Using the Correct Honorific and Title in AI-Generated Text

English has almost no grammatical honorific system, so an English-mediated translation has nowhere to put one. The result is not a slightly less polite sentence. It is a sentence from which a fact about the relationship between two people has been deleted, and the reader cannot tell that anything is missing.

An honorific is information, not decoration

The habit of thinking about honorifics as politeness markers — a dial from casual to formal — is the source of most of the errors here. In the systems below, the choice of honorific encodes specific, recoverable facts: the relative seniority of speaker and referent, whether the referent is inside or outside the speaker’s group, the setting, and sometimes the speaker’s own stance toward what they are saying. Those are propositions. A translation that renders them all as neutral English has dropped propositions, not tone.

This is why it matters for generated text specifically. A model translating into English has no grammatical slot for the information and drops it, which is defensible. A model translating out of English into Japanese or Korean has to invent it, because the target language has no neutral option — every sentence commits to a level. The model will pick one from context, that choice will be invisible to a monolingual English reviewer, and it can be badly wrong in a way that reads as rude or sarcastic rather than as a mistake.

Japanese: suffixes and three kinds of keigo

The visible layer is the name suffix. さん (-san) is the general respectful default. 様 (-sama) is markedly higher and is what a business writes to a customer, and what goes on an envelope. 君 (-kun) and ちゃん (-chan) mark familiarity and carry age and gender implications. 先生 (-sensei) attaches to teachers, doctors, lawyers and other licensed professionals; 先輩 (-senpai) marks someone senior to you in the same institution. Using no suffix at all — 呼び捨て, yobisute — is itself a marked choice indicating intimacy or, in the wrong context, contempt.

Underneath that is keigo, which is not a suffix system but a restructuring of the verb, and it has three branches that a flattened translation conflates. 尊敬語 (sonkeigo) elevates the person being spoken about. 謙譲語 (kenjōgo) lowers the speaker and their group, which elevates the listener by contrast. 丁寧語 (teineigo) is the ordinary polite register marked by the -masu and desu forms. The same action takes a different verb in each: what you do is suru, what an honoured person does is nasaru, and what you do for them is itasu.

The consequence for business writing is the in-group / out-group rule, and it is the one most reliably got wrong by generated Japanese. When speaking to a customer about your own colleague, you use humble forms for the colleague and drop the honorific from their name — even if that colleague is your boss. A model that has learned “address seniors respectfully” will attach 部長 or 様 to its own side and produce a sentence that a Japanese business reader will immediately mark as written by someone outside the culture.

Korean: the level is in the verb

Korean has suffixes too — 씨 (-ssi) as a neutral respectful form, 님 (-nim) as a higher one attached to names and to titles — but the load-bearing part of the system is in the verb, and it is obligatory. Every Korean sentence is inflected for a speech level, so there is no way to write a sentence without choosing one.

Two mechanisms operate at once and they are independent. The speech level is about the listener: 합쇼체 (formal, deferential — the register of a news broadcast or a formal announcement), 해요체 (polite, the everyday courteous default), 해체 (plain, used with close friends and juniors). The subject honorific is about the person being talked about, and is marked by the infix 시 (-si-) inserted into the verb stem. You can speak politely to a listener about a subject you do not honorify, or the reverse, and the two dials are set separately.

There is a further layer of suppletive vocabulary: honorific nouns and verbs that replace their plain equivalents entirely. 밥 (rice, meal) becomes 진지; 있다 (to be) becomes 계시다; 먹다 (to eat) becomes 드시다 or 잡수시다. A model generating Korean that inflects the verb correctly but keeps the plain noun produces a mismatch that reads as inconsistent rather than as merely informal, and this mixed-register output is a characteristic artefact of generated Korean.

German: a pronoun and a title that is part of a name

German has a much smaller system and it still defeats generated text regularly, which is useful because it shows the problem is not about how elaborate the system is. The core is the T-V distinction: du for informal address, Sie for formal, with the formal form capitalised and taking third-person plural verb agreement. There is no neutral option. Every German sentence addressing a person picks one, and the wrong pick is not a register error but a social claim.

The switch between them is itself an event with a name — duzen and siezen as verbs, and the offer of the du being conventionally the senior person’s to make. A model rewriting a German email “more warmly” may switch a customer from Sie to du as part of the rewrite, which reads as a presumption rather than as warmth.

The second German feature is academic title. In German-speaking countries a doctorate is conventionally used in address and in written correspondence — Sehr geehrte Frau Dr. Müller — and in Austria in particular the use of titles is notably more thorough. Dropping the title from a salutation is not neutral; it is a small deliberate-looking omission. The correct salutation form varies by language in ways that salutation format by language takes separately, and it is worth treating as structured data rather than as prose the model composes.

What the flattened translation loses

Take a Japanese sentence in which a speaker uses humble forms for their own team and 様 for the customer. Translated to English it becomes a plain declarative. What has gone is: which side of the transaction the speaker is on, that this is a customer-facing communication rather than an internal one, and the relative seniority the speaker is asserting. If that English is then translated onward — into Korean, say — none of it can be recovered, and the Korean will be generated at whatever level the model guesses from surrounding cues. This is the pivot-language problem: information that has no representation in the pivot is lost irrecoverably, not degraded.

The practical consequences for anyone building on this:

• Do not round-trip through English. Where a target language has an obligatory honorific system, translate from the source directly. A pivot through a language with no such system guarantees the level is guessed rather than carried.
• Make the level an explicit parameter. If you are generating Japanese or Korean, state the relationship in the prompt — who is speaking, to whom, in what setting, and which side is the in-group — rather than asking for “polite” output. Polite is not a setting the language has.
• Keep honorific suffixes attached to names as data. If a name is stored with 様 and the template adds its own suffix, you get a doubled honorific, which is a visible error. This is the same argument for treating names as structured fields made in why name order gets reordered.
• Have a native speaker review register specifically. Grammatical correctness and register correctness are separate checks, and a reviewer asked whether the text is “correct” will usually answer the first question.