A European law firm saved €26,500 and 13 days on one project—by doing MTPE right. A publisher lost money—by doing it wrong.¶
340 pages of contracts. Three languages. One week to deliver.
Scenario A: Book a traditional LSP. Standard rate: €0.18-€0.30 per word. Timeline: 18-25 business days. Bill: €45,000.
Scenario B: Upload to an in-house MTPE system with a built terminology base. Assign two certified legal editors. Turn it around in five days. Bill: €19,500. Savings: 56%.
Then there’s Scenario C: A small publisher tries to translate a cookbook via light MTPE with a mass-market AI tool. Machine renders “zesty” as “spicy-spicy,” strips the friendly tone, and produces output that takes longer to edit than rewriting. Back to human translation. MTPE failed.
The difference between Scenarios A and C isn’t the technology—it’s knowing when to use it and how to use it right.
Machine translation post-editing (MTPE) is powerful and economical. It’s also a minefield when applied to legal work without guardrails. Let’s talk about where the line is, what can go wrong, and how to set up a system that actually works.
What post-editing is, and why lawyers suddenly care¶
Post-editing is when a human reviews machine output, fixes errors, and polishes the text for publication.
Simple in theory. Brutal in practice: machines translate fast but thoughtlessly. They miss context, misunderstand legal terms (consideration, breach, force majeure), confuse acronyms (LLC, NDA, IP), and fumble on ambiguous words. “Power of attorney” could mean a lawyer’s authority (correct) or the energy wielded by attorneys (absurd). Machines often choose wrong.
A post-editor—in legal translation, this is a certified legal translator or a practicing lawyer who translates—reviews the machine output and fixes: - Grammar and punctuation. - Terminology: so “consideration” stays consistent as “exchange” or “review,” never bouncing between the two. - Dropped sentences (machines lose them often). - Semantic errors: the translated phrase has flipped the meaning of the original.
Why does this matter to law firms? Because traditional human translation is slow and expensive. A contract translation might run €0.18-€0.30 per word, take 18-25 days, and cost tens of thousands of euros. MTPE cuts that by a third or more. And if done right, the quality stays identical.
But “done right” is the operative phrase.
The numbers: how MTPE actually saves time and money¶
Three metrics that shift the calculation:
Speed: From 18-25 business days (human-only) to 5-10 days (MTPE). Research from 2025 shows a post-editor spends an average of 40-54 seconds reviewing one segment of machine output, versus 2-3 minutes writing a translation from scratch. On a 340-page contract, that’s the difference between the third week and Friday of the same week.
Cost: 35-60% savings depending on document type. One European law firm cut its annual translation budget by 42% after moving routine contracts to MTPE. Not all contracts—only the boilerplate-heavy ones.
Consistency: Set up a terminology database once, and the machine enforces it everywhere. “Force majeure” translates the same way on page 1, page 100, and page 340. “Breach of contract” never wavers. Humans forget; machines don’t.
That’s the upside. Now the risk.
Why 38% of unreviewed machine-translated legal documents contain critical errors¶
When researchers in 2025 analyzed machine-translated legal texts, they found something chilling: 38% contained errors that fundamentally change meaning. Not “the prose is awkward.” Errors of this magnitude:
- A dollar amount flipped (€50,000 instead of €500,000).
- A condition inverted (“when the client breaches” became “if the client pays”).
- A meaning reversed (“not prohibited” became “prohibited”).
- A party omitted (“both parties” became “one party”).
These errors don’t get footnoted. They invalidate contracts and lose lawsuits.
Why does this happen?
Machines are pattern-recognition engines, not comprehension engines. They don’t read a contract like a lawyer, understanding obligations. They look at training data—billions of text snippets—and pick the variant that appeared most often nearby. If the training set had 10,000 examples of “consideration” as “exchange” and 1,000 as “review,” the machine picks the first. But in edge cases, in unfamiliar syntax, in specialized legal language, it falls apart. And legal documents are edge cases by definition.
When a human expert then reviews? Errors plummet to 2%. Not magic—just expertise.
Light PE vs. full PE: where savings become liability¶
There are two flavors of post-editing.
Light post-editing: Fix grammar, punctuation, and obvious blunders. Goal: make the text readable and meaning preserved. Time: cut in half. Cost: ~50% cheaper than full PE. Risk: medium-level errors might slip through.
Full post-editing: Reviewer checks every sentence on semantics, terminology, context, and legal accuracy. Takes about as long as a full human translation but with less cognitive load, so the editor can focus. Cost: close to full translation but with faster turnaround. Risk: minimal if the editor is qualified.
For legal documents destined for court, a government registry, an embassy, or a settlement agreement—always full PE. Light PE is for internal memos, vendor correspondence, and procedural notices where mistakes have local impact.
Where MTPE works for lawyers (and where it breaks)¶
Two categories.
MTPE works for: - Standard commercial contracts (lease templates, NDA boilerplate, purchase agreements). - Procedural documents (service descriptions, regulatory notices, findings). - High-volume, repetitive documents (consent forms, terms of service). - Internal correspondence and legal memoranda. - Discovery and legal research packs (working materials).
Why? The text repeats, so the machine learns. Stakes are medium; an error costs thousands, not millions. You often have the original to double-check.
MTPE doesn’t work for: - Litigation documents headed to court. Judges read translations with a magnifying glass. One error tips the case. - Patents. A single mistranslation leaves the patent holder unprotected in half the jurisdictions. - Settlement agreements, divorce decrees, and compromise documents. These are about money and structure—errors lead to appeals and re-litigation. - Expert reports and affidavits attached to cases. Translation errors undermine the expert’s credibility. - Contracts with local counsel. You’ve outsourced responsibility but kept the risk.
How to reduce risk: a framework for managing MTPE projects¶
If you’re rolling out MTPE, here’s how to do it without the regulatory hangover.
Step 1: Risk-based content segmentation
Before MTPE, split the document into sections by legal risk. High-risk content goes to a judge; low-risk is bureaucratic boilerplate. Then: - High-risk = remove from MTPE, translate manually. - Medium-risk = light PE only if a senior attorney then reviews. - Low-risk = light PE + spot-check via automation.
Step 2: One terminology database for everything
Populate a shared glossary before MTPE starts. “Consideration” → “exchange,” “breach of contract” → “violation of contract” (not just “violation”). The machine enforces it; editors verify compliance. A database of 300-500 core terms is enough; 5,000+ is gold for large LSPs.
Step 3: Qualified editors and QA sampling
A post-editor for legal text isn’t just someone fluent in two languages. They need to be: - Certified in ISO 18587:2017 (the international standard for professional MTPE). - Experienced in legal translation (minimum 2-3 years). - Fast at spotting machine errors without rewriting the whole passage.
Quality assurance: pull ~5% of completed documents and have a senior lawyer review them. Errors flag that the editor is under-qualified or the glossary is incomplete.
Step 4: Technology and infrastructure
You need a platform that: - Enforces terminology consistency (TM + glossary integration). - Supports comments and revision tracking. - Doesn’t store documents on public servers (critical for legal work). - Measures quality against standards like MQM (Multidimensional Quality Metrics).
Platforms: memoQ, Trados, Phrase, Lokalise, or specialized MTPE tools like Argos Translate for privacy-first deployments.
Confidentiality and liability: who pays if something breaks¶
This conversation must happen before you sign anything.
If you upload a legal document to free ChatGPT, DeepL, or Google Translate, you’re granting those platforms the right to use your data for model training. A secret agreement, a patent specification, a medical record—all end up on servers used to train millions of AI systems. Then your law firm gets hit with a GDPR fine for data leakage, and you caused it.
Solution: Use MTPE vendors who guarantee non-disclosure and non-retention agreements, or deploy a self-hosted system on your own infrastructure. Legal documents are sensitive data—treat them like it.
The second part: liability. When you hire an LSP, they’re responsible for quality. When you buy an MTPE service and hire your own editors, you’re responsible. That means lawsuits over poor system selection, inadequate QA, unqualified editors—that’s now your problem. Civil liability. Regulatory fines.
So if you’re deploying MTPE, the bare minimum is documentation proving you selected the system correctly, trained editors properly, and ran QA rigorously. Without that, you’re exposed.
Real cases: when MTPE works and when it doesn’t¶
Case 1: European Law Firm (340 pages, 3 languages)
Constraint: Contracts for a major client. €45,000 via LSP. 18-25 days. Deadline: one week.
Solution: Loaded documents into Trados with their own translation memory and glossary. Assigned two in-house legal editors with prior experience translating these contracts. Each document took 1-2 hours of editing.
Outcome: Delivered in five days at €19,500 (50% of the LSP quote). Three languages in parallel. Quality matched the LSP because the editors knew the contracts deeply.
Case 2: Small Publisher (Cookbook)
Constraint: 150 pages, culinary terminology, required a warm, friendly tone.
MTPE attempt: Machine rendered “hot dish” as “heated plate,” “zesty” as “very spicy,” and flattened the tone entirely. Light PE took longer than rewriting. Generic MTPE couldn’t capture style or nuance.
Outcome: Switched back to human translation, using AI as a rough draft to save 10-15% of time. MTPE wasn’t the right tool here.
Lesson: MTPE thrives on repetition, formality, and neutral tone. It fails when you need personality, nuance, or scarce context.
FAQ: What translation teams ask about post-editing¶
When is MTPE safe for legal work, and when is it risky?
MTPE is safe for low-stakes, repetitive content—internal memos, vendor correspondence, standard templates. It’s risky for anything that reaches a judge, a regulator, or a central authority. Those need human eyes.
How much time does MTPE actually save?
40-60% of a full human translation timeline. Full translation: 18-25 days. MTPE: 5-10 days. Assumes you already have a mature system (glossary, trained editors, QA process).
Light vs. full post-editing—what’s the real difference?
Light PE: grammar, obvious mistakes, readability. Full PE: every sentence checked for accuracy, terminology, and legal meaning. Legal documents need full PE almost always.
Who really benefits from MTPE for legal documents?
Agencies translating high-volume, standard documents (purchase orders, employment contracts, ongoing client work). Freelancers using it as a working draft when clients ask for lower rates.
How do I keep legal documents confidential?
Never use free public AI tools. Use MTPE vendors with non-disclosure guarantees, or run a local system. Legal documents are sensitive—treat them that way.
Will a court accept a machine-translated document?
Depends on the jurisdiction. Many require certified translations signed by a sworn translator. Machine output without certification won’t be admitted as official evidence. It’s fine as a working document for your lawyers, but not for the court.
Bottom line:
MTPE is a powerful tool, but not for everything. Use it for low-stakes, repetitive work where mistakes are recoverable. For high-stakes documents—keep humans in control. The hybrid approach—MTPE for routine work, humans for critical language—is how modern translation teams cut costs without sacrificing quality.