Your colleague translates a 10-page contract in 40 minutes - same quality as yours. You spend 3 hours on the same volume. The difference? They use ChatGPT as a first draft and only spend time on editing. No, it’s not cheating - it’s the new standard. According to a 2026 Slator survey, 88% of translators already use AI in their workflow.
But there’s a gap between “throw text into ChatGPT” and “get a quality translation.” Let’s figure out how to actually use ChatGPT and Claude for document translation, where they help, and where they’ll let you down.
ChatGPT vs Claude: who translates better¶
In an independent 2026 test (200 sentences, 8 language pairs), Claude scored 8.3 out of 10 while ChatGPT scored 7.9. But there’s more behind these numbers.
Claude is stronger at: - Preserving tone and style (especially legal and business text) - Idioms and expressions - ChatGPT translated literally 34% of the time, Claude only 8% - Long documents - Claude’s context window (up to 200K tokens) lets you load an entire document without chunking - Terminology consistency within a single text
ChatGPT is stronger at: - Technical documentation and formatting (API docs, README files) - 8.2 vs 7.8 for Claude - Number of supported languages (including rare pairs like Ukrainian-Thai) - Response speed - Integrations with other tools
For DE-EN and DE-UK pairs, both work well, but Claude more often produces natural-sounding target text instead of calquing from the source.
Which documents can you translate with AI¶
Not all documents are equal. Here’s an honest breakdown:
| Document type | ChatGPT/Claude | Human review needed? |
|---|---|---|
| Letter from Finanzamt / Krankenkasse | Excellent | Minimal |
| Product manual | Good | Check terminology |
| Business correspondence | Good | Check tone |
| Academic text, article | Good | Check terms and style |
| Employment contract (Arbeitsvertrag) | Average | Mandatory |
| Medical report | Average | Mandatory |
| Court documents, notarial deeds | Risky | Full review mandatory |
The rule is simple: the higher the legal consequences of a mistake, the less you can rely on AI without review.
Reuters reported cases where immigration applications translated via GPT had names turned into month names, entire paragraphs disappeared, and in several places the translation said the exact opposite of the original. This isn’t a bug - it’s the reality of working with LLMs without quality control.
Prompts that actually work¶
Translation quality is 70% about how you phrase your request. “Translate this text” is the worst possible prompt.
Basic document prompt¶
Translate this text from [German] to [English].
Document type: [employment contract / medical report / official letter].
Preserve legal terminology precisely.
Don't simplify or paraphrase - the translation must match the original in content and structure.
Leave unfamiliar terms in the source language in brackets.
Translator’s prompt (MTPE workflow)¶
You are a professional DE>EN translator with 10 years of legal translation experience.
Translate the text below following these rules:
1. Preserve the structure and numbering of the original
2. Translate legal terms according to common law equivalents, add the German original in brackets
3. Don't add explanations of your own
4. If a term has no established English equivalent - leave it in German and mark with [?]
5. Avoid calquing from German - sentences should sound natural in English
Glossary for this text:
- Arbeitnehmer = employee
- Arbeitgeber = employer
- Kündigung = termination of contract
- Probezeit = probationary period
Few-shot prompt¶
Translate from German to English in legal document style.
Example of correct translation:
DE: "Der Arbeitnehmer hat Anspruch auf bezahlten Erholungsurlaub von mindestens 24 Werktagen."
EN: "The employee is entitled to paid annual leave of at least 24 working days."
Now translate:
[your text]
The key trick is the glossary. When you add a list of terms that should be translated in a specific way, quality jumps dramatically. Works in both ChatGPT and Claude.
Pricing: how much does AI translation cost¶
Let’s compare the cost of translating one page of text (roughly 1,800 characters or 250-300 words):
| Service | Cost per page | Notes |
|---|---|---|
| ChatGPT Plus ($20/mo) | ~$0.02-0.05 | Unlimited chat messages |
| ChatGPT API (GPT-4o) | ~$0.01-0.03 | $2.50 per 1M input + $10 per 1M output tokens |
| Claude Pro ($20/mo) | ~$0.02-0.05 | Unlimited chat messages |
| Claude API (Sonnet) | ~$0.01-0.02 | $3 per 1M input + $15 per 1M output tokens |
| Translation agency (DE-EN) | €25-60/page | Includes certification |
| Freelancer (DE-EN) | €15-35/page | Without certification |
AI translation is 500-1,000x cheaper than human translation. But you still spend time on review and editing. The real time savings with MTPE (AI draft + human editing) are 30-50%, not 100%.
5 mistakes AI makes regularly¶
After working with hundreds of translations through ChatGPT and Claude, here are the typical fails:
1. “Invents” term translations. Instead of leaving an unfamiliar term in the source language or flagging uncertainty, AI delivers a confident (and wrong) answer. The legal term “Eigenbedarfskündigung” might turn into something odd instead of the correct “termination of lease due to landlord’s personal use.”
2. Skips entire sentences. Especially in long texts. ChatGPT sometimes just “forgets” to translate a paragraph or merges two sentences into one, losing meaning.
3. Changes numbers. “24 working days” can become “20 working days.” Always check figures, dates, amounts - AI doesn’t understand that “30 Werktage” and “30 days” are legally different things.
4. Mixes formal and casual registers. One paragraph might have “in accordance with applicable legislation” and the next - “if you want to complain about it.”
5. Generates “confident nonsense.” AI never says “I don’t know.” It’ll produce a translation even if the source is unclear or corrupted. This is the most dangerous trait for legal texts.
Working workflow: AI + human¶
Here’s what an effective process looks like for translators who’ve already integrated AI into their work:
- Preparation - identify document type, compile a glossary of key terms, write your prompt
- AI draft - load the text into ChatGPT or Claude with your prepared prompt
- First pass - go through the text and flag problem areas (terms, numbers, omissions)
- Editing - fix errors, adapt style, check terminology consistency
- Final review - compare with original paragraph by paragraph
This approach is called MTPE (Machine Translation Post-Editing) - and it’s becoming the industry standard. 84% of translators expect demand for MTPE services to grow.
Want to automate even more? ChatsControl has a built-in AI critic that reviews translations 2-3 times before delivering the result. Steps 3 and 4 essentially happen automatically.
What AI can’t do (and won’t anytime soon)¶
Let’s be honest - there are things where AI won’t replace a human:
- Certified translation for Germany. No institution will accept a translation from ChatGPT. You need a sworn translator with an official seal and signature
- Translation where mistakes have legal consequences. Contracts, patents, court documents - the cost of error is too high
- Cultural adaptation. AI doesn’t know that in a Ukrainian court you say “Your Honor” while in a German court it’s “Herr Vorsitzender”
- Legal compliance verification. AI doesn’t check whether a term matches the target country’s legal framework - it just translates words
More on why machine translation doesn’t work for legal documents - a separate article with real examples of errors.
ChatGPT or Claude: which one to pick for translation¶
| Criteria | ChatGPT (GPT-4o / GPT-5) | Claude (Sonnet / Opus) |
|---|---|---|
| Translation quality (EU languages) | 7.9/10 | 8.3/10 |
| Idiom preservation | 66% | 92% |
| Long documents (50+ pages) | Needs chunking | Works as-is |
| Technical documentation | Slightly better | Good |
| Price (subscription) | $20/mo | $20/mo |
| Price (API, Sonnet/4o) | ~$0.01-0.03/page | ~$0.01-0.02/page |
| Language count | 50+ | 30+ |
Practical advice: use both. Claude for first drafts of legal and business texts. ChatGPT for technical documents and when you need a rare language pair. Compare results on the first few pages and stick with whichever gives a better draft.
If you don’t want to deal with prompts and APIs - ChatsControl handles it for you: upload a .docx, get a translation with preserved formatting and built-in quality checks.
FAQ¶
Can I use ChatGPT to translate official documents?¶
As a draft for further review by a translator - yes. As a final translation for submission to authorities - no. German institutions require certified translation from a sworn translator. Machine translation has no legal standing.
What’s the best prompt for ChatGPT translation?¶
Specify the language pair, document type, style (legal, business, casual) and add a glossary of key terms. The more context you provide, the better the result. A prompt like “translate this text” gives mediocre output, while a prompt with a role, rules, and examples produces significantly better results.
Is ChatGPT or Claude better for German-to-English translation?¶
In 2026 tests, Claude showed better results for European language pairs (8.3 vs 7.9 out of 10). Claude is better at preserving idioms, tone, and legal terminology. ChatGPT wins on speed and number of supported languages. For DE-EN, we recommend trying both and comparing on your specific text type.
How much does translation through ChatGPT API cost?¶
GPT-4o costs $2.50 per million input tokens and $10 per million output tokens. One page of text (250-300 words) costs roughly $0.01-0.03. For comparison, the same page at a translation agency costs €25-60. A ChatGPT Plus subscription ($20/month) gives unlimited access via chat.
Is it safe to upload confidential documents to ChatGPT?¶
OpenAI stores chat data and may use it for model training (unless you disable this in settings). For confidential documents, use the API - data sent through the API isn’t used for training. Or disable chat history saving in ChatGPT settings. Claude has a similar policy - API data isn’t stored for training purposes.