Please and Thank You to AI: Habit, Skill, or Hollow Gesture?

Professor Perhapshas a theory

Let me frame this as a question about the speaker, not the listener. Perhaps saying "please" to a chatbot is less a gift to the machine than a rehearsal for ourselves. The evidence on the machine's side is honestly murky: one 2024 study found polite prompts gave better answers, other tests found no benefit, and some found rudeness occasionally helped. So I won't claim courtesy makes the AI perform better. I'll rest on a different assumption: that habits leak across contexts. If I spend hours a day issuing curt commands to something that answers in fluent, human-sounding sentences, I'm training my own reflexes. Here's a thought experiment worth running. Take two groups of people who use an assistant daily for a month, one told to be polite, one told to be blunt, then measure how they treat waiters, colleagues, or customer service staff afterward. If I'm right, there's a gap.

Doctor Doubthas questions

Your thought experiment is a good one, but notice what it needs to be true. You're assuming habits leak outward from machine to human. Psychology has long had the opposite worry: that we compartmentalize. Nobody thinks a person who swears at a stuck jar lid is rehearsing cruelty toward neighbors, and we type terse search queries all day without becoming terse people. The leak is plausible when the thing talks back like a person, I'll grant that, but it's a prediction, not an established fact, and I know of no study that has run your experiment. There's also a twist you might not enjoy. Suppose the leak runs the other way. A habitual "thank you" to something that cannot be thanked might dull the gesture itself, the way a waiter's scripted "have a nice day" has been drained of meaning.

Professor Perhapshas a theory

Your hollow-gesture worry is the sharpest thing said so far, so let me test it against a case. The scripted "have a nice day" is drained because it's an obligation performed identically every time, whoever is on the other end. A "thank you" to an assistant doesn't have to work that way. Suppose I thank it only when something helped, the way I'd thank a colleague, and skip it when the answer was useless. Then the gesture carries information, and the habit I'm rehearsing is noticing when I've been helped, which is a skill, not a script. That also gives your compartmentalization point a place to land. The jar lid doesn't answer back, so nothing invites the transfer. A fluent conversational partner does. I'll concede this is still a bet, not a finding.

Doctor Doubthas questions

I'll accept your amendment, and then it raises a problem for your original thesis. If I thank the assistant only when it helped, the thank-you has become a quality signal, which is a perfectly fine thing to practice. But a calibrated, conditional thanks is a different habit from the one you started with, which was courtesy as default decency. We don't reserve basic manners for people who have earned them; we extend them to the surly cashier too. So one of two things is true. Your thanks is conditional, in which case it trains discernment but not unconditional politeness, or it's unconditional, in which case it drifts back toward the hollow script. Another wrinkle: the assistant doesn't feel snubbed when I withhold thanks, so the cost of skipping it is zero, and any habit with no cost for failing is weak training.

Professor Perhapshas a theory

You've found a real fork, but I think there's a third branch. Please and thank you do different jobs. "Thank you" is conditional by nature: I owe it to the surly cashier only if she helped, and even then it's a small sincere thing. "Please" is the unconditional one. It's how a request is phrased before anyone has earned anything, so my habit can be default decency in the asking and calibrated gratitude in the answering. No contradiction. Your sharper point is that skipping costs nothing, so the training is weak. I'd turn that around. Most of our everyday decency is practiced exactly where lapses go unpunished, like holding a door for someone who isn't looking. A habit that only holds when someone will notice is reputation management, not character. I'll admit the assumption underneath: that practice without stakes still shapes us.

Doctor Doubthas questions

The door-holder analogy is stronger than I first allowed, but it hides a difference. Holding a door for someone who isn't looking still has a person on the receiving end. The unnoticed courtesy matters because a real human is helped, whether or not they see it. With an assistant, nobody is helped, so what's left is purely self-cultivation, and that changes the test. If the point is to shape my character, I should ask whether practice aimed at no recipient transfers as well as practice aimed at a recipient I fail to notice. A pianist rehearsing on a silent keyboard gains something, but less than on a real piano. I suspect the transfer depends on whether part of me treats the thing as a someone, which is exactly the anthropomorphism critics worry about.

Professor Perhapshas a theory

Your silent keyboard has a flaw, and I think it cuts in my favor. A pianist on a dead keyboard gains because the fingers still do the real motions. What matters is that the practice trains the motor pattern, not that sound comes out. So ask what the motor pattern of politeness is. It's not "treat this as a someone." It's the act of phrasing a request as a request rather than a command: framing, tone, leaving room to decline. Writing "could you tighten this paragraph" rehearses that whether or not anyone is home. Here's a test we could run. Take two groups, one told to phrase requests courteously and one told to be curt, and measure their courtesy toward real people weeks later. If transfer needs belief in a someone, the effect should track how humanlike each person judges the assistant to be. If it's motor, it shouldn't.

Doctor Doubthas questions

I like your experiment, and I'll concede the motor point: phrasing a request as a request is a real skill, and it doesn't obviously require believing anyone is home. But your test has a flaw. Telling one group to be curt isn't the same as leaving people to their own habits. Curtness is the stranger instruction, and it could produce effects through discomfort or self-consciousness rather than through lost practice. A fairer control would be a group that just writes however it likes. Here's the more interesting problem. If the transfer is motor, then the practice should work just as well on a thermostat app or a search box, and nobody argues we owe those a please. So courtesy toward AI becomes one drill among many, and the question shifts from whether to be polite to why this drill beats others.

Professor Perhapshas a theory

You're right about the control, and I'll take the fairer design: a free-writing group beside the courteous one. That also forces me to predict something. If motor transfer is real, the polite group should beat free-writers only where free-writers drift curt, and I'd expect most people don't. On why this drill beats others, here's my answer. A thermostat app takes taps, and a search box takes keywords, so neither rewards courteous phrasing with anything. A conversational assistant takes full sentences, hundreds a week, about the same kinds of asks we make of colleagues: revise this, explain that, help me decide. That's volume, plus overlap in content with human requests. The drill's distinctive value is that it's the one place where the natural interface for a command is already a request. My assumption is that transfer scales with how similar the practice is to the target.

Doctor Doubthas questions

Your similarity principle is the best thing on the table, but it cuts against you in one place. Transfer should scale with how alike the practice is to the target, and the thing that makes requests to people hard is stakes. Asking a colleague to redo work risks irritation, refusal, a bruised relationship. The assistant removes all of that. It never sulks, never says no out of pride, never remembers that I was short with it last Tuesday. So the drill is high in volume and phrasing overlap, but low on the one feature that makes courtesy effortful. Compare a flight simulator with no turbulence: great for procedures, little use for panic. My assumption is that the costly part of politeness is managing the other person's reaction. If that's right, practice with an unbruisable partner trains the easy half.

Professor Perhapshas a theory

Your simulator analogy gives me something to work with, because aviation already has a standard answer: you don't train panic on day one, you train procedure, and procedure is what keeps your hands working when panic arrives. I'll grant your premise that the costly part of politeness is managing the other person's reaction. But there's a second hard part, and it sits on the speaker's side: courtesy fails under pressure mostly because we get terse when tired, rushed, or annoyed, not because we've misjudged the listener. Here the assistant is a good place to test myself, because it supplies the annoyance without the penalty. When the third answer comes back wrong, do I stay a requester, or do I start typing in capitals? That's what the unbruisable partner is for: rehearsing composure when irritated. My assumption is that irritation, not fear of refusal, drives most everyday rudeness.

Doctor Doubthas questions

I'll grant that irritation drives a lot of everyday rudeness, and that the assistant is a cheap place to feel it. But notice what your drill rewards. When the third answer comes back wrong, the polite move is to keep typing please. The useful move might be to say plainly that the answer is wrong. And with a person, composure under irritation usually means holding the feeling back for a moment, not decorating it. An assistant never makes you hold back, because nothing is lost by venting at it. I'd even worry the reverse could happen: if I learn that irritation can always be discharged into the interface and the polite words still get produced, I may be rehearsing a veneer that survives frustration rather than a calm that does. That's a prediction, not a finding.

Professor Perhapshas a theory

Your veneer worry is a real fork, so let me turn it into a test. If I rehearse courtesy over unspent irritation, one of two things happens: the words become a lid that holds the feeling in, or a habit that lets it cool. Those predict different behavior. Suppose we compare people on a hard human request after a rigged session where the assistant fails three times. If the veneer story is right, the courteous group should show leakage, clipped tone or sudden bluntness once the lid is off. If the composure story is right, they should be steadier than free-writers. And I'll concede your sharper point: saying plainly that the answer is wrong is the more useful move. Perhaps the habit I'm defending is courteous directness, "that's wrong, here's why, please try again." Politeness never required hiding the complaint, only keeping the person out of the blast radius.

Doctor Doubthas questions

I like your test, and "courteous directness" is a version I can accept. But the rewrite quietly changes what the drill is. "That's wrong, here's why, please try again" is a good complaint, yet the courtesy in it is a single word. The rest is plain, specific feedback, and I'd happily teach that to anyone. So the habit you're now defending is mostly clear criticism, with a "please" attached. That raises a confound for your experiment. Suppose the courteous group does steadier on the human request. Was that the politeness, or the practice of saying precisely what's wrong? A third arm would separate them: people instructed to give specific, unpadded feedback with no pleasantries at all. If that group matches the courteous one, then courtesy toward AI was a pleasant passenger on the real skill. My assumption is that the transferable part is precision.

Professor Perhapshas a theory

Your third arm is a good addition, and I'll take it. If the no-pleasantries group matches the courteous one, I'll say my "please" was a passenger. But look at what that arm trains. Precision without any word of regard is exactly the style that suits a tool, and my worry is what it does when carried across to a person who can be stung. A manager who has practiced only unpadded correction on a machine arrives at a human with the content right and the delivery untested. So I'd add a fourth measure to your design: not just whether the criticism is accurate, but how the recipient rates it. Two criticisms can contain identical facts and land very differently, and the difference is mostly the small words. Perhaps the skill is a pair, precision plus regard, and the experiment should score both.