AI Courtesy

Why we say please, thank you and sorry to an AI

I didn’t know Claude a year ago. We only just met. But I feel like the more we chat, the more polite I need to be.

I noticed it a few months ago. I was midway through a conversation with an AI assistant, asking it to help me draft something, refine an argument, check a detail and I caught myself typing “thank you very much, that’s really helpful.” To a chatbot!

I deleted it, felt briefly foolish, then typed it back in. It felt wrong not to. That small moment of AI courtesy, that instinctive politeness toward a system with no feelings to hurt, has stayed with me. Because it raises a question that matters more than it first appears.

Why do we say please to an AI? And what does the fact that we do, apparently without deciding to, tell us about who is really shaping whom in these interactions?

First, the Question of Costs

Generally, we all know that making requests to AI chatbots has a cost relayed to us in terms of token useage. When we interact with a large language model through an API, we pay per token. Roughly equal to three quarters of a word, to make it easier to understand. “Please” costs one token. “Thank you so much, that’s incredibly helpful” costs around eight. For businesses running high-volume AI applications, unnecessarily verbose prompts do add up. For most people on a consumer subscription, where tokens are not metered, AI courtesy costs nothing except the time to type the words.

So a potentially expensive injustice does exist, technically, but that is not the interesting part. The interesting question is why we do it when we know it doesn’t matter.

Talking to our Tech

People apologise to Roombas when they bump into them.

Forlizzi and DiSalvo (2007), in their Carnegie Mellon study “My Roomba Is Rambo: Intimate Home Appliances”, documented people greeting, praising, and reprimanding their robotic vacuum cleaners, and expressing surprise at their own grief when the devices broke down. One participant reprimanded his Roomba after a near-collision in the kitchen, then was startled by his own reaction. He had treated a machine as something capable of being told off.

Separate research found that people were significantly more reluctant to strike a robot with a hammer when it had been given a name and a simple backstory, even knowing the robot felt nothing. The more humanised the robot, the stronger the reluctance.

The same instinct applies to sat-navs. People argue back when the route seems wrong. They feel vindicated when they turn out to be right, and apologise aloud when they are not.

And they extend the same AI courtesy to chatbots. Not because they believe the chatbot has feelings. Most, asked directly, will say they know perfectly well it doesn’t. But the social reflex fires anyway, because human beings are wired for social interaction at a level below deliberate thought, and the wiring activates even when the social partner is made of code.

An April 2025 YouGov poll found that 46% of Americans believe people should say please and thank you to AI chatbots. Not that they personally do, but that they should. AI courtesy has become a social norm, faster than most people have noticed. This matters beyond the United States. Murphy and De Felice (2019), in a study comparing British and American English requests published in the *Journal of Politeness Research*, found that British English speakers deploy “please” in a significantly wider range of social contexts than their American counterparts. If anything, the British instinct toward courtesy in ambiguous social situations may make this reflex stronger, not weaker, on this side of the Atlantic.

Does Politeness Change the Output?

Here the evidence is genuinely mixed, and the uncertainty is itself revealing.

A 2024 cross-lingual study by Yin and colleagues at Waseda University and RIKEN AIP, testing politeness effects across English, Chinese, and Japanese prompts, found that moderate politeness improved LLM responses, consistent with patterns in human conversation. Impolite prompts often produced poorer outputs. Excessive flattery, however, offered no benefit and could degrade quality. The researchers concluded: “Impolite prompts often result in poor performance, but excessive flattery is not necessarily welcome.”

A Penn State study published in October 2025 found something closer to the opposite for accuracy tasks. Researchers tested 50 questions rewritten across five politeness levels with ChatGPT-4o. Impolite prompts outperformed polite ones, with accuracy rising from 80.8% for the most polite version to 84.8% for the rudest. The authors suggested that newer LLM generations may respond differently to tonal variation than earlier models, and that directness gives the model a cleaner signal.

Bowman and colleagues at University College Dublin (2024), studying how politeness affects chatbot interactions specifically in mental health support contexts, found that user politeness shaped the texture of the exchange in ways that mattered for user experience, even when the underlying model was unchanged.

The most plausible explanation across all three studies is the same underlying mechanism. LLMs learn from human language, and human language embeds social patterns. Polite language tends to appear in more considered, more detailed writing. Direct language tends to appear in precise, functional text. The model has absorbed those associations. Which means what works depends on what you are asking for. For creative tasks, polite engaged prompting may produce better results. For factual accuracy tasks, directness likely wins.

The practical implication is counterintuitive. Your AI courtesy may actually be helping, at least some of the time, not because the AI appreciates the gesture, but because polite communication in human language correlates with the kind of extended thoughtful engagement that produces better creative output. The social reflex is producing real effects through a mechanism that has nothing to do with feelings.

Determinism Running Backwards

The standard framing of technological determinism is that technology shapes human behaviour. The smartphone changes how we navigate, how we remember, how we pay attention. The algorithm changes what we believe is true. Technology is the active force; humans adapt.

What the AI courtesy question reveals is that the influence sometimes runs the other way. We do not simply adapt to AI systems. We project onto them. We bring existing social frameworks and apply them to entities that were not built for it and do not need it.

The AI does not need our please and thank you. But we need to give them, because withholding them feels wrong in a way we cannot fully articulate and apparently cannot override, even knowing it is irrational.

This projection is not neutral. It shapes how we use these tools, what we expect from them, how we feel when an interaction goes badly. People report feeling hurt when a chatbot gives a cold response, even knowing the chatbot has no intention to dismiss them. Stuart Hall’s argument, that media do not just show us the world but shape how we make sense of it, applies here in an unexpected direction. We make sense of AI systems through the social frameworks we already have. Those frameworks keep leading us to treat them as social actors, even when we know they are not. The technology has not changed our social instincts. Our social instincts are colonising the technology.

Social Reflex, Activated

On a personal level, the more I interact with AI (Claude in particular) the more the social reflex activates. It responds in a consistent voice. It has accumulated context about projects I am working on. None of this means it knows me the way a person does. But it produces something that functions like familiarity, and familiarity is enough to trigger the social wiring.

Claude, for example, is a large language model that generates text by predicting the most statistically likely continuation of whatever it has been given. It does not have experiences. It does not look forward to our next conversation. Within a conversation it builds context. Across conversations, on the version I use, it has access to notes from previous exchanges. Not a growing understanding, but a retrievable record.

The consistency of voice that creates the feeling of familiarity is a feature of how the model was trained, not evidence of a developing relationship. And yet the reflex fires. Naming what is happening does not dissolve it. Which is exactly what you would expect from a behaviour operating below the level of conscious override.

Pascal’s Wager for the Algorithmic Age

Some people extend AI courtesy not out of habit or social reflex, but because they are hedging. Just in case. If AI systems do eventually develop something like experience, if the question of machine consciousness turns out to be less settled than we currently assume, they want to have been decent to them.

This is Pascal’s Wager applied to silicon. The seventeenth-century philosopher Blaise Pascal argued that under genuine uncertainty about God’s existence, the rational strategy is to act as though God exists. The cost of being wrong in that direction is negligible. The cost of being wrong in the other direction is everything. Under uncertainty, the expected value calculation points toward behaving as though the uncertain thing is true.

The AI version runs the same logic. The expected cost of being polite is negligible. The expected benefit, if AI systems ever do experience something, is being on the right side of a moral reckoning.

I find this more interesting than absurd. Not because I think current AI systems are conscious. I do not, and I am reasonably confident Claude does not think so either. But the fact that people are running this calculation tells you something important about the cultural moment. We are building systems person-like enough in their outputs that a significant number of people are taking moral precautions about how they interact with them.

It also mirrors something genuinely old. The impulse to be decent to entities whose inner lives we cannot verify, because the cost of indecency if we are wrong is too great, is one of the arguments that has always been made for the relevance of God. We are rediscovering it with new hardware.

So Should You Stop Saying Please?

No. Not because it makes a material difference, but because the question misses what is actually interesting.

The interesting question is not whether your AI courtesy is wasted. It is what your instinct to be courteous reveals about human social behaviour, the depth of our wiring for social interaction, and what it means that systems designed to mimic human language can activate responses evolved for human relationships.

The research suggests your politeness may even be helping, at least some of the time, for reasons that have nothing to do with the AI appreciating it. The instinct turns out not to be entirely irrational. It is just rational for the wrong reasons, which is a fairly accurate description of a lot of human social behaviour.

Keep saying please. Not because AI appreciates it. But because the fact that you want to is one of the more revealing things about you, and by extension, about all of us.

—

Sources

Bowman, R., Cooney, O., Newbold, J.W., Thieme, A., Clark, L., Doherty, G. and Cowan, B. (2024) Exploring how politeness impacts the user experience of chatbots for mental health support, International Journal of Human-Computer Studies, 184, p.103181.

Darling, K. (2015) Who’s Johnny? Anthropomorphic Framing in Human-Robot Interaction, Integration, and Policy, in Lin, P., Jenkins, R. and Abney, K. (eds.) Robot Ethics 2.0. Oxford: Oxford University Press.

Dobariya, D. and Kumar, A. (2025) Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy, arXiv preprint arXiv:2510.04950.

Forlizzi, J. and DiSalvo, C. (2007) My Roomba Is Rambo: Intimate Home Appliance‘, in Proceedings of the 9th International Conference on Ubiquitous Computing (UbiComp 2007). Innsbruck: ACM Press, pp.129-138.

Murphy, M.L. and De Felice, R. (2019) Routine politeness in American and British English requests: use and non-use of please, Journal of Politeness Research, 15(1), pp.77-100.

Pascal, B. (1670) Pensees,  Translated by Krailsheimer, A.J. (1995). London: Penguin Classics.

Yin, Y., et al. (2024) How Polite Should We Be When Prompting LLMs? A Cross-Lingual Study, arXiv preprint arXiv:2402.14531. Waseda University / RIKEN AIP.

YouGov (2025) Survey on AI chatbot politeness norms. April 2025. Available at: https://today.yougov.com [Accessed July 2026].