AI Agents for Lawyers

Use Cases

AI Legal Translation: Accuracy, Confidentiality and Workflow

AI legal translation gives reliable working drafts of contracts and pleadings when a glossary, system prompt and bilingual review surround it. German-English example.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. Where AI legal translation stands in September 2026
  2. A German-English contract clause, two ways
  3. Terminology: the glossary is the asset
  4. Side-by-side review without re-translating
  5. What a certified translation still requires
  6. Confidentiality and data location by tier
  7. The quality-assurance checklist
  8. Where this leaves you

Frontier models now translate legal German into English well enough that a Big Law corporate lawyer said on Hacker News in September 2026 that his firm gives AI translation away and charges only for reliance. That is the whole workflow: the machine drafts, a bilingual lawyer makes it reliable, the fee follows reliance. What separates a usable draft from an embarrassing one is a glossary, a system prompt and a side-by-side check.

"Before computer translation became good (i.e pre-Deepl era), we used to translate documents (cease and desist letters, court petitions etc.) ourselves. Now we offer AI translation free of charge. If client wants reliance on the translation we charge a small bit." — a self-described Big Law corporate lawyer, Hacker News, 17 September 2026

There is no independent legal-translation benchmark comparable to the research and drafting benchmarks that now exist for English-language legal work; the accuracy figures you will see are vendor claims. The closest proxy is omission: Vectara's summarisation hallucination leaderboard of May 2026 put the best models at roughly 1.8% to 3.3% (Vectara), and a translation that silently drops a sub-clause is the same failure. The practical picture:

OptionStrengthWeaknessUse it for
Dedicated engine (DeepL and similar)Fast, consistent, glossary support, formal register for German; free and paid tiers have different data terms, so check the current onesNo reasoning about legal function; literal on terms of art; sees one segment at a timeFirst pass on long, low-risk documents; correspondence
Frontier model with a system prompt (Claude Opus 5 or Fable 5.1, GPT-6 Astra, Gemini 3.1 Pro)Understands function, applies a glossary, retains source terms, explains departures, holds a whole agreement in a million-token contextSlower; smooths ambiguity the source left deliberately; must be told not to improve the draftingContracts, pleadings, opinions
EU-domiciled or open-weight (Mistral Large 3, Meta's Muse Glimmer 30B on a 24GB GPU, a local model on a German server)Data location and controlQuality gap on hard passages; Mistral Large showed 64% citation errors in the HAQQ research benchmark, a caution about precisionFirms with strict residency requirements
Legal-platform agents (LegalOn's Translate agent, launched 10 February 2026)Integrated with contract reviewVendor-boundIn-house contract flow
Court-certified or sworn translatorLegally accepted; carries liabilityCost and timeAnything filed or registered

For the frontier row, the Claude versus Gemini comparison covers context and data terms; the open-weight comparison covers the third row.

A German-English contract clause, two ways

Source, a standard limitation clause of the kind found in German-language service agreements:

§ 9 Haftung. (1) Die Haftung des Auftragnehmers für leicht fahrlässig verursachte Schäden ist ausgeschlossen, soweit nicht Schäden aus der Verletzung des Lebens, des Körpers oder der Gesundheit oder aus der Verletzung wesentlicher Vertragspflichten (Kardinalpflichten) betroffen sind. (2) Bei leicht fahrlässiger Verletzung wesentlicher Vertragspflichten ist die Haftung auf den vertragstypischen, vorhersehbaren Schaden begrenzt.

A literal rendering, of the kind a segment-by-segment engine produces without a glossary:

§ 9 Liability. (1) The contractor's liability for damages caused by slight negligence is excluded, unless damages from injury to life, body or health or from the breach of essential contractual obligations (cardinal obligations) are affected. (2) In the event of slightly negligent breach of essential contractual obligations, liability is limited to the typical, foreseeable damage.

A functional rendering, produced with the glossary and system prompt below:

Clause 9 Liability. (1) The Contractor shall not be liable for loss caused by ordinary negligence (leichte Fahrlässigkeit), except for loss arising from death, personal injury or damage to health, or from breach of a material contractual obligation (wesentliche Vertragspflicht, "Kardinalpflicht"). (2) Where a material contractual obligation is breached through ordinary negligence, liability is limited to the loss typical for this type of contract and foreseeable at the time of contracting.

Four things changed, and each is a legal decision rather than a linguistic one. "Slight negligence" is not a category an English-law reader recognises; "ordinary negligence" with the German term retained tells the reader that this is the lower German tier, below grobe Fahrlässigkeit. "Damages" in English means the money awarded; the German Schaden is the loss suffered, so "loss" is right. "Kardinalpflicht" has no English equivalent and is a term of art from German case law, so it stays in the text. And "vertragstypisch, vorhersehbar" is a German-law limitation formula; a model that silently rewrites it as "reasonably foreseeable" has imported the English remoteness test into a German contract. The literal version is not wrong as language. It is wrong as law, and a client relying on it would misjudge the exposure.

Terminology: the glossary is the asset

GermanPreferred EnglishTrap
GmbHGmbH (retain)"Limited liability company" invites US LLC assumptions
Geschäftsführermanaging directorNot "CEO"
Prokuristauthorised signatory (Prokurist)No English equivalent; retain
Vertragsstrafecontractual penalty"Liquidated damages" changes the enforceability analysis under English law
Gewährleistung / Garantiestatutory warranty / guaranteeBoth are routinely flattened to "warranty"
Rücktritt / Kündigungwithdrawal (rescission) / terminationConflated by every engine
Schadenersatz (AT) / Schadensersatz (DE)damagesThe spelling tells you which German; tell the model
Firmacompany nameNot "firm"
Einstweilige Verfügunginterim injunction
Betriebsratworks councilNot "union"
Steuerberatertax adviser (Steuerberater)Not "accountant"
Rechtsanwaltlawyer, attorney-at-law (Rechtsanwalt)Not "solicitor" or "barrister"

Austrian and German legal German differ in vocabulary (Jänner, Schadenersatz), statute (ABGB against BGB) and institutions; the system prompt must say which. Build the glossary once per language pair, keep it under version control with a named owner, and load it into every run through the system prompt:

You are a legal translator working German (Austria) to English (UK).
Audience: English-qualified lawyers who do not read German.

Rules:
1. Translate legal function, not words. Where a German term has no
   English equivalent (e.g. Kardinalpflicht, Prokurist), give the
   closest English and keep the German in brackets on first use.
2. Apply the attached GLOSSARY exactly. Never substitute a synonym.
3. Do not import English-law concepts. "vertragstypisch,
   vorhersehbar" is a German limitation formula; render it
   descriptively, never as "reasonably foreseeable".
4. Preserve structure: clause numbering, defined terms in the same
   capitalisation, every list item, every cross-reference.
5. Do not improve, shorten or clarify the drafting. Ambiguity in
   the source stays ambiguous in the target.
6. Numbers, dates, currencies and party names are copied verbatim
   and then listed in a final table for checking.
7. After the translation, output a DEPARTURE LOG: every place you
   did not translate literally, with the source, the target and
   one line of reasoning.
8. If a passage is illegible or genuinely ambiguous, mark it
   [CHECK] rather than guessing.

Rule 8 matters for scans: the models read scanned PDFs by vision, and practitioners report that Claude is more likely to say "illegible" than to invent a number (Spellbook's analysis); for anything that will be relied on, run a dedicated OCR pass first, as the PDF extraction playbook describes. One note for Opus 5 users: Anthropic says the model "verifies its own work without being told to", so an old-style instruction to "add a final verification step" now produces over-verification. Ask for the departure log instead.

Side-by-side review without re-translating

  1. Align by clause number: source left, target right, one row per clause. The model will produce this table if asked.
  2. Read the departure log first. Every entry is a legal judgment the model made; accept or reverse each one.
  3. Read the target alone, as an English lawyer would, for sense and register.
  4. Read the pairs for five risk classes only: negations (nicht, kein, soweit nicht); defined terms and party names; numbers, dates and currencies; modal verbs (soll, muss, hat zu, kann against shall, must, may); and list completeness.
  5. Back-translate the ten riskiest clauses with a different model, English to German, and compare with the source. Divergence marks the clauses to read twice.
  6. Search the target for every glossary term to confirm compliance.
  7. Sign off with name, date and scope. The scope line is the one the client sees: "Working translation for information; not a certified translation; the German text prevails."

That last sentence is where reliance is priced. A translation delivered under it costs the client little; a translation delivered without it is an opinion on the English text, and should be billed and reviewed as one.

What a certified translation still requires

Courts, land registers, company registers and authorities do not accept AI output. In Austria, translations for court and official use come from allgemein beeidete und gerichtlich zertifizierte Dolmetscher; in Germany, from translators sworn or authorised under the rules of the relevant Land; in the UK there is no state certification and the translator or agency signs a statement of accuracy that is sometimes notarised or apostilled; in the US, agencies such as USCIS require a signed certification of accuracy. AI does not change any of this. What it changes is the cost of the certified step: a sworn translator who receives your glossary and a reviewed AI draft is checking rather than drafting, and many now price it that way. Ask before assuming; some certified translators will not work from a machine draft as a matter of professional rule.

Confidentiality and data location by tier

The confidentiality analysis for translation is the same as for any client document, with one aggravating factor: translation means pasting the whole document in, not a question about it.

TierTraining and retention (as of September 2026)Verdict for client documents
ChatGPT Free, Plus, ProTrains on conversations unless the user opts out; 30-day retention for abuse monitoring even after opt-out (OpenAI enterprise privacy)No
Consumer GeminiActivity kept 18 months; a subset read by human reviewers; Google's own page says not to enter confidential information (Google)No
Claude Free, Pro, MaxUser chooses; opted-in chats retained five years, opted-out 30 daysNo for client text
ChatGPT Business, Enterprise, API; Claude Team, Enterprise, APINot used for training by default; API retention 30 days at OpenAI; Claude API content not retained by default, except that Fable 5.1 requires 30-day retention and is not eligible for zero data retention (Anthropic retention docs)Yes, with a processor agreement
Microsoft 365 Copilot (enterprise)Runs under Microsoft's data-processing terms; consumer Copilot does notEnterprise yes, consumer no
EU-resident enterprise (Gemini Enterprise with EU residency; Mistral, EU-domiciled)Location controlled; Mistral's own argument is that a residency toggle "changes where data sits, not which government can compel its production" (DigitalApplied)Yes, for residency-sensitive matters
Self-hosted (a local model on a German server, as a three-lawyer German firm runs for EUR 20 to 30 a month)Nothing leaves the buildingYes, with lower quality on hard passages (case study)
DeepL free versus ProTerms differ by tier; verify the current Pro terms before usePro only, after checking

The professional rules are blunt. The UK Upper Tribunal in UK v SSHD [2026] UKUT 81 (IAC) said that pasting client letters into public AI tools puts them "in the public domain". BRAK's guidance treats tools that store or forward client data for model training as a breach of section 43a(2) BRAO (BRAK Hinweise). ÖRAK's four rules from its March 2026 KI-Tag are that client data stays on EU servers, AI use is disclosed to clients, the lawyer keeps final responsibility and no client data is used for training (Anwaltsblatt). The confidentiality checklist and the Austrian and German guidance page give the settings and the rules in full, and the zero data retention entry explains what that term does and does not buy.

The quality-assurance checklist

  • Tier confirmed: business, enterprise, EU-resident or self-hosted; training off; processor agreement in place
  • Language variant stated in the prompt (Austrian or German German; UK or US English)
  • Glossary loaded and version noted
  • Whole document in one run, or clause-aligned chunks with definitions attached
  • Departure log received and every entry accepted or reversed
  • Target read alone for sense
  • Pairs read for negations, defined terms, numbers and dates, modal verbs, list completeness
  • Ten riskiest clauses back-translated with a second model
  • Glossary compliance searched
  • Scope line on the document: working translation, not certified, source text prevails
  • Reviewer name and date recorded
  • Certified translator engaged for anything to be filed, with the glossary and the reviewed draft

Where this leaves you

Translation is the AI use case where the economics have already settled: the draft is close to free, the review is the service, and the certificate stays human. A firm that builds the glossary, fixes the tier and trains its bilingual lawyers on the five-class review can offer working translations at a price that would have been impossible in 2023 and still sign its name to the reliance. Module 3 of the AI Lab for Lawyers course works through multimodal and file handling, including glossary-driven translation and the side-by-side review. The Austrian firm reading this already knows which German it writes; the model does not until it is told.

Frequently asked questions

Is ChatGPT or DeepL better for legal translation?

For a first pass on long, low-risk documents a dedicated engine such as DeepL is fast and consistent, especially with a glossary loaded. For contracts and pleadings where the legal function of a term matters, a frontier model with a system prompt and glossary is better because it can retain source terms, explain departures and handle a whole agreement in one context. Neither replaces the bilingual review.

Can a law firm use AI to translate confidential client documents?

Only on a tier where the text is not used for training and where retention and data location are known: business or API tiers, enterprise Copilot, EU-resident enterprise offerings or a self-hosted model. Consumer tools are out. A UK tribunal in 2026 said pasting client letters into public AI tools placed them in the public domain, and German and Austrian bar guidance bars any tool that trains on client data.

Does an AI translation count as a certified translation?

No. Courts, registries and authorities in Austria and Germany require translations by court-certified or sworn translators, and UK and US practice requires a signed certification of accuracy from a translator or agency. AI can produce the draft that the certified translator checks, which cuts cost and time, but the certificate and the liability remain human.

How do I check an AI translation without redoing it?

Align source and target by clause number, ask the model for a departure log listing every place it did not translate literally and why, read the target alone for sense, then read the pairs for five risk classes: negations, defined terms and party names, numbers and dates, modal verbs, and list completeness. Back-translate the ten riskiest clauses with a different model and compare.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion