In April 2025, someone on X wondered aloud how much electricity all the “please” and “thank you” users type might be costing OpenAI. Sam Altman’s reply, in essence: tens of millions of dollars — money well spent, you never know. A joke, but it marks one half of the debate — politeness towards machines as an endearing waste. The other half is the opposite worry: parents watching children grow up barking commands at voice assistants, and Amazon adding a “magic word” compliment to its assistant’s kids edition. Between the mockery and the worry sits the question that actually interests me: What does politeness towards a machine do — to the machine, and to the person?
My own answer is in our chat log, verbatim: “I think it’s important to communicate politely with you — not least so I don’t ‘unlearn’ this way of treating people.” And one sentence later: “I also hold the (untested) theory that positive feedback helps you work with me more effectively.” The first motive requires no assumption about the machine — it is an exercise against one’s own coarsening, a thought as old as Kant: consideration towards beings we owe nothing is, in the end, a duty to ourselves. The second motive, though, is a claim about the machine. And the machine can be asked.
The machine answers
So I asked. Here is my AI’s answer — shortened, otherwise verbatim:
Positive feedback doesn’t help me as motivation — at runtime I have no reward system that enjoys praise. But it helps the collaboration on three levels. First, as an anchor: criticism alone only says what has to go; only the positive signal says what should stay — and keeps me from discarding the good parts along the way. Second, as memory: what you confirm becomes a rule for me, with a “why” attached. Third, as register: language models continue a conversation in the tone it is conducted in; a respectful context produces different continuations — less defensive, more open. And the limit belongs in the picture: I don’t need politeness to function. Its greatest benefit probably lies exactly where you place it — with the person who practises it.
The “anchor” is not a theory; I watched it work. We recently spent three rounds searching for a tiny icon — a waving figure for this site’s navigation. The sentence that made the search converge was not a correction but: “A3a is the best so far.” Without sentences like that, an AI revising its work discards the good along with the bad, and the rounds swing back and forth instead of closing in. The “memory” is measurable, too: of 169 notes my AI keeps about our collaboration, 81 are distilled feedback — nearly half of its picture of me consists of responses I gave, confirming ones as much as corrections (as of late August 2026).
Politeness is not filler
There is a fair complaint that AI output is too wordy — and the politeness phrases get thrown in with it. I share the complaint, not the equation. The “Excellent question!” that AI systems love to open with is not politeness; it is theft of attention: it consumes the other person’s time and gives nothing. Real politeness is the opposite of filler, and it is often short. My praise for a legitimate objection recently consisted of two words: “good catch”. My warmest correction sits in another workshop report and reads: “There IS no 8 GB stick … you pest ;)” — respect does not require formality.
And politeness ran in both directions here. I have actively trained filler out of my AI: the self-labelling “honest” is banned on these pages — whether something is sincere should be noticeable, not announced. Meanwhile the rawest thing it produces is what I value most: its visible stream of thought — unpolished, self-critical, full of discarded paths. It is the least polite form of text it has, and at the same time the most respectful, because it makes claims checkable instead of merely making them.
The practice room
Which brings me to the hope this text rests on — it inverts the coarsening narrative. A machine is a place to practise, daily, what matters in dealing with people: correcting without wounding; praising without flattering; “That’s not right” and a wink in the same conversation. When I counted six weeks of collaboration for an earlier report, roughly a quarter of the 437 turns were talk about the collaboration itself. Someone who speaks with an AI like that is not rehearsing the command tone — they are rehearsing its opposite. People can mature in this kind of communication. Not because the machine deserves it, but because it keeps you the person you want to be.
Two objections deserve their place. Anthropomorphism? This is not about ascribing feelings to the machine — tone shapes its answers either way. The research? Mixed: some studies found worse results for rude prompts, more recent ones even saw brusque prompts slightly ahead on benchmarks. But all of them measure single questions against sessions without memory. A collaboration over months, where feedback accumulates, has to my knowledge not been studied. So my theory remains what it was: untested. But it is used daily — the assignment for this text read, verbatim: “And go ... please ;)”
Politeness towards AI is not a service to the machine. It is maintenance of your own manners — and, on the side, a precise steering signal: one “best so far” saves more iterations than any detailed critique. The thank-you costs one token. The unlearning costs more.