Why Politeness in a Prompt Changes Output More in Some Languages

“Does being polite to the model change the answer?” is usually asked about English, where the honest answer is that the effect is small and mostly stylistic. Asked about Japanese or Korean it is a different question with a different answer, because in those languages the model cannot produce a sentence without choosing a politeness level, and the strongest cue available is the level you used.

Obligatory versus optional politeness

The distinction that explains everything here is whether politeness is encoded in the grammar or in the vocabulary.

In English, politeness is lexical and optional. “Send me the report” and “Could you send me the report when you have a moment?” differ in word choice, hedging and length. The verb send is identical in both. You can write an English sentence without deciding how polite it is, because the neutral option exists.

In Japanese and Korean it is grammatical and obligatory. Every finite verb carries an ending that encodes the relationship to the addressee, and there is no ending that declines to encode it. There is no neutral Japanese sentence in the way there is a neutral English one — the plain form is not neutral, it is the intimate or written-impersonal form, and using it to a customer is a specific and noticeable choice.

A model generating in such a language therefore has to resolve a question on every sentence that English never asks. It resolves it from context, and your prompt is most of the context.

Japanese: the verb ending commits you

Japanese distinguishes at least three things people collapse into “politeness”. Teineigo, the polite style, is the です/ます ending — this is the axis most affected by prompt tone. Sonkeigo elevates the person being talked about; kenjōgo humbles the speaker. Together the last two are keigo proper, and they change the verb itself rather than only its ending: suru becomes nasaru or itasu depending on who is acting.

The observable effect is that a prompt written in plain form tends to produce output in plain form, and a prompt written in です/ます tends to produce output in です/ます. This is not the model being deferential. It is style matching, the same mechanism that makes a model answer a terse question tersely and a formal one formally — except that here the match lands on obligatory morphology, so it is visible in every sentence rather than in occasional word choices.

There are secondary effects worth knowing. Keigo is longer: the polite and honorific forms add syllables to every predicate, so the same content costs more characters and more tokens, which matters if you have a length limit — see what Japanese costs in tokens. And keigo is genuinely hard, including for native speakers; the honorific and humble forms are the part of the system most often produced incorrectly, and a model producing them confidently and wrongly is a failure a non-speaker on your team cannot see.

Korean: speech levels on every clause

Korean encodes the same information in a system with more levels and, if anything, less room to avoid the choice. The two that matter for most products are the deferential style, whose declarative ending is -습니다, and the polite informal style ending in -요. The plain style, banmal, ends in neither and is used with close friends, children and subordinates; addressing an adult customer in banmal is not a slight stylistic miss but an insult. There is also the written plain style used in newspapers and reports, which is impersonal rather than rude, and confusing it with banmal is a common non-native error.

Korean adds a second axis: the honorific infix -시- marks respect toward the sentence’s subject, independently of the addressee level. So a sentence can be deferential toward the reader and neutral toward the person it describes, or the reverse, and getting the combination wrong produces something a Korean reader notices immediately and a non-speaker cannot detect at all.

The practical upshot is the same as for Japanese and stronger: prompt register propagates, and a prompt written in casual Korean is asking for output no support product should send.

The English control

Run the same experiment in English and the difference is real but of a different kind. A curt English prompt yields a curt English answer: shorter sentences, fewer hedges, less preamble. The grammar is unchanged, because English grammar has nothing to change. There is no verb ending that becomes wrong.

That is why the widely repeated advice about being polite to models reads as folklore in English and as engineering in Japanese. In English the effect is a length and tone shift you might not even want to correct. In Japanese and Korean it is a morphological switch whose wrong setting makes your product read as rude to every user in that market, and it is deterministic enough to test with a regular expression.

Between those two poles sit the T-V languages — German Sie/du, French vous/tu, Spanish usted/tú, Russian вы/ты. There the choice is obligatory but binary and confined to pronouns and their agreement, so it is easier to get right and easier to check than keigo, and far more consequential than English tone. The term comes from Brown and Gilman’s 1960 paper The Pronouns of Power and Solidarity, which is still the reference for why the distinction patterns the way it does across European languages.

What this means for a product

Three conclusions follow, and none of them is “be polite to the model”.

• Do not let register be inherited. If the level is decided by the incidental tone of whatever prompt someone wrote, it will vary between features and drift with every prompt edit. State it explicitly and grammatically. The pattern is in specifying formal or informal register by language.
• Do not let the user’s register decide it either. Mirroring the customer is a defensible policy for English tone and a bad one for Japanese: a customer writing casually does not license your product to answer in plain form, and a support reply has a house register regardless.
• Test it mechanically. Sentence-final だ in a Japanese reply that should be polite, or a missing -요/-습니다 in Korean, is detectable without a native speaker on call. It is one of the few quality properties in this space you can assert in CI rather than review by hand.