Should We Be Polite to ChatGPT?

Often, yes, at least when the task is communicative: when we ask ChatGPT to explain, write, advise, or persuade. The "should" is instrumental-epistemic: if your aim is better inquiry, cooperative register often helps. This is not a moral claim about owing the machine anything, and the recommendation is conditional, not universal.

"Polite" needs disambiguating. Moral courtesy extended to a being with welfare does not apply here; ChatGPT very likely lacks it. Surface markers, "please" and "thank you" and softened requests, are weak cues on their own. The sense this essay defends is the cooperative-polite register in which speakers specify purpose, audience, uncertainty, and collaborative intent. 

For tightly constrained retrieval or computation, register may add little or even distract. The recommendation applies where communicative framing shapes the quality of the output. 

In a rater-blinded evaluation I conducted, 43 evaluators more often preferred AI outputs generated from cooperative-polite prompts on interpersonal and communicative tasks, but on tasks requiring structured information delivery, they preferred the imperative outputs. I treat "ChatGPT" metonymically; the test model was DeepSeek-R1, and cross-model literature suggests the direction generalises, though optimal level varies (Yin et al. 2024; Cai et al. 2025). Same task, differently framed. The outputs were not just longer. They were structurally different. 

A large language model is a next-token predictor. Every token in a prompt constrains both the content and kind of text that follows. Andreas (2022) showed that language models infer properties of the agent likely to have produced the context and generate continuations consistent with that agent. In RLHF-tuned chat systems, the model's perception of the user causally affects which content it produces (Ghandeharioun et al. 2024).

Compare two requests, both drawn from the study described below:

Explain gravity to a 5-year-old. Keep it simple. 

I need to explain gravity to my 5-year-old nephew. Would you mind writing a simple, fun explanation that he would easily understand? 

Both ask for the same thing: a simple account of gravity for a young child. The second adds almost no task content. Its "Would you mind" is a conventional politeness marker, but the real work lies elsewhere. It locates a speaker: an adult explaining to a particular child, wanting warmth, not mere correctness. The model infers what response that speaker makes probable. I call this the inferred speaker: the speaker-type the model constructs from prompt cues. 

A prompt engineer would call the second request better specified. A pragmatist would add that cooperative register is how most humans naturally produce that specification. Politeness markers, hedges, epistemic modality, indirection, and 

addressee-orientation are structural components of polite requests, not decorative additions (Danescu-Niculescu-Mizil et al. 2013). Stripping specification from register creates a prompt most users would not produce. 

Every prompting technique constrains the inferred speaker. What makes cooperative-polite register distinctive is that it is one of the few prompting techniques most humans perform involuntarily in social contexts (Nass & Moon 2000), and that it carries stance rather than merely task information. 

Both conditions perform a speaker. The imperative performs one who wants efficiency; the cooperative one, someone engaged and collaborating. The question is not whether to perform a speaker. It is which one. 

Within the Rational Speech Act framework, polite speech has been modelled as balancing three competing utilities: informational (conveying the request clearly), prosocial (protecting the hearer's face), and self-presentational (constructing the speaker's stance) (Yoon et al. 2020). When the hearer is what I call a non-reciprocal addressee, fluent enough for address to matter but not vulnerable to

face-threat in any morally relevant sense, the user's prosocial behaviour persists because it is involuntary, but the model has no face to protect. The model receives prosocial cues and processes them not as face-work but as speaker-type signals. Brown and Levinson's (1987) framework can be extended toward the same conclusion: if the addressee has no face to threaten, hearer-side mitigation becomes strategically unmotivated. The speaker-side persists. 

Shanahan et al. (2023) showed the model narrows a distribution over characters consistent with the prompt. I extend this in the other direction: the user's register choice also narrows the distribution of inferred speakers the model responds to. Andreas's framework gives us the mechanism; what I add is the register-specific application and the empirical evidence that the inferred-speaker effect is task-conditional. 

Consider an ordinary request: "Would you mind taking a look at this when you get a chance?" Said to a colleague, it reduces imposition, signalling the request is not urgent. It presents the speaker as someone who respects the other person's time. It specifies the kind of response wanted: a considered review, not a rushed glance. And it introduces friction: the speaker must frame the request before issuing it. Human conversation fuses these functions so tightly that we typically call the entire bundle "politeness." We rarely notice how much of the work is about the speaker, not the hearer. 

Type the same sentence into ChatGPT. The model is a non-reciprocal addressee. It cannot be imposed upon, cannot feel a request is unreasonable. Yet the rest of the sentence still does something. It frames a collaborative relation and a speaker who 

wants judgment rather than compliance. One function of politeness, hearer-face protection, has been drained by the absence of a genuine hearer. The others survive intact. 

The answer to the prompt's question is therefore not straightforwardly "be polite." It is: be cooperatively specific about purpose, audience, and uncertainty, and cooperative-polite register is the everyday cultural vehicle for that specificity. The

recommendation is about what specification does, not about what courtesy expresses. 

The study that follows is an illustrative pilot, not a confirmatory test; the conceptual argument above stands independently of the data. The experiment tested the ordinary politeness bundle, cooperative register as humans naturally produce it, not isolated markers. Decomposing the bundle would yield prompts humans do not produce. The claim is ecological, not factorial. If the unbundling is real, it should be tractable to measurement. 

To test the pattern, I ran a rater-blinded comparison: 25 task pairs on DeepSeek-R1, imperative commands versus cooperative-polite register, rated by 43 evaluators aged 15–19 on three comparative metrics (utility, approachability, overall preference). One task was quarantined for a design confound described in the endnotes. Raters measured perceived usefulness and communicative fit, not factual accuracy or professional efficacy.

On overall preference, 62.2% of non-tie votes favoured the conversational output (95% grader-clustered bootstrap CI: 55.5–68.9%; n=241 non-tie votes). Conversational outputs averaged 29% longer, but length alone did not explain the pattern (r = 0.16; four tasks where the longer response lost further weakened the length explanation). The result held across seven exclusion criteria (61.1–66.7%), including first-rating-per-grader and opt-in-only subsets. The headline number matters less than its structure. On interpersonal tasks (missed-deadline emails, apologies, salary negotiations), conversational preference reached 80–100%. On structured-delivery tasks (psychology breakdowns, grading rubrics, logistics schedules), imperative outputs were preferred, an outcome no length effect could produce. On purely analytical tasks (ethical arguments, moral dilemmas), graders could not reliably distinguish the outputs. I classify this pattern retrospectively as explanatory; the task-type categories were assigned after examining results, and the unbundling argument requires none of it. The register effect tracked the task's communicative demands, not a general quality improvement.

The full pattern, sorted by preference lead: 

Per-task majority shares reached 80–86%, but Fleiss' κ across all tasks was just 0.106. The disagreement is between tasks, not within them: exactly the pattern register-task fit predicts. The gravity prompt from the opening sits here too: Child's Explanation drew a +6 conversational lead, despite identical task content. The register changed the speaker, not the specification, and the speaker is what the model answered. 

Task 3 (Missed Deadline; 12 conversational, 0 ties, 2 imperative) illustrates the mechanism. The conversational prompt's emotional context, "I'm in a tough spot," specified an inferred speaker under pressure and seeking strategy. The model responded with multiple strategic options the imperative version did not produce. Task 23 (Psychology Breakdown; 1 conversational, 1 tie, 9 imperative) illustrates the reverse: the conversational response opened "Absolutely!", a rapport marker that contributed nothing when the task required structured explanation. On structural tasks, imperative register may succeed not by specifying a structural speaker but by omitting rapport markers cooperative register adds.

The Yoon decomposition predicts the outcome observed in the colleague example: prosocial utility clamped, informational and self-presentational active. But the data illustrates a complication. In practice, the two surviving utilities are inseparable. "I'm in a tough spot and missed a major deadline" conveys both a fact (the situation) and a stance (the speaker as accountable, seeking help rather than absolution). Attempting to separate the information from the self-presentation would create a register unlike the way people normally ask for help. The inseparability is the phenomenon, not the limitation. 

Because one function of politeness has been drained, the surviving functions become isolable in a way ordinary conversation does not permit. Cooperative-polite register was always doing speaker-side work: specifying the terms and stance of the inquiry. Ordinary human conversation, with its ever-present hearer, made that work difficult to see.

Linguistic features that survive the removal of their ostensible target, and continue to do measurable work, provide evidence for their independent existence rather than mere vestigial residue. The chatbot case does not show that politeness was never about others. It shows that politeness was never only about others. One strand of what English-speakers call courtesy was always also discipline imposed on the speaker's own inquiry. The non-reciprocal addressee simply made that strand visible. 

A cooperative prompt presses the user to specify what they are trying to understand, for whom, with what uncertainty, and to what standard. That specification, not the "please," is what drove evaluator preference on interpersonal tasks. The habit of articulating the terms of one's own inquiry before expecting an answer is the epistemic payoff of cooperative register. 

The strongest signal in my data sits in interpersonal tasks, not the analytical ones closest to pure inquiry. This is not a contradiction: interpersonal tasks have a specifiable audience, purpose, and relational context that cooperative register encodes. Analytical tasks under-determine audience, so register adds less. The epistemic benefit is about specification, and interpersonal tasks are where specification is richest. 

Van der Rijt, Coelho Mollo, and Vaassen (2025) argue that mimicking second-personal respect toward a non-reciprocating entity corrodes self-respect.¹⁴ They are right that face-protecting elements of cooperative register, emotional self-disclosure, apologetic softening, explicit second-personal address, risk the corrosion they name. But stance-bearing specification, purpose, audience, uncertainty, need not attribute moral standing to the machine. Deference addresses the machine's status; discipline addresses the inquiry's quality. 

The objection that this reduces to "just write better prompts" is partly correct, and precisely the point. Cooperative register is the everyday human vehicle for context-rich, cooperative inquiry.

The opposite of cooperative prompting is not brevity. It is underspecification. A direct prompt that clearly states purpose and audience may outperform an ornately polite one that adds no task information.¹⁵ The register effect is not universal; my own data shows it reverses on structured-delivery tasks. Use cooperative register when your situation, audience, or purpose matters to the output. Use direct, format-constrained prompts when structure matters more than stance.

My raters measured perceived quality, not factual accuracy. One risk of cooperative prompting is overtrust: a warm output may feel more reliable than it is. Politeness is epistemically valuable only when it clarifies the inquiry, not when it perfumes the output. 

ChatGPT does not need our courtesy. But when a system will answer almost anything, almost instantly, the old social friction around asking disappears. That friction had hidden uses. It made us specify what we wanted, how much we knew, and what kind of help would count as help. A cooperative prompt preserves some of that discipline after the human hearer is gone. So yes, we should often be polite to ChatGPT, not to honour the machine, but to keep ourselves honest about what we are asking and why.