AI Legal Translation: What Works, What Needs a Human, and How to Prompt It

The most-used AI tool in Bavarian law firms is not ChatGPT. In the Bayerischer AnwaltVerband survey of 558 lawyers, most in firms of one to ten people, DeepL came first at 70.7% of AI users and ChatGPT second at 69.2%. Only 37.9% of firms used AI at all; those who did were translating.

That is the honest picture of AI legal translation in Europe: it is the entry drug. Lawyers who would never paste a Schriftsatz into a chatbot have pasted counterparty letters into a translator for years. The habit escapes most firm policies and should not escape the discipline: a translation tool is a language model with a narrower interface. It trains or it does not, it invents where it lacks an equivalent, and it will render your defined terms three different ways across forty pages unless you stop it.

Translation is the entry drug: what the DACH numbers show

The Bavarian figures are not an outlier. In the FTI Consulting and Relativity General Counsel Report for 2026, 40% of legal departments were using or experimenting with AI for analysing foreign-language materials, behind summarisation (83%), clause identification (63%) and transcription (53%). A&O Shearman’s Harvey customer page reports staff saving two to three hours a week on summarisation and translation (a vendor page, but consistent with the surveys). Greg Siskind, listing ten AI uses for RFP responses in ABA Law Practice, makes translation the tenth. It leads because the input is complete and the lawyer can check the output at her desk. Nobody has to verify a citation, which is also why it is where lawyers stop being careful.

Working translation or certified translation: a decision table

A working translation lets you read, advise and negotiate. A certified or sworn translation is one a court, registry or authority will accept, and in most jurisdictions only an authorised translator can produce it.

An AI draft of a foreign judgement is a fine way to decide whether enforcement is worth pursuing; it is not the document you attach to the application.

DeepL versus ChatGPT versus Claude for legal text

DeepL is a translation product; ChatGPT and Claude are general models that translate. Frontier models carry an advantage on legal text that Anthropic’s interpretability research explains: “the shared circuitry increases with model scale, with Claude 3.5 Haiku sharing more than twice the proportion of its features between languages as compared to a smaller model”.

Two caveats. Translation is not legal reasoning: on the LEXam benchmark of 7,537 law-exam questions in English and German, GPT-5 scored 70.2 and Claude 3.7 Sonnet 62.9, while EuroLLM-9B managed 22.95; a model can translate a Gutachten fluently and still get the law wrong, so Lulius’s rule applies: “Allgemein-LLM für Sprache, spezialisierte Legal AI für Recht.” And general models are, in Darwin Gray’s words, “often geared towards a US audience”; expect American terminology in an English-law document unless you say otherwise.

Terminology control: glossary first, translation second

Defined terms are where AI legal translation fails silently. A forty-page agreement defines “Closing” and “Affiliate” once; a model translating in one pass renders them consistently for ten pages and then, as the context fills, starts varying them (the context window guide explains why). The fix is a glossary before the model sees the document, extracted from the source or from a dual-language agreement you have already signed, then blocks with the glossary attached to each.

Translate the following block from [German] into [English] for an [English-qualified] lawyer. Binding glossary (use these renderings verbatim, including capitalisation): <glossary>...</glossary>. Preserve clause numbering, defined-term capitalisation and cross-references exactly. Where a term has no glossary entry, translate it and list it under "NEW TERMS" at the end so I can add it before the next block. Do not summarise, omit or reorder anything.
<block>...</block>

Run it block by block, updating the glossary after each.

False friends and untranslatable concepts

The hard part is concepts that exist in one system and not the other, and the model’s instinct to produce a fluent equivalent regardless.

• Verschwiegenheit. “Confidentiality” loses the criminal-law weight of section 203 StGB and section 9 RAO.
• Treu und Glauben. “Good faith” is a trap: the common-law reader hears a narrow implied term, the civil lawyer a principle that reads obligations into every contract.
• Consideration. No civil-law equivalent; the model produces a fluent German word and the German reader wonders why the recital exists.

Translate <text> from [source language] into [target language] for a lawyer qualified in [target jurisdiction]. Where a source concept has no direct equivalent in the target system, translate it as literally as clarity allows and add a numbered translator's note [TN] giving the source concept, the closest target-system concept and how they differ. Never substitute a concept that changes legal effect. Keep the original term in brackets after the first rendering.

One prompt Harvey showcased at the A&O launch in 2023 shows the same problem from the drafting end: “Draft me an email to my Silicon Valley private equity client regarding what is the difference between what constitutes a material adverse effect in an M&A deal under Delaware law as compared to New York law”: a translation task in everything but name.

Foreign-language review in discovery and diligence

The FTI figure is about review: a data room in four languages, or a production where the key emails are in Portuguese. The due diligence workflow and the generative e-discovery guide cover the platforms; the translation-specific rule is to extract in the source language and quote it, so the reviewer verifies against the original and has the passage to hand when opposing counsel disputes the rendering.

For each document in this batch, whatever its language: Language | Document type | Date | One-line English summary | Does it contain [a change-of-control trigger]? (Yes/No, with the original-language passage quoted verbatim and an English rendering beneath it) | Confidence. Never translate a quoted passage without also giving the original. Mark any document you could not read fully as PARTIAL.

Confidentiality: which translation tools train on your text

The distinction that matters is not DeepL versus ChatGPT but consumer versus commercial tier. ChatGPT Free, Plus and Pro and Claude Free, Pro and Max train on conversations by default unless you opt out; Business, Team and Enterprise tiers do not. For DeepL, read the terms of your tier; the burden is yours.

For German and Austrian lawyers the test is stricter and criminal. BRAK’s guidance is that only “abstract” prompts allowing no inference about a specific mandate belong in ChatGPT-type tools, and that a provider’s mere possibility of access counts under section 203 StGB; ÖRAK calls mandate data in public or unsecured AI systems “standesrechtlich unzulässig”. The BRAK, DAV and ÖRAK guidance sets out the section 43e BRAO contract a translation provider needs, the anonymisation guide covers the alternative, and the EU data residency comparison shows where each vendor keeps the data.

Cost, tokens and the translate-back-translate-spot-check workflow

Models read text as tokens, and morphologically rich languages such as German need “substantially more tokens per word” than English. A German document therefore costs more, runs slower and fills more of the context window than its English translation; on metered pricing ($10 per million input tokens and $50 per million output for Claude’s frontier model) that favours blocks over pasting the whole Akte.

The AI legal translation workflow:

1. Anonymise or use a contracted tier. Placeholders and an offline key, or a no-training tier under a section 43e BRAO-grade contract.
2. Glossary first, then translate in blocks with the glossary attached, collecting new terms after each.
3. Back-translate a sample. In a fresh chat, translate three clauses of the output back to the source language and diff them; have the model rate each difference cosmetic, material or changes legal effect.
4. Spot-check the defined terms. Search the output for every glossary entry and confirm the rendering is identical each time.
5. Decide the product. Working copy: done. Anything filed or signed: a bilingual lawyer’s full read, and a sworn translator where the forum requires one. Never confuse the working copy with the certificate.

As a Vienna partner working in German and English most days, I run this loop constantly; it is one of the live demonstrations in AI Lab for Lawyers, on an anonymised bilingual document with the glossary built on screen.

Where to go next: the DACH legal AI tools overview covers the platforms built for German-language work; the prompt library holds these prompts beside the other document workflows in the use-cases cluster; and the overview of how lawyers use AI puts translation next to summarisation and review. The live sessions of AI Lab for Lawyers make the loop a reflex.