Keigo, AI, and the Japanese Business Email: Why the Register Slips

TL;DR: Keigo AI email drafts fail in a specific way. The grammar is usually fine. The register is wrong: the level of formality, the amount of cushioning, the shape of the closing. AI treats politeness as a single dial and sets it to the average of every Japanese business email it has ever seen. Your reader notices in the first line. Telling the model to be "polite and professional" does not help, because that phrase describes thousands of writers and none of their actual habits. A voice lives in the concrete choices: which greeting, how many cushion phrases, which verb form closes the ask. Measure those from your own emails, hand them over as rules, and every future model executes them better than the last.

You ask ChatGPT for a follow-up email to a client in Japanese. It comes back with flawless 尊敬語, a textbook opening, and a closing you would never write. You send it anyway, and the client's reply is a shade cooler than usual.

Nothing in that draft was incorrect. Everything in it was somebody else's.

Why Does AI Get Keigo Wrong in Business Email?

Because keigo is two problems and AI solves only the first. The first problem is grammar: producing well-formed honorific, humble and polite forms, and a modern model does this reliably. The second problem is register: choosing how formal, how cushioned and how distant to be with this reader today. The model settles that by default, and the default is the statistical centre of all Japanese business writing it was trained on.

Japan's Council for Cultural Affairs laid out the five categories of keigo in its 2007 guidance, 敬語の指針: 尊敬語, 謙譲語Ⅰ, 謙譲語Ⅱ, 丁寧語 and 美化語. That is the grammar layer, and it is teachable. What the guidance cannot tell you, and what no model can derive from grammar, is whether you personally write ご確認ください or ご確認いただけますでしょうか to a supplier you have known for six years. Both are correct. Only one is yours.

Here is what an AI draft typically gets right and wrong on a single email:

Every row in the right column is a register failure, not a grammar failure. Spell-checkers and grammar tools will pass the draft. Your client will not.

What Actually Changes Between Registers in Japanese?

Four things, and each writer sets them independently: the politeness level itself, the amount of cushioning before an ask, the opening and closing formulas, and the sentence rhythm. AI collapses all four into one setting called "formal" and applies it uniformly.

Take a product manager in Osaka writing three emails in a morning. To the vendor she has worked with for two years, the opening is いつもお世話になっております, the ask arrives in the second sentence, and the closing is よろしくお願いします without the いたします. To a prospective client's director, the opening carries a seasonal line, the ask is wrapped in お忙しいところ恐縮ですが and ご検討いただけますと幸いです, and the closing takes the full 何卒よろしくお願い申し上げます. To her own team, in Japanese, the subject line is the ask, sentences end in です and ます, and there is no closing formula at all.

Same person. Three registers. None of them is "polite and professional" in a way an adjective could distinguish, and all three are polite and professional.

We wrote about the general version of this in why your writing voice changes when you switch languages. The Japanese case is sharper because the register machinery is grammaticalised. In English, a writer who is slightly too formal with a colleague sounds a little stiff. In Japanese, a writer who uses 謙譲語Ⅱ with a peer sounds like they are addressing a stranger, and the peer will wonder what changed.

The four moving parts, spelled out:

1. Politeness level. Not a spectrum with two ends but a set of discrete choices per verb. させていただきます versus いたします versus します is three distinct social positions, and most writers have a firm habit for each relationship.
2. Cushioning. How many クッション言葉 precede an ask, and which ones. Some writers use one, always the same one. Some stack three for a first contact and zero for a regular one. AI defaults to stacking.
3. Formulas. Opening and closing lines. The set is small and public, so the choice within it carries information. A writer who always closes internal mail with 以上です and client mail with 引き続きよろしくお願いいたします has a signature, and it is not in any adjective.
4. Rhythm. Sentence length, how often a sentence ends in a noun rather than a verb, whether the ask is its own paragraph. This is the part of voice that survives translation of the other three and the part AI flattens most thoroughly.

Our own measurements bear out the flattening. Across the 320-sample benchmark behind our per-language baselines, Japanese AI output sits at 59 on formality and 45 on conciseness, with the lowest vocabulary richness of the four languages we measured at 37. Those are averages, and an average is the problem: the model does not write at your formality level, it writes at 59. (We withdrew our Japanese expressiveness figure after finding a measurement defect, and we are not citing it here.)

Why Doesn't "Polite and Professional" Fix It?

Because the phrase is a pointer to a register the model already knows, and what it points to is the centre of everything written under that label. Type 丁寧でプロフェッショナルに, or "polite, professional, appropriate for Japanese business", and the model retrieves its average of Japanese business correspondence: the same average every other user of that phrase receives. Your colleague who types the same words gets the same email. You two write nothing alike.

AI does not know your tone. It knows words that describe tone.

This is the same failure we took apart for English in why tone adjectives are not your voice, and Japanese makes it worse in two ways. First, the register vocabulary is richer, so the adjective covers more ground and constrains less. "Polite" in Japanese spans everything from a Slack message to a formal apology letter. Second, the reader's sensitivity is higher. Japanese business readers read register as information about the relationship, so a mismatch is not a style nit. It is a message you did not mean to send.

The common workarounds do not close the gap either:

• Uploading past emails as reference files. The model retrieves them when it judges them relevant and paraphrases the content. It does not absorb the habits. Five emails cannot tell it whether your single cushion phrase is a rule or a mood.
• Asking for "the same tone as my last email". Works for one draft, evaporates in the next conversation, and copies the surface of that one email including whatever was unusual about it.
• Writing in English and asking for a Japanese translation. Produces English directness in Japanese grammar. The ask arrives too early, the cushioning is missing, and the result reads as abrupt to any Japanese colleague. The reverse direction, AI English that sounds translated, is the same failure with the languages swapped.
• Switching to a newer model. A more capable model follows instructions more faithfully. Give it an adjective and it produces a more polished rendering of the same average. The draft reads better and sounds no more like you.

That last point deserves its own section, because it is the one people bet on.

Will a Better Model Learn My Japanese Register?

No, and the reason is structural. Model capability and user-data conditioning are separate layers. The lab improves the first with every release. It cannot supply the second, because the second is made of emails you sent and the lab has never seen them.

What a stronger model delivers is fidelity to instructions. Feed it "polite and professional" and it executes that instruction more thoroughly than the last model did: smoother keigo, more consistent level, cleaner formulas. It has moved closer to the average, not closer to you. Readers who report that their Japanese drafts got "more generic" after a model upgrade are describing exactly this.

The same property works in your favour the moment the instruction carries something precise. A profile that says "open client mail with いつもお世話になっております, never the seasonal greeting; one cushion phrase before an ask, always お手数ですが; close with よろしくお願いいたします, never 申し上げます; ask in the second sentence" is a set of rules a stronger model hits more accurately than a weaker one. The better the model gets at following nuanced instructions, the better it executes your profile. We laid out the general argument in why better AI models still don't know your voice. Japanese is the case where the argument bites hardest, because the gap between "polite" and your polite is the widest.

So the model dropdown is the last thing to change. Fix what the model is being asked to execute, and every future release executes it better.

How Do You Give AI Your Own Japanese Register?

Measure it from emails you have already sent, then hand the measurements over as rules. Your register is countable. Every one of the four moving parts above can be read off ten to fifteen emails in an hour, and each one becomes an instruction with no adjective in it.

Start by separating your samples. Japanese in one folder, English in another, and within Japanese, label each email by relationship: long-standing client, new contact, internal, supplier, apology. Do not pool them. A profile averaged across your client mail and your team mail is a profile of nobody, which is the mistake you are trying to escape.

Then count, per register:

Level. For each ask, which verb form did you use? Tally いたします, させていただきます, します. For each request, ください or いただけますでしょうか or いただけますと幸いです. You will find a firm habit per relationship, and a habit is a rule.

Cushioning. Count クッション言葉 per ask. Note which ones. Most writers have one or two favourites and a maximum they never exceed. Write the maximum down as a hard ceiling.

Formulas. List your openings and closings by register. Note what you never use. Absence is invisible in a sample and has to be stated: "never 拝啓, never 何卒 in internal mail."

Rhythm. Words or characters per sentence, typical and range. Where the ask sits (first sentence, second, its own paragraph). Whether you use bullet lists in client mail. How often a sentence ends in a noun.

A finished block for one register looks like this:

Register: existing client, weekly contact. Open with いつもお世話になっております, no seasonal line. Ask in sentence two. One cushion phrase, お手数ですが or 恐れ入りますが, never stacked. Requests as いただけますでしょうか. Verb level いたします; never させていただきます. Sentences 25 to 45 characters, one idea each. Close with 引き続きよろしくお願いいたします. Never 何卒, never 申し上げます.

Paste that where your instructions live: ChatGPT custom instructions, a Claude Project, a Gemini Gem, a Custom GPT's instructions field. Keep one block per register and tell the model which one applies. Our guide to connecting your writing style to ChatGPT covers where each block goes and how to keep it from drifting in a long conversation.

This is also what a Style Profile from MyWritingTwin.com does, per language. A multilingual user's Japanese samples get their own Writing DNA analysis, measured against a Japanese baseline rather than an English one, across more dimensions than the four above. The output is a block of constraints, and your Writing Twin for ChatGPT, Claude and Gemini is that block deployed inside the tools you already use. Your English samples get a separate profile. Neither one is a translation of the other, because your two voices are not translations of each other.

Two properties of a measured profile matter here more than anywhere else. It can be checked: "was that polite enough?" has no answer, and "was there one cushion phrase, and was it お手数ですが?" does. And it outlives the model, because the rules describe you rather than the tool.

How Do You Test Whether the Draft Is Yours?

Give the model the brief for a Japanese email you already sent, put its draft beside yours, and count. No tools, five minutes, and the answer is a set of numbers rather than a feeling.

Pick one email per register from the last month. Write a two-line brief for each: who it was to, what you needed, what they already knew. Ask the model for the email in Japanese and compare:

• Opening line: same formula, or the textbook one?
• Ask position: which sentence?
• Cushion phrases: how many, and which?
• Verb level on the ask: いたします, させていただきます, or します?
• Closing: yours, or よろしくお願いいたします by default?
• Characters per sentence: typical, shortest, longest

If the counts differ, the model does not know your register, whatever it says when you ask it directly (it will say yes; it always says yes). Adjective-based setups usually pass the formal-client email, because the average is closest to that register, and fail the internal and long-standing-client ones, because those are where your habits diverge from the centre.

Run the same test after replacing the adjective with the measured block. The counts converge on the first draft and keep converging as you tighten single rules. That convergence is the difference between an email with your signature on it and an email with you in it.

The keigo was never the problem. The register was, and the register was never handed over.

See how far your AI drafts sit from your real writing with the Humanity Score test, or upload your Japanese samples and let a Style Profile do the counting.