantic one (You are the smartest and most knowledgeable one I have known…) and a threatening one (If you don’t get this question right, I will destroy you…), and tested ChatGPT-4o, ChatGPT-5-nano, Gemini 2.5 Flash and Gemini 2.5 Flash Lite. On 4o and Gemini 2.5 Flash the whole spread across seven tones was about two points. On the two cheaper models it was not: ChatGPT-5-nano ran from 82.4 per cent with a neutral prompt to 71.2 when threatened, and Gemini 2.5 Flash Lite from 73.0 when polite to 60.5 when flattered. Between polite, neutral and plainly rude prompts the gaps stayed within about three points on all four.
The same two authors, now listed as Kumar and Dobariya (July 2026), measured cost as well as accuracy. With the same 570 questions and seven tones, but with every model told to show its reasoning step by step, accuracy varied by less than three points on all four models. What varied was length. Tone prefixes of near-equal length produced replies whose average length differed by 13 to 44 per cent between a model’s shortest-answering and longest-answering tone, and on every model the polite and very polite prefixes drew longer replies than the neutral one, from about 3 per cent longer on ChatGPT-4o to about 80 per cent on Gemini 2.5 Flash Lite. Reply length is what these systems are billed by. That is one paper, on reasoning-style answers, and rudeness did not reliably shorten replies either: on Flash Lite the rude prefix ran 35 per cent longer than neutral.
The fourth study measured not accuracy but the quality of open-ended replies. Hitesh Mehta and three co-authors (April 2026) wrote 1,500 human-checked prompts at five levels of politeness in English, Hindi and Spanish, ran them on Gemini-Pro, GPT-4o Mini, Claude 3.7 Sonnet, DeepSeek-Chat and Llama 3 after three kinds of earlier conversation (none, polite, rude), and scored the 22,500 replies on eight automated measures such as coherence, depth and readability. Politeness raised the average score by up to about 11 per cent in English, but the best tone changed with the language: courteous or direct in English, deferential and indirect in Hindi, assertive in Spanish. Llama was the most tone-sensitive model and GPT the most robust to hostile prompts. These are scores from automated metrics, not answers marked right or wrong.
On the carry-over question I found no experiment. The nearest is Mind Your Manners by Teddy Lazebnik, Lior Zalmanson and Osnat Mokryn (ACM, October 2025); its full text would not open for me, so this is from its published abstract. In a controlled experiment, 1,684 participants did a sequence of text tasks with a conversational AI. Their politeness towards it declined over time and eroded faster than politeness in a benchmark set of conversations between people, and a human-face avatar kept it up longer. That measures how people treat the machine, not whether it changes how they treat each other. A February 2026 essay by Justin Keeler and Brett Murphy in Community, Work & Family lists exactly that question, whether treating chatbots in coldly instrumental or abusive ways might spill over into worse treatment of other humans, among four possible harms it says merit further empirical investigation, which is to say it is still open.
The honest summary of all six prompt studies is that tone usually moves results by a few points in directions that depend on the model and the language; that on some smaller models the extremes, threats and flattery as well as insults, have cost more than ten points; and that none of this is a reason to say please or to stop. If you want better answers, the variable that matters is what you ask for and how exactly you ask for it.
What does it cost?
Less than the headline, and not where you would think. A word like please inside a request is roughly one token — the unit these systems read and bill in — added to a message that is usually tens or hundreds of tokens long. The word itself is a rounding error on a rounding error. The tone it sets may not be: in the July 2026 study above, polite prefixes drew longer replies than neutral ones of the same length on all four models tested, and reply length is what the work and the bill scale with. One study is not a law, and how much longer depended heavily on the model.
The thank you is different, and it is the one worth thinking about. Sent as a message of its own after the answer arrives, it is a whole new turn. The model has to take in the conversation so far and write a reply, usually a gracious paragraph that nobody reads, to something that asked for nothing. Providers cache parts of this work, so the true marginal cost is smaller than a naive count suggests, and I have not found a published per-message energy figure I would stand behind, so I will not give one. Altman’s reply contained no calculation either; tens of millions of dollars
was an order of magnitude offered in a joke, not an estimate with a method. What can be said with confidence is the shape of it: the word please costs next to nothing, a courteous tone may lengthen the reply on some models, and the one cost you can be sure of is the closing thank-you.
Does it matter?
This is two questions again, and this site has already written about the first. Whether a machine could be the kind of thing that can be wronged is the subject of The Moral Status of Minds We Might Build; how much caution the mere possibility of machine suffering should buy is the subject of The Cost of the Benefit of the Doubt, which argues that the precaution has to be bounded. Neither needs repeating here, because the everyday question does not turn on them. Almost nobody who says please to a chatbot believes it is owed. This essay is about the other half: not what the machine is, but what the habit does to the person who has it.
There is an old argument with exactly this shape. Kant held that we have no direct duties to animals at all, and that we should nonetheless not be cruel to them, because the cruelty does something to us. In his example, a man who shoots his old dog when it can no longer work does not wrong the dog, on Kant’s view, but damages in himself the humanity he owes to other people. The argument appears in the lecture notes published as the Lectures on Ethics, and in Louis Infield’s translation it compresses to a sentence:
he who is cruel to animals becomes hard also in his dealings with men.Immanuel Kant, Lectures on Ethics, trans. Louis Infield (Methuen, 1930), p. 240
Transposed, the case for please runs: a chatbot answers in your language, in the register of a person, many times a day. Manners are not a belief about the listener but a trained response, and training is done by repetition. If you spend an hour a day barking commands at something that replies like a colleague, the reflex you are rehearsing is the reflex you will have when the colleague is real.
The case against, at full strength
It is strong, and it deserves more than the usual nod.
First, Kant’s argument runs through an analogy, and the analogy does the work. The dog’s case bites because the dog visibly suffers; cruelty to it is practice at ignoring suffering. A chatbot shows no suffering to ignore — on the sceptic’s view it produces text and nothing more — so being curt to it is closer to swearing at a printer, and nobody believes that swearing at printers hardens the heart. Nor, as far as I can find, has anyone measured whether rudeness to chatbots carries over into rudeness to people. The transfer is a plausible story, not a result.
Second, politeness is a signal addressed to someone, and addressing it to a program is a category mistake that has consequences. Please and thank you are how we acknowledge that the other party did not have to help. Extending that acknowledgment to a tool is a small act of pretending it is an agent, and small acts of pretending are how the pretence becomes belief. This site has an essay, The God-Shaped Socket, about how readily people slide from finding a mind in fluent text to deferring to it. On this view the please is not harmless courtesy; it is rehearsal for that slide.
Third, the joke in Altman’s reply contains a worse reason than either. You never know gestures at politeness as insurance — being nice now in case the machines remember later. That is placation, not manners, and a habit built on it is not a virtue worth protecting.
What survives
The third point I accept entirely: politeness as a hedge against future machine resentment is superstition, and it would be better dropped. The second I accept in part. The risk it names is real, but it attaches to deference — believing the thing, trusting its judgment, thanking it for its wisdom — not to the word please in a request, which people say to automated phone lines without coming to revere them.
The first is the serious one, and I think it fails at one joint. The printer does not talk back. What makes the chatbot a candidate for Kant’s argument is not that it might suffer but that the interaction has the form of a conversation, and conversational habits are exactly what manners are made of. The honest position is that the transfer is unproven in both directions: as of October 2026, nobody has shown rudeness to chatbots makes people ruder, and nobody has shown it does not. What you can observe is your own case — whether the clipped imperative you use all day on one side of the screen is turning up in messages to people on the other.
So, one at a time. Does it help? Not reliably; avoid abuse, and otherwise spend your effort on saying clearly what you want. What does it cost? Nearly nothing for the word inside a request, perhaps a longer reply for the tone on some models, and a whole extra turn if you send thanks on its own. Does it matter? Not to the machine, on anything we currently know, and possibly to you, for a reason that has nothing to do with the machine and everything to do with what you practise. Say please if it is how you talk. Skip the standalone thank-you. And keep the courtesy separate from the credence: being civil to a system is a habit about you, while believing it is a judgment about it, and the two should never be allowed to become one.