TL;DR
Paste 3,000 words into ChatGPT, say “translate this to Japanese,” and you get something fluent, grammatical, and tonally wrong about half the time: branded terms get translated, the author’s voice flattens, idioms turn into Wikipedia-sounding paraphrases. The fix is a four-step workflow — build a glossary, translate section by section with that glossary pinned, run a back-translation diff on suspect paragraphs, then have a native speaker read the result. This gets you roughly 90% of the way; the last 10% is human judgment.
What this covers
This is the workflow I use when ChatGPT output has to be publishable, not just intelligible. ChatGPT (GPT-5.5, default in ChatGPT since April 23, 2026) handles 58+ languages well and trades blows with DeepL and Gemini 3.1 Pro on raw fluency, but its real advantage over a one-click translator is control: you can pin terminology, anchor tone, and audit the result. None of that happens automatically — you have to drive it.
Who this is for
Anyone whose translation will be read by a native speaker who cares: a marketer localizing a newsletter, a researcher publishing in a second language, a founder writing investor updates for a non-English board. If the audience just needs the gist, paste-and-pray is fine. If a reader will spot a clumsy verb choice or a brand name rendered as a generic noun, you need a workflow.
Typical jobs where it pays off:
- Localizing marketing copy where tone and brand terms both matter.
- Translating a 1,000–5,000-word long-form piece while preserving a specific authorial voice.
- Producing bilingual documentation where terminology must stay consistent across many files.
- Domain text (legal, medical, technical) where a wrong term carries real cost.
Pick the model before you start
As of June 2026 the ChatGPT picker on Go, Plus, Pro, and Business is Instant / Thinking / Pro, not separate version numbers. For translation:
| Mode | Use it for | Notes |
|---|---|---|
| GPT-5.5 Instant | General marketing and editorial copy | Fast; fine for most prose |
| GPT-5.5 Thinking | Legal, medical, technical, or high-stakes copy | Slower, fewer silent meaning shifts; set thinking time to Extended for dense source |
| GPT-5.5 Pro | The hardest passages you plan to publish unedited | Pro tier only |
A practical limit to know: in-app context on Plus is roughly 320 pages of text; the full 1M-token window is only on the $200 Pro plan. Long documents still get translated more cleanly section by section than in one giant paste — the model holds tone better over short spans — so chunking is the right call regardless of your context budget.
Before the first prompt:
- List the terms that must NOT be translated (brand names, product SKUs, code identifiers, person names). Write them down — this becomes your locked glossary.
- Name the tone you want preserved (punchy, formal, conversational, marketing-glossy) and find a 100-word sample of that tone in the target language to anchor on.
- Open a Project for the language pair so the glossary and memory stay scoped, instead of polluting your general chat.
Step by step
1. Build a glossary from the source. Paste 500–1,000 words and ask:
Extract every term in this text that needs a deliberate translation choice:
brand names, product names, technical terms, idioms, recurring phrases.
Output as a table: source term | suggested translation | reason | leave-as-is (Y/N).2. Review the glossary by hand. Lock brand names as leave-as-is. Pick a single rendering for each repeated technical term. This is the step that prevents an inconsistency surfacing on page 6.
3. Translate one section at a time (500–1,500 words per turn), pasting the locked glossary at the top:
Translate the section below from EN to JA.
Glossary (use exactly): [paste table]
Tone: conversational, like a founder writing to early users.
Preserve paragraph breaks. Do not summarize or shorten.4. Run a back-translation check on suspect paragraphs:
Translate this JA paragraph back to EN literally, preserving structure.
Don't smooth it out — I want to see the drift.Compare to the original. Drift on factual nouns or numbers is a real problem. Drift on adjectives is usually fine.
5. For idioms or culturally loaded phrases, ask for three alternatives with a one-line note each, then pick:
Give me 3 ways to render "we shipped it warts and all" in JA marketing copy,
each with a one-line note on the connotation it carries.6. Final pass: a native speaker reads it. The workflow gets you to ~90%; the last 10% is human judgment on register and rhythm a model still misses.
A prompt that produces honest output
You are translating EN to [target language] for a published piece.
Constraints:
- Use the glossary I pasted. Don't invent new renderings for terms in it.
- Preserve the author's voice: [paste 100-word voice sample].
- If a phrase is ambiguous in the source, flag it inline as [AMBIGUOUS: ...]
instead of guessing.
- Do not paraphrase to shorten sentences. Match length where possible.
- Numbers, dates, and proper nouns must round-trip unchanged.The [AMBIGUOUS: ...] flag is the load-bearing line. It catches the cases where the model would otherwise silently smooth something into a half-meaning, which is exactly the kind of error back-translation can miss.
Quality check
- Spot-check three random paragraphs by back-translating literally. Watch factual nouns and numbers, not stylistic shifts.
- Run a find-replace pass on glossary terms — every locked term should appear in its locked rendering, no exceptions.
- Read the first and last paragraphs out loud in the target language. If they sound like a translation, the voice prompt didn’t take.
- Confirm brand names, URLs, and product SKUs survived intact. These are the most common silent edits.
ChatGPT vs DeepL vs Gemini for translation
No single tool wins everywhere. As of June 2026:
| ChatGPT (GPT-5.5) | DeepL | Gemini 3.1 Pro | |
|---|---|---|---|
| Languages | 58+ | ~33 | 100+ |
| Glossary control | Prompt-based, unlimited | Built-in: Free/Individual = 1 glossary / 5 entries; Advanced ($34.49/mo) = 2,000 entries; Ultimate ($68.99/mo) = 10,000 | Prompt-based |
| Tone / voice steering | Strong | Limited (Pro has tone presets) | Strong |
| Best for | Glossary + tone + domain instructions | Fast, clean first-pass on conventional text | Very long documents (1M-token context) |
| Free tier | Yes (tight limits) | Yes (1,500 chars/translation, ~50k chars/mo) | Yes |
DeepL is often cleaner for short, conventional text and ships a real glossary feature. ChatGPT wins when you need tone control, domain instructions, or reasoning about an ambiguous phrase. A common pattern: DeepL for the first pass, then ChatGPT for the polish and the awkward 10%. For documents too long for ChatGPT’s in-app context, Gemini 3.1 Pro’s 1M-token window can hold the whole thing at once — see ChatGPT vs Claude vs Gemini.
Reuse the workflow
- Save the per-language Project with its glossary; reuse it for every piece into that pair.
- Keep a
translation-voice-samples/folder with the 100-word anchor texts for each tone you commonly target. - For team use, share the glossary as a CSV so engineering, marketing, and support all use the same rendering. DeepL’s Team plan exposes shared glossaries natively if you standardize on it for first-pass work.
Common mistakes
- Pasting the whole document at once. Long passes lose tone consistency by the end; the model drifts toward its generic house style.
- Skipping the glossary step. Brand names get translated; technical terms get rendered three different ways across the doc.
- Asking for “natural” without a voice sample. “Natural” defaults to the model’s house style in that language, which is bland.
- Trusting back-translation to be lossless. It isn’t — but drift on factual content is a red flag worth investigating.
- Using Instant for legal or medical text. The stakes don’t match the tradeoff; switch to Thinking.
- Forgetting low-resource languages are weaker. ChatGPT’s training in some Southeast Asian and African languages is thinner; quality drops and back-translation drift rises. Bring in a human reviewer earlier.
FAQ
- Is ChatGPT better than DeepL for translation?: It depends on the text. DeepL is often cleaner for short, conventional copy and has a built-in glossary (up to 2,000 entries on its Advanced plan). ChatGPT wins when you need tone preservation, domain instructions, or handling of an ambiguous phrase. Many people use DeepL for the first pass and ChatGPT for the polish.
- Can I translate code comments and docs the same way?: Yes, but pin a glossary that explicitly says “leave code identifiers, variable names, and string literals unchanged.” Code-adjacent prose is where translators go wrong most often.
- Does the model preserve markdown and HTML?: Mostly, but it occasionally drops a tag or rebalances headings. Diff against the source after translating.
- Which model mode should I use?: GPT-5.5 Instant for general copy, GPT-5.5 Thinking (Extended thinking time) for legal, medical, or technical material. GPT-5.5 Pro only for the hardest passages you plan to ship unedited.
- What about voice mode for translation?: Useful for travel-grade conversation. Not for publishable output — there’s no glossary, no back-translation, no audit trail.