If you have an app translated by DeepL or ChatGPT and a Japanese friend told you "it's fine but it doesn't really sound Japanese" — this article explains why, with concrete examples. It is also the article we send to teams who are about to ship MTPE output and are not sure that hiring a native writer is worth it.
For the broader picture, see the complete guide to Japanese app localization.
The Short Version
Modern MT produces Japanese sentences that are:
- Grammatically correct.
- Lexically plausible.
- Structurally English.
A native speaker reads "structurally English" Japanese the way you read a sentence written by someone who learned English from a textbook. Every word is right; the rhythm is wrong; trust drops.
For internal tools, this is fine. For consumer apps where retention depends on the first impression, it is the difference between an app that converts and an app that quietly does not.
Failure Mode 1: Subject Insertion
English requires an explicit subject. Japanese drops the subject when context is clear. MT keeps the subject.
❌ あなたはあなたのアカウントを削除することができます。 ✅ アカウントを削除できます。
The first line is correct Japanese. It is also exhausting. Two possessive pronouns ("you" → あなた, "your" → あなたの) appear in a sentence that, in natural Japanese, has zero pronouns.
The same problem at scale, in a settings screen:
❌ あなたはあなたの通知設定を変更することができます。あなたはまた、あなたのアカウントを削除することもできます。 ✅ 通知設定の変更や、アカウントの削除ができます。
The MT version is 50% longer, says nothing extra, and reads as alien. The native rewrite collapses two sentences into one and removes every redundant pronoun.
Failure Mode 2: Loanword Choice
Japanese has both native (kanji-based) and loanword (katakana) forms for many concepts. MT picks based on training-data frequency, not contextual fit.
| English | Common MT output | Natural app copy | Why |
|---|---|---|---|
| Settings | セッティング | 設定 | "セッティング" is correct but reads as amateurish in UI |
| Share | シェア | 共有 | "シェア" works in conversation; UI uses 共有 |
| Login | ログイン | ログイン | Loanword wins here |
| Photo | フォト | 写真 | "フォト" is brand-y; "写真" is the default |
| Update | アップデート | 更新 / アップデート | Either works depending on context |
| Calendar | カレンダー | カレンダー | Loanword is universal |
The pattern is not "always pick the kanji" or "always pick the loanword." It is "pick what shipping Japanese consumer apps in your category actually use." MT cannot do this without the corpus knowledge of the specific category.
Failure Mode 3: Register Drift
Japanese forces you to choose a register (politeness/formality level) for every sentence. MT chooses fresh for each sentence in isolation. The result drifts within a single screen.
A real onboarding screen we audited (paraphrased):
❌ ようこそ！はじめましょう。 アカウントを作成してください。 利用規約に同意します。 続けるには確認してね。
Read aloud:
- 「はじめましょう」 — polite-neutral.
- 「作成してください」 — polite-neutral, slightly stiffer.
- 「同意します」 — declarative, plain form.
- 「確認してね」 — casual.
Four sentences, three registers. To a Japanese reader this is instantly identifiable as "this app was translated, not written." The native rewrite picks one register and stays in it:
✅ ようこそ。さっそく始めましょう。 アカウントを作成してください。 利用規約に同意のうえ、次へ進んでください。
Failure Mode 4: Idiom Transplant
Marketing copy uses idioms. MT translates them word-for-word. The result is sometimes meaningless, sometimes hilarious, always non-native.
❌ ゲームを変える機能。 (literal: "function that changes the game") ✅ 使い方が変わる、新機能です。 ("A new feature that changes how you use [it].")
❌ 私たちはあなたをカバーしています。 (literal: "we cover you") ✅ あなたに必要なものはすべて揃っています。 ("Everything you need is here.")
❌ それは一石二鳥です。 ("kill two birds with one stone" — direct) ✅ ひとつで両方できます。 ("You can do both with one [thing].")
The third example is interesting because the Japanese idiom 一石二鳥 exists and is correct. But it is too literary for app copy. A native writer rewrites for register, not just for meaning.
Failure Mode 5: Politeness Inflation
MT, especially when prompted with "translate politely," over-applies honorifics. The result reads like a department store announcement.
❌ お客様におかれましては、こちらからご確認いただけますでしょうか。 ✅ こちらでご確認いただけます。
The first version uses three honorific markers (「お客様におかれましては」, 「ご確認」, 「いただけますでしょうか」). It is grammatically valid for very high-stakes business correspondence. In an app, it is suffocating.
The reverse problem also happens — MT undershoots politeness for finance/medical apps where keigo is expected. We cover register selection in keigo in app copy.
Failure Mode 6: Wrong Voice for the Brand
MT has no model of your brand voice. A meditation app and a payment app should sound radically different in Japanese. MT outputs both in the same default register.
A meditation app onboarding line:
❌ MT default: アプリへようこそ。プロフィールを設定してください。 ✅ Calm-brand: ようこそ。最初に、簡単な質問にお答えください。 ✅ Energetic-brand: ようこそ！まずはプロフィールから始めよう。
All three are correct. Only one matches a given brand. MT cannot pick.
Failure Mode 7: Length Blindness
Mobile UI has hard length constraints. Buttons fit ~6–10 Japanese characters; tab labels fit 4–6; navigation titles fit 12–16. MT outputs the most "complete" translation, which often overflows.
"Save" → ❌ 保存する → ✅ 保存 "Continue" → ❌ 続行する → ✅ 続ける / 次へ "Add to cart" → ❌ カートに追加する → ✅ カートに追加 "Sign up with Apple" → ❌ Appleでサインアップする → ✅ Appleで登録
The native writer optimizes for "what fits the button while still feeling natural." MT optimizes for "complete sentence." This is one of the highest-impact failure modes because it directly breaks layout.
Failure Mode 8: Cultural Reference Mismatch
App copy occasionally references Western cultural touchstones. MT translates the words; the reference does not land.
"Set it and forget it." → ❌ 設定して忘れてください。
Direct translation; the imperative "forget it" sounds dismissive. Native rewrite:
✅ 一度設定すれば、あとはおまかせ。
This kind of rewriting is the gap between translation and localization. AI is structurally bad at it because it solves "preserve meaning of input string" rather than "communicate the underlying intent to a Japanese audience."
Why Post-Editing Doesn't Fix This
The natural reaction is "fine, run MT and have a native editor clean up." This is called MTPE (Machine Translation Post-Editing). It works for European languages. For Japanese consumer apps, it doesn't, for two reasons:
-
Editors anchor. Given a draft to fix, editors fix what's wrong rather than rewrite from scratch. The English-shaped sentence structure survives. We measured this in a side-by-side test: MTPE output had 60% more sentences with explicit subjects than from-scratch native writing of the same source produced.
-
It's not faster. A good Japanese writer working from English can produce native output at roughly the same speed an editor can fix MT. The only thing MTPE saves you is the first read of the English source — minor.
When AI Is Useful for Japanese
We use AI in our workflow, just not as the translator. Useful applications:
- Glossary extraction. Pull all UI terms from the source codebase for translator briefing.
- Consistency checking. Flag where the same English string was translated differently across files.
- Quick first-draft for non-customer-facing content (analytics dashboards, internal tooling).
- Length-checking translated strings against UI constraints.
- Generating multiple variants for A/B testing.
The line is: AI is good at scaffolding around the translation. It is not good at being the translation for any string a paying user will read.
How to Tell If Your Existing Japanese Copy Is MT
Quick audit — read your app's Japanese copy aloud (or have a native speaker do it). If you hear any of these, you have an MT problem:
- Frequent 「あなた」 (you) and 「あなたの」 (your).
- Sentences ending with 「することができます」 instead of 「できます」.
- Mixed registers within the same screen (some 「ます」, some plain form).
- Loanwords where the kanji term is more natural for the category.
- Long, complete sentences on buttons that should be 2–4 characters.
- Direct translations of English idioms.
- Honorifics applied evenly across casual and formal contexts.
Three or more of these and the copy is not ready to ship.
Where to Go Next
- Keigo in Apps: When to Use Polite, Plain, or Casual Japanese
- Common Mistakes in Japanese App Localization
- How to Make App Copy Sound Native in Japanese
- Japanese UX Writing: Microcopy That Feels Native
If you want a quick read on how MT-influenced your existing Japanese copy is, send us your strings file and we will return a same-day audit identifying the failure modes above with line-numbers and suggested rewrites.