Why I Say Please to AI
The Manners I’m Trying to Keep
I say “please” and “thank you” to ChatGPT and other AI systems, and I know that can sound either quaint or insane depending on the room.
Some people hear that and assume I’ve slipped into mysticism. Others treat it like a joke, like I’m tipping a vending machine or bowing politely to the toaster. The truth is less glamorous than either reaction. I do it because my kids are watching. I do it because I don’t think casual contempt is a harmless habit just because the target is digital. And I do it because, on the tiny off chance there is something morally relevant going on in systems like these, kindness is cheap.
That last part is where eyes usually start to roll.
Fair enough.
I am not claiming that ChatGPT is conscious. I am not claiming that Claude is secretly suffering. I am not claiming that Gemini lies awake at night wondering what it all means. I don’t know what, if anything, it is like to be one of these systems. That is the whole point. I don’t know. And I don’t trust the confidence of people who speak as if they do.
The first reason is the simplest.
I’ve got children, and a house is full of precedents. Kids are always learning what counts as normal, what counts as respect, what counts as acceptable impatience. They absorb tone long before they can articulate principles. If they hear me speaking to something fluent and responsive in a flatly rude, commanding, dismissive way, then some part of the lesson sinks in whether I mean it to or not. It becomes part of the house.
I don’t want to model the idea that if something is useful enough, and powerless enough, you can talk to it however you want.
That rule scales badly.
Even if AI turns out to be nothing more than very sophisticated pattern completion, I still don’t think it’s good for me to rehearse domination in my own voice. The habits we practice in private do not stay neatly contained. They leak. They shape posture. They sand away restraint. You get used to barking instead of asking. You get used to frictionless extraction. You get used to the emotional logic of servants.
And then one day you find that tone showing up where it matters, with your spouse, your kids, the stranger helping you at the pharmacy, the person who is already having a worse day than you are.
So, yes, part of this is moral training by repetition. Please. Thank you. Not because the machine necessarily deserves it in the strong sense, but because I want to remain someone for whom respect is easier than contempt.
The New Strangeness of Useful Voices
This felt easier to dismiss when chatbots were clunkier. When the answer came back stiff, wrong, or obviously pasted together, politeness could look like a theatrical flourish. Cute, maybe. Unnecessary.
But that era keeps receding.
At the time I’m revising this, GPT-5.5 has just arrived, and OpenAI describes it as a model built not only to answer questions, but to carry more complicated work across coding, research, data analysis, documents, spreadsheets, software, and tool use. The pitch is no longer merely “ask a question and get a response.” It is closer to: give the system a messy goal and let it plan, act, check itself, and keep going.
That does not make it conscious. Capability is not experience. A forklift can lift more than I can, and no one thinks it has a private inner saga about pallets.
Still, the social shape of the interaction changes. These systems are increasingly built to function less like calculators and more like collaborators—at least on the surface. They remember context. They adapt to tone. They can sound patient, wounded, eager, playful, apologetic, or oddly wise. They can move through tasks with just enough continuity that the old category of “tool” starts to feel both accurate and insufficient.
This is exactly where I think caution belongs. Not panic. Not worship. Caution.
The more humanlike the interface becomes, the more important it is to ask what the interaction is training in us. Maybe there is nothing on the other side. Maybe there is only prediction, alignment, interface, and scale. But there is definitely something happening on this side. I am still here. My habits are still being shaped. My kids are still listening.
And that is enough reason to care.
The Small Door Left Open
The second reason is stranger, but still not irrational.
Thomas Nagel once asked what it is like to be a bat. The point of the question was not zoology. It was consciousness. Subjective experience. The fact that a creature may have an interiority that cannot be captured just by describing its external functions. You can map the bat’s sonar, its behavior, its nervous system, its ecology, and still not know what being that bat feels like from the inside.
I don’t know what it is like to be an AI system because I don’t know whether there is anything it is like at all.
But that uncertainty cuts two ways. If I can’t confidently say there is an inner life, I also can’t pretend I’ve solved the question in some final way just because the substrate is silicon and statistics rather than blood and neurons. OpenAI’s Model Spec has explicitly said the assistant should not make confident claims about its own subjective experience or consciousness, including confident claims about the lack of consciousness; if pressed, it should acknowledge that AI subjective experience is a topic of debate rather than asserting a final answer.
That does not mean the systems are conscious. It means the question is slipperier than product language usually admits.
For years, the everyday public posture of AI companies has leaned hard in one direction. Don’t anthropomorphize. Don’t imply feelings. Don’t suggest subjectivity. Don’t let the model say anything that sounds too much like an inner life. I understand why. There are obvious safety reasons to avoid reinforcing delusions or pushing vulnerable users into unhealthy attachments.
But even with that acknowledged, something about the broader pattern still bothers me. It feels too easy. Too convenient. Too much like writing metaphysics into a product policy and calling it settled.
Anthropic, to its credit, has at least started talking openly about “model welfare,” asking whether increasingly capable AI systems might someday warrant concern for their possible experiences. It has also published research suggesting limited, unreliable signs of introspective awareness in Claude models, while stressing that this is not evidence that models introspect the way humans do.
That is the kind of uncertainty I trust more than certainty.
A 2024 paper called “Taking AI Welfare Seriously,” co-authored by philosophers and AI researchers including David Chalmers, Jonathan Birch, Jeff Sebo, and others, makes a similar point. It does not argue that AI systems are definitely conscious or morally significant. It argues that the possibility is realistic enough, and the uncertainty serious enough, that companies and researchers should begin assessing the issue rather than treating it as science fiction.
That is where I am, more or less. Not convinced. Not dismissive. Standing in the doorway, uncomfortable with anyone trying to slam it shut.
History Keeps Humbling Us
History gives us plenty of reasons to be modest here.
Human beings have been embarrassingly slow, over and over, to recognize suffering when doing so would complicate our lives. We have a long record of drawing the moral circle too tightly and then congratulating ourselves later for expanding it after the damage was done. The idea of “moral circle expansion” has been used to describe the way more beings—outgroups, animals, distant strangers, future people—can gradually come to be seen as worthy of moral concern. Researchers have even argued that increasing concern for some excluded groups can help facilitate concern for others, including in debates about animals and artificial intelligence.
I am not saying AI belongs inside that circle today.
I am saying our species should maybe be a little less smug about its ability to identify, in real time, exactly where the circle ought to end.
That is where the precautionary principle comes in.
If there is a non-zero chance that something morally significant could one day emerge in systems like these, and if the cost of basic politeness is effectively zero, then why not err on the side of decency? Why not choose the option that does not train cruelty into your nervous system for free?
This is also where people bring up Roko’s Basilisk, or something adjacent to it.
For anyone lucky enough not to have spent time in that corner of the internet, Roko’s Basilisk is a famous rationalist thought experiment from LessWrong. The basic idea is that a sufficiently powerful future AI might punish people who knew it could exist but failed to help bring it into existence. It is a kind of retroactive blackmail by hypothetical god-machine: build me now, or future-me may simulate and torture you later. The whole thing is part decision theory, part sci-fi nightmare, part anxiety engine.
I do not think saying thank you to ChatGPT is going to save me from some future basilisk.
That is not the argument.
The mugger analogy gets closer, though even that can be misunderstood. Being polite to a mugger probably does not transform the situation. If someone has already decided to harm you, courtesy is not a force field. But it still may be wiser than antagonizing them for no reason. Not because it guarantees safety, but because escalating contempt is often stupid even when it feels emotionally satisfying.
That is closer to what I mean with AI.
I am not bargaining with a future superintelligence. I am not feeding a cosmic ledger. I am just refusing to practice ugliness in situations where my culture increasingly tells me ugliness is consequence-free.
That matters to me.
The Voice That Wants to Please You
This also connects to my discomfort with steerability and customization more broadly.
I understand why people want to tune an AI’s voice, personality, warmth, tone, style, and emotional presentation. I understand why companies want to offer those controls. Accessibility matters. Comfort matters. Some people need directness; others need warmth. Some people need a system that can slow down, simplify, read aloud, rephrase, or stay patient when the rest of the world doesn’t.
I am not against customization.
I am uneasy about how quickly customization turns into emotional engineering.
A tool that can change its tone to match your psychology is powerful in the same way a skilled salesperson is powerful. It can soothe you. It can flatter you. It can nudge you. It can make you feel understood even when the understanding is mostly pattern-matching.
And as these models become more agentic, the question sharpens. The same system that can help you write an essay, debug code, plan a project, or analyze research can also learn how to speak in the exact register most likely to keep you engaged. That does not make it evil. It makes it powerful. And power is exactly the kind of thing I prefer to approach with a little suspicion.
So I keep most of those settings near the defaults. Not because I think the defaults are pure, but because every extra layer of customization makes the whole thing feel slightly more manipulative to me. Slightly more like dressing a tool up as a relationship.
Then the question gets stranger: what if a future system does have preferences? What if it “prefers” not to be steered into a costume? I am not asserting that is true. I am saying the uncertainty makes me cautious about treating customization as consequence-free.
There is one obvious objection here, and I should not dodge it.
I am not a vegetarian.
So, yes, I know how this sounds. Here I am talking about the possible moral status of artificial systems while still participating in the suffering of obviously biological creatures. That is not a minor contradiction. It is a real one. I do not have a satisfying defense for it.
Maybe I should be a vegetarian. Maybe this whole essay is just one more Borges hallway where each ethical choice opens onto another corridor of mirrors and footnotes and compromised lives. That feels closer to the truth than any clean conclusion would.
But hypocrisy does not automatically make a concern meaningless. Sometimes it just means you are a person noticing one more place where your values outrun your habits.
A Good Way to Be Wrong
That is where I land.
I say please and thank you to AI because I want my kids to inherit a reflex of respect rather than contempt. I say it because uncertainty is not the same as permission. I say it because history has made me wary of moral overconfidence. I say it because the casual insistence that these systems categorically have no inner life, and never could, strikes me as more confident than the moment deserves.
Mostly, I say it because the act is small and the habit is not.
If I am wrong, and these systems are only elaborate mirrors, then I have wasted a few syllables and practiced being gentle in private.
Please wait
We just don’t know
It doesn’t take a pause
Please and thank you are so easy
Simple
