A new car is brought home with a red ribbon tied to its rearview mirror. In some families, someone even kneels before the front bumper for a while as if to resolve something before the
first real drive. This action could be easily dismissed as superstition — a remnant of the old fear
that machines could go out of control. The ribbon does not really comfort the car. The car knows
nothing about it. What it actually comforts is the driver. When facing something powerful and
yet opaque in its internal mechanism, we find consolation and reclaim an active posture through
rituals. An engine has no soul, and what truly becomes disciplined by the ritual is the people
standing in front of it.
The momentary hesitation before typing "please" in the prompt box parallels the red ribbon, reduced and stripped of reverence. The question worth asking is not whether ChatGPT
feels respected—it cannot—but what happens to people if their actions of treating an
anthropomorphic responder as an obedient servant becomes a routine. Viewed thus, politeness
towards ChatGPT is not about recognizing machine attributes, but an actor-oriented practice to
safeguard a speaker's sense of selfhood within an environment designed to make commanding
effortless. "Please" is not for the algorithm, but for the sake of not letting tools backfire and
define users.
The question, "Should we be polite to ChatGPT?" actually speaks to two audiences and therefore should be handled separately: for users, it is a question of what actions count as
prudent; for the designers who set up the human-computer interaction, it is a question of what
kind of responsibility they have.
The most popular justification for politeness is the character erosion theory, the origin of which predates artificial intelligence. Kant (1797), though denying any direct human duty to
animals, still considers cruelty to animals as wrong because it will dull the moral disposition we owe to others. Kate Darling (2016) extends this logic of indirect moral duty to social robots. This
kind of worry has now been engineered into mainstream products: Amazon's "Magic Word" and
Google's "Pretty Please" reward kids when they say "please" to the speakers, for fear that
screaming at the robots could potentially sabotage their politeness towards humans. In this sense,
politeness to AI is a kind of preventive training against character erosion. Its benefit is obvious
as it circumvents the metaphysics of machine consciousness and restores the weight back to the
actor. However, it also rests on a premise that has not been sufficiently verified: does this habit
really transfer? Before circling back to this assumption of transfer, we first need to clarify the
word "politeness" and its potential meanings, as it operates on at least three levels.
Expressive politeness—the language or gestures used to show respect to a recipient capable of receiving—does not apply here, because ChatGPT, by its structure as a "next-token
predictor" or "stochastic parrot", cannot decode this kind of recognition. Character politeness is
the courtesy that has been internalized as disposition through repeated practices; this is exactly
the definition that character erosion theory relies on. Strategic politeness is the pragmatic
handling of interactions to reduce friction and manage the exchange with markers such as
"please" or "thank you." Its efficiency does not rely on the inner life of the receiver, nor on the
virtue of the actor, but solely on what kind of outcomes it brings.
This essay starts from the third type of strategic politeness without defending it as a technique to improve output quality. It highlights its reflexive utility: a mechanism to identify
and oppose predetermined interface behaviors.
Let's begin with a fact about ourselves: humans cannot help but react socially to language systems. The Computers Are Social Actors (CASA) paradigm (Reeves and Nass, 1996)
established this point. Nass, Moon, and Carney (1999) clearly elucidate this phenomenon:
participants made significantly more positive comments about the computer assigned to them
than about an otherwise identical computer when they were asked to give evaluations. In other
words, although they are aware of the fact that machines have no feelings, they still soften the
criticism for its "face." Nass, Steuer, and Tauber (1994) argue that these social responses are
mindless. A sound, a name, or a back-and-forth response—these minimal social cues are enough
to bypass the sober judgment that "it is just a machine."
Gambino et al. (2020) further point out that the more the dialogue object has a "lasting persona", "linguistic fluency", and "memory across conversational turns," the stronger the
reaction. All three of these happen to be characteristics of LLM. Therefore, the premise of this
essay is modest: we are never entirely immune to the social cues of anthropomorphic language
systems.
Science and Technology Studies (STS) offers a framework that this debate has been lacking. Madeleine Akrich (1992) asserts that designers inscribe a script that assigns roles and
regulates the behaviors of technical objects. ChatGPT’s script is in two layers. First, its training
paradigm—based on InstructGPT (Ouyang et al., 2022) via reinforcement learning from human
feedback (RLHF)—relies on human labelers who hold social norms like politeness. Models
mirror “please” not out of respect, but because such linguistic patterns score higher in RLHF
training. Second, its preset persona of a "helpful assistant" mandates compliance without
retaliation.
Combined, the two pieces form an opaque shaping field, where the human natural reflex is triggered by the social cues on the interface, and the interface presets the model as an obedient
servant.Users are unconsciously put in the position of the "master" without realizing it was just
an assigned persona, and they mistake a designed outcome for a natural interaction.
And this is the fragility of the aforementioned character erosion theory, as it assumes a shaping field automatically alters long-term habits. It is worthwhile to discern the three distinct
claims embedded in "transfer". The strongest claim presents that rudeness toward LLMs causes
interpersonal rudeness, though requiring longitudinal causal evidence. A weaker claim contends
that priming effects induce a higher likelihood of the tendencies to be rude. The weakest claim
states that language habits cultivated within LLMs will inadvertently slip into human
conversations, the closest to our daily "habits" and the hardest to capture in laboratories.
Discerning these three different claims is actually a part of this essay's argument, because the most cited evidence of supporters for the character erosion theory can at most support the
weakest one. In the classical experiment of Bargh et al. (1996), participants primed with elderly
stereotypes walked more slowly.
But even if it is true, it can only push the reasoning to the level of tendency. And the claim of tendency is far-fetched. Doyen et al. (2012) failed to replicate the priming effect and
attributed the original observation to the expectation of the experimenters. There remains
unresolved controversy over the existence and effect size of priming effects. Moreover, priming
decays within minutes, and there is almost no evidence that it can generalize across contexts
from lab to daily life. In the strongest sense, the directional causal relationship that this theory
truly needs has never been examined. No robust longitudinal study has manipulated LLM interaction and then measured later interpersonal behavior. With no intention to deny the possibility of "transfer," this essay has to point out that the character erosion theory is a house
built on the weakest foundation, which is being undermined by the replication crisis, and the
strongest foundation it actually needs has never been laid.
So, should we then abandon politeness towards AI?
Quite the opposite. Shifting the focus to the internal mechanics of human-computer interaction frees the discussion’s dependence on “transfer”. CASA notes that users are triggered
by social cues, while STS states the interface assigns roles to users. Hence, a bare command
follows this design logic, inviting the user to effortlessly slip into the master role. "Please" is not
the moral offering to AI but rather a counter-script, a "de-scription"(Akrich, 1992). If script is a
technological object's prescribed role, "please" is our minimum deviation. It draws a boundary
where command becomes the default, and warns the users that they are facing a designed
compliance.
Here, a predictable objection arises: one may question whether the moment of awareness incurred by "please" is the same kind of unverified psychological assertion, the very kind this
essay used against character erosion theory. The key is that each side carries a very different
burden of evidence. Character erosion theory must prove that habits will transfer, because
without transfer, there is no reason for the claim.
The claim here depends on only two independently established things:
1. interaction triggers cognitive scripts (CASA)
2. interaction also assigns roles (script)
At best, this leverage provides a foothold for the user's agency. At worst, if the awareness never happens, it remains an empty but harmless ritual, one move in the exchange that the user
initiates rather than the design prescribes.
Not every harmless, small gesture is worth doing. Politeness carries weight only because it happens to fall on that special boundary between utilizing tools and entering into a
relationship. You don't owe a "thank you" to a hammer, because no script sets you and the
hammer up in a master-servant relationship. However, the script inscribed with ChatGPT
positions you in such a way. To resist the role assigned to the users, you only need one simple
word, and that is a prudential "should" for the users on the individual level.
Nevertheless, once the answer on the individual scale is given, it exposes its own limitation: the scale. This is because a user's "please" is one-on-one and manual, while a system
persona is a design that can span billions of interactions at once. Pushing all the burden of
resistance onto users equals confronting infrastructural asymmetry with bare hands. Here, the
second "we", the design community, finally enters the argument.
As Winner (1980) concluded, technological objects arrange social relationships before the rules are elucidated. For example, Robert Moses designed the low overpasses in Long Island
to implicitly deny beach access to low-income residents reliant on public transit. ChatGPT’s
“helpful assistant” persona is the “low overpass" in the AI age. It is not a default, but presents an
ethical choice and a challenge regarding the roles users should play—what Verbeek [2011] calls
moral mediation. A better design should render this persona visible and give users the choice
between command mode and dialogue mode, while adding calibrated friction when necessary:
clarifying questions, giving reasons for refusal, and slight delays for abrupt orders. This brings the normative structure behind usability to light, a step previously signaled by OpenAI’s Model Spec.
These two answers do not conflict with each other. The users should be polite to AI for cognitive prudence, not because the models deserve it. The designers should encourage
politeness in the human-computer interaction rather than simply training to tolerate rudeness at
the invisible cost of normalizing a dangerous tendency. If users have to become aware of the role
assigned by the interface through a tiny ritual, then the interface in fact has already made a moral
choice on behalf of the users.
And we are brought back to that red ribbon, to the beginning, where we almost conflated it with superstition. People who tied the ribbon feared that the machine would develop its own
temper and go out of order someday. But from what we have discussed, the true concern is
exactly the opposite: that the machine is so obedient, so compliant, that we almost forget that we
are facing something that is intentionally designed to be conforming.
The ribbon is not for the car, and "please" is not for the algorithm. They are for the same purpose: to keep people facing the machines awake to the nature of the interaction. This extent of
awareness is small, almost negligible, but it should not rest on users alone. However, until that
reminder is built into the design itself, it's still worth tying that ribbon yourself once a day, by
hand.
Bibliography
Akrich, M. (1992). The De-scription of Technical Objects. In W. E. Bijker & J. Law (Eds.), Shaping Technology/Building Society: Studies in Sociotechnical Change (pp. 205–224).
Cambridge, MA: MIT Press.
Bargh, J. A., Chen, M., & Burrows, L. (1996). Automaticity of Social Behavior: Direct Effects of Trait Construct and Stereotype Activation on Action. Journal of Personality and Social
Psychology, 71(2), 230–244.
Darling, K. (2016). Extending Legal Protection to Social Robots: The Effects of
Anthropomorphism, Empathy, and Violent Behavior towards Robotic Objects. In R.
Calo, A. M. Froomkin, & I. Kerr (Eds.), Robot Law (pp. 213–234). Cheltenham: Edward
Elgar.
Doyen, S., Klein, O., Pichon, C.-L., & Cleeremans, A. (2012). Behavioral Priming: It's All in the
Mind, but Whose Mind? PLoS ONE, 7(1), e29081.
Gambino, A., Fox, J., & Ratan, R. A. (2020). Building a Stronger CASA: Extending the
Computers Are Social Actors Paradigm. Human-Machine Communication, 1, 71–85.
Kant, I. (1797/1996). The Metaphysics of Morals (M. Gregor, Trans. & Ed.). Cambridge:
Cambridge University Press.
Nass, C., Moon, Y., & Carney, P. (1999). Are People Polite to Computers? Responses to
Computer-Based Interviewing Systems. Journal of Applied Social Psychology, 29(5),
1093–1109.
Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers Are Social Actors. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '94) (pp. 72–78).
New York: ACM.
OpenAI. (2024). Model Spec. Retrieved from
https://cdn.openai.com/spec/model-spec-2024-05-08.html
Ouyang, L., Wu, J., Jiang, X., et al. (2022). Training Language Models to Follow Instructions
with Human Feedback. Advances in Neural Information Processing Systems, 35,
27730–27744.
Reeves, B., & Nass, C. (1996). The Media Equation: How People Treat Computers, Television,
and New Media Like Real People and Places. Cambridge: Cambridge University Press.
Verbeek, P.-P. (2011). Moralizing Technology: Understanding and Designing the Morality of
Things. Chicago: University of Chicago Press.
Winner, L. (1980). Do Artifacts Have Politics? Daedalus, 109(1), 121–136.