A pharmaceutical company translated a licensing agreement with a machine translator to cut costs. One mistranslated term changed the scope of the license - and the company spent months in a legal dispute that cost tens of times more than a professional translation would have. This isn’t an isolated case - a 2025 study comparing human and AI translations of legal texts found that human translations significantly outperformed AI across all quality criteria.
But does that mean machine translation of legal documents is completely useless? No. The answer is more nuanced and depends on what you’re translating, why, and what happens if there’s an error. Let’s figure out when you can trust AI translation and when you absolutely can’t.
Trust levels: from “fine, good enough” to “don’t even think about it”¶
The biggest mistake is treating machine translation as all-or-nothing. There’s a clear spectrum, and understanding it will save you both money and headaches.
Green zone: trust is fine¶
Getting the gist of a document. You received a 30-page Mietvertrag (rental agreement) from a German landlord and want to understand what it says before spending money on a professional translation? Throw it into DeepL or ChatGPT - for understanding the general structure and key terms, that’s enough. You’ll grasp what the contract is about, the timelines, and the amounts involved.
Internal communication. Need to quickly explain the contents of a letter from a German partner to your colleague or manager? Machine translation handles this. An accuracy of “about 80%” is acceptable here, because you can always follow up and clarify.
Pre-negotiation analysis. Want to prepare for a meeting and understand the other party’s position from their document? Machine translation gives you the big picture.
Yellow zone: trust with limitations¶
MTPE (machine translation post-editing). An experienced legal translator takes the machine translation as a draft and carefully reviews it. Research shows MTPE cuts working time by up to 63% - but only when the editor understands legal terminology in both legal systems. A junior translator might miss that “shall” was translated as “can” instead of “is obligated to” - and that error stays in the final text.
Notes for your lawyer. If you’re translating a document to show your attorney and discuss strategy - that’s acceptable, but make sure to tell them it’s a machine translation. They need to know the terms might not be 100% accurate.
Red zone: never trust¶
Court documents. A US federal court has already declared Google Translate unreliable even for obtaining simple consent from a person. Courts in Germany, France, and Argentina reject machine translations as inadmissible evidence.
Contracts for signing. One mistranslated word can turn a legal obligation into a voluntary action. We covered specific terminology traps in German contracts in detail.
Visa and immigration documents. An inaccurate translation in a visa application can cost months of waiting and resubmission. One emigrant shared on a forum: “Saved 200 euros on translation, used Google Translate. Got rejected, lost 3 months, and paid for a proper translation anyway.”
Notarial or sworn certification. No notary or sworn translator will put their seal on a machine translation. And most official procedures in Germany require exactly that - a certified translation.
Google Translate vs DeepL vs ChatGPT: which is better for legal texts¶
Not all AI translators are equal - and the difference matters a lot for legal texts.
Google Translate¶
Supports 249 languages, but it’s the weakest option for legal texts. It regularly confuses modal verbs (shall/may/must), which have fundamentally different meanings in contracts. The free version stores and analyzes everything you input - a direct confidentiality violation for legal documents.
DeepL¶
According to the Intento 2025 benchmark, DeepL outperforms Google Translate in 65% of language pairs, especially for European business texts (German, French, Spanish). It handles formatting better - when tested with a 20-page French contract, DeepL preserved tables, headers, footnotes, and even signature blocks. DeepL Pro doesn’t store your texts on its servers, which matters for confidentiality.
But legal terminology accuracy is still an issue - research shows only 53% of legal text translations through DeepL are fully accurate.
ChatGPT and Claude¶
LLMs (large language models) are a new class of tools that work differently from Google Translate or DeepL. They understand context better and can account for legal specifics if you write the right prompt.
In Lokalise’s 2025 evaluation, Claude scored highest - 78% of its translations were rated “good.” ChatGPT, when translating a legal contract, mixed formal and colloquial registers and used an informal term for “liability” (Haftung) that a German lawyer would immediately flag. Claude correctly used formal legal terminology in the same test.
But even the best LLM doesn’t replace a human translator for legal documents - it makes a better first draft for subsequent post-editing.
| Tool | Legal term accuracy | Confidentiality | Legal standing | Best for |
|---|---|---|---|---|
| Google Translate | Low | Data stored | None | Quick overview |
| DeepL Free | Medium | Data stored | None | Overview, drafts |
| DeepL Pro | Medium | Data not stored | None | Drafts for MTPE |
| ChatGPT | Medium-high | Depends on plan | None | Contextual drafts |
| Claude | Highest among AI | Not stored on free tier | None | Drafts for MTPE |
| Professional translator | 95-99% | NDA | Yes (if sworn) | Official documents |
Why AI fails specifically with legal texts¶
Legal texts aren’t just “difficult texts.” They have three characteristics that AI still can’t handle reliably.
Modal verbs are everything¶
In everyday language, “shall,” “may,” “must” are nearly interchangeable. In a legal text, they represent three fundamentally different levels of obligation. “The Seller shall compensate” - the seller is obligated to compensate. “The Seller may compensate” - the seller can but doesn’t have to. Research found 24 instances of incorrectly translated modal verbs in a single set of contracts run through Google Translate.
In German, it’s even trickier: “muss” (obligated), “soll” (should), “kann” (can), “darf” (permitted) - four levels instead of three, each with its own legal weight.
Chameleon words¶
“Consideration” in everyday English means “thinking about something.” In contract law, it means “something of value exchanged between parties.” In one international fraud case, the mistranslation of this exact term led to a claim being dismissed.
“Default” isn’t “the standard setting” - it’s “failure to fulfill obligations” or “a judgment entered when the defendant doesn’t appear.” “Gesellschaft” isn’t “society” - it’s “company” in a business context. Machine translators pick the most common meaning and miss the mark.
Legal systems don’t convert¶
A GmbH isn’t an LLC. Yes, both translate to “limited liability company,” but minimum share capital, management rules, and founder liability are fundamentally different. Translating an employment contract (Arbeitsvertrag) requires understanding both legal systems, not just swapping words.
Confidentiality: the hidden cost of free translation¶
When you paste contract text into a free online translator, it ends up on the company’s servers. And this isn’t a theoretical risk.
In 2017, Norwegian oil company Statoil (now Equinor) used the free service Translate.com - and confidential documents became publicly accessible. Google explicitly states in its terms of service that it can use submitted content to improve its services.
For legal documents, this is a double hit. First, you’re violating client confidentiality. Second, if the document contains personal data (names, addresses, identification numbers), you’re violating GDPR - and that means fines up to 20 million euros or 4% of annual global turnover.
What about paid versions?¶
DeepL Pro and Claude API don’t store texts and don’t use them for model training. ChatGPT Team and Enterprise also offer confidentiality guarantees. But the accuracy problem doesn’t go away - you just get the same errors, only confidentially.
ISO 17100: why “I checked it with AI” doesn’t count¶
There’s an international translation quality standard - ISO 17100. It explicitly excludes machine translation from its scope: raw machine translation output plus post-editing is not translation under ISO 17100.
For machine translation with post-editing, there’s a separate standard - ISO 18587. And it requires the post-editor to hold translator qualifications (a relevant degree or at least 2-5 years of professional experience).
Why does this matter? If your document is needed for official purposes, the receiving institution may require a translation done according to ISO 17100. Machine translation, even edited, doesn’t qualify under this standard.
Practical algorithm: trust or not¶
Before translating a legal document with AI, ask yourself three questions:
1. What are the consequences of an error? - I just won’t understand one clause of the contract → green zone, AI is fine - I might be misunderstood in negotiations → yellow zone, AI + review - The document goes to court/embassy/notary → red zone, professional only
2. Who will read the translation? - Only me → green zone - My lawyer or colleague → yellow zone - Court, embassy, counterparty → red zone
3. Does the document contain confidential data? - No → use any tool - Yes → paid versions with privacy guarantees only, or a service with data protection
If even one answer falls in the red zone - don’t risk it. Professional translation costs start at 30-60 euros per page. A legal dispute over a mistranslation can cost thousands to millions.
What to do when you need a legal translation¶
For your own understanding. Use DeepL Pro or Claude - they’ll give you a solid grasp of the document’s contents. Just remember: it’s a draft, not a finished translation.
For working with your lawyer. Do the AI translation, but make sure to note that it’s machine-translated. Your attorney needs to know the terms might not be 100% accurate.
For official submission. No shortcuts here - you need a sworn translator (for Germany) or a notarized translation (for Ukraine). You can find sworn translators in the justiz-dolmetscher.de database.
For quick document translation with preserved formatting - ChatsControl translates your .docx or .pdf in minutes. For preliminary review and internal communication - it’s the perfect fit. For official legal documents that require certification - use the certified translation service.
FAQ¶
Can I trust ChatGPT to translate a contract?¶
For understanding the general content - yes, ChatGPT and Claude are actually better than Google Translate because they account for context. But for signing or official submission - no. No AI translation has legal standing, and courts don’t accept them as official documents. Use AI as a first step, then bring in a professional translator.
How is DeepL better than Google Translate for legal texts?¶
DeepL is more accurate for European languages (outperforms Google in 65% of language pairs per the Intento 2025 benchmark), better preserves document formatting, and the paid version doesn’t store your texts. But even DeepL achieves only 53% fully accurate translations of legal texts - meaning nearly half need corrections.
How much does professional legal translation cost compared to AI?¶
AI translation is free or starts at $20/month for a paid version. Professional legal translation runs 40-80 euros per page in Germany, 300-500 UAH per page in Ukraine. The price gap seems large, but the pharmaceutical company that saved on translating a licensing agreement lost months to legal disputes. A fine for an error in one regulatory document ran to a million dollars.
What is MTPE and is it a good middle ground for legal documents?¶
MTPE (Machine Translation Post-Editing) is when AI produces the first draft and a professional translator carefully reviews and corrects it. For legal texts, this works only if the post-editor has legal specialization and meets ISO 18587 requirements. MTPE cuts time by up to 63%, but responsibility for quality lies with the human, not the machine.
Does Germany accept translations done by AI?¶
For official procedures (visa, court, notary, Anerkennung) - no. You need a translation certified with the signature and seal of a sworn translator (beeidigter Übersetzer). Machine translation has no legal standing regardless of how accurate it is. For internal use or personal review - there are no restrictions.