Using AI for bilingual writing without sounding like a machine
Literal accuracy is the first pass, not the last. A bilingual draft fails when it is correct and still sounds like nobody would say it.
Deni AI team
Why bilingual drafts still need a human
Deni AI ships in English and Japanese, so we see the same machine-voice problems visitors see. A model can move meaning across languages quickly. It is much worse at deciding how close the reader is, how much explanation they need, and which words should stay untranslated.
The tell is not a grammar error. It is a sentence that is technically right and socially wrong: an English paragraph that explains what the Japanese never said, or a Japanese sentence that sounds like a translated press release.
Pass one: meaning
Ask the model to preserve claims, numbers, names, and constraints. Tell it not to improve the argument yet. The job is transfer, not rewrite.
Pass two: voice
In a second prompt, ask only for tone: shorter sentences, natural particles, fewer nominalizations, or a less salesy English rhythm. Do not let it reopen the facts.
Pass three: a real reader
Read the result as the person who will receive it. If a phrase would feel stiff in Slack or too casual in a contract, fix that yourself. Models average the internet. Your audience is not the internet.
Keep a do-not-translate list
Product names, legal defined terms, and error strings often should stay in the source language. Put those in the prompt so the model does not invent a local equivalent.
Failures we keep seeing
Over-explaining. English drafts often add a helpful clause that was never in the Japanese. That extra clause can change a promise. If the source did not say it, the translation should not say it.
Register mismatch. A model will happily turn a blunt internal note into polite customer copy, or the reverse. Tell it the relationship: teammate, customer, lawyer, or public reader.
False friends in product language. Words like agent, memory, workspace, and free plan do not have one perfect pair. We keep a glossary and force the model to use it. Without that, every page slowly drifts.
Prompt skeleton we reuse
- Audience and relationship: who reads this, and how formal should it sound?
- Do-not-translate list: product names, legal terms, and error strings.
- Preserve: numbers, dates, claims, and anything that looks like a promise.
- Do not add explanations that are missing from the source.
- After the meaning pass, run a second prompt that may change rhythm only.
When to stop using the model
Legal pages, commercial disclosure, and anything that creates a customer obligation should be reviewed by a person who can read both languages. AI can propose a first pass. It cannot sign the policy.
If a sentence is doing social work — an apology, a refusal, a joke — write it yourself. Those sentences are short. The risk of a wrong tone is higher than the time you save.
Common questions
Can I translate a whole site in one prompt?
You can draft it that way. You should not ship it that way. Split by page purpose: UI chrome, legal text, and marketing copy fail in different ways.
Is Japanese to English harder than English to Japanese?
They fail differently. English often gets too long and explanatory. Japanese often gets too stiff or too casual for the relationship. Review for the failure you actually see.
Should I use the same model for both languages?
Use whatever model handles the source text faithfully, then do a voice pass. If two models disagree on a claim, check the source sentence instead of blending the translations.