You open the MT output of a 40-page distribution agreement. The text reads cleanly - fluent, idiomatic, no obvious grammar issues. Quick read-through, fix a few awkward sentences, deliver. Two months later the client’s lawyer flags a clause where “shall not be liable” came out as “shall be liable” - a single dropped negation that inverted a liability cap worth €3 million.
That’s the specific thing about legal post-editing. The errors that cause the most damage are usually the ones that look like they aren’t errors at all.
Why Legal MTPE Fails Where General MTPE Succeeds¶
Technical documentation and product manuals have a useful property for MTPE: errors are visible. A wrong part number, a mistranslated UI element, a mangled measurement - these surface quickly during review.
Legal text doesn’t work that way. Machine translation consistently produces output that reads as fluent while containing errors that are invisible to anyone without legal background in both jurisdictions. Researchers describe this as the detection problem - linguistically clean output creates a strong impression of accuracy that masks critical distortions.
A 2025 analysis of legal MT output by Artlangs found critical errors in 38% of samples reviewed before any human post-editing. When the post-editor had domain expertise, that rate dropped to under 2%. The delta isn’t about how carefully someone reads. It’s about whether they have the legal background to recognize the problem.
Three categories of legal text create the highest MT failure rates:
Jurisdiction-specific terminology. Legal concepts are not universal. “Good faith” in common law doesn’t map directly to “Treu und Glauben” in German law - each term has different scope, different prerequisites, different case law behind it. “Force majeure” has different trigger conditions in French, English, and German contract law. MT translates the words. It cannot translate the legal concept behind them.
Operative versus descriptive language. MT handles descriptive sentences better than operative legal drafting. “The seller represents that…” functions differently from “The seller warrants that…” in an English contract - different legal consequences, different remedies on breach. The post-editor needs to know that functional difference to catch the substitution.
Complex negation structures. Legal drafting frequently uses layered negated constructions: “shall not be required to,” “nothing in this clause shall be construed to prevent,” “not without prior written consent.” These cause MT errors at a higher rate than affirmative clauses - and when negation is dropped or inverted, the meaning reverses entirely.
One of the greatest risks associated with modern machine translation is not the number of errors it produces, but rather how difficult those errors can be to detect. Linguistically fluent translations create a strong impression of accuracy while potentially containing critical distortions.
There’s also a data problem. Standard MT evaluation metrics like BLEU score surface-level linguistic similarity, but legal accuracy depends on clause structure, operative terminology, and formulaic meaning - none of which BLEU captures. This means the MT engine’s self-reported quality metric is genuinely unreliable as a signal for legal content.
Document Risk Tiers - What to Accept for MTPE and What to Decline¶
Not all legal documents carry equal risk. A practical framework for deciding which documents to post-edit at all:
| Tier | Documents | Appropriate workflow |
|---|---|---|
| High risk | Contracts, court filings, patent claims, sworn/certified translations, settlement agreements | Full human translation or full PE with attorney-level review - never light PE |
| Medium risk | Due diligence packages, regulatory submissions, corporate governance documents | ISO 18587-compliant full PE with domain-specialist second reviewer |
| Lower risk | Internal legal memos, discovery documents for informational use, reference materials | Full PE acceptable with legal background requirement for post-editor |
Artlangs’ analysis of legal MTPE workflows states it plainly: “pure machine output or light post-editing on operative contract language is still unacceptable.” Their recommended approach: MTPE for discovery, due-diligence packs, and internal memos; full human or certified review for anything that carries binding legal weight.
Sworn and certified translations are a hard line. In many jurisdictions - Germany, France, Austria, Argentina, and others - certified translations are only legally valid if produced or certified by a sworn/court-recognized translator. No post-editor’s review replaces that certification, regardless of quality. “MTPE of a sworn translation” is a contradiction in terms: the legal validity belongs to the human translator’s certification, not the MT pipeline. Courts in these jurisdictions may reject machine-translated documents as inadmissible unless a recognized sworn translator has certified them.
Patents get their own risk category. Research from the IPR Group notes that in patent translation, an inadequate rendering of a claim can alter the scope of protection - leading to rejection of the application or exposure in future litigation. Patent claims use a highly specific drafting structure that MT regularly distorts, and the overlap of technical and legal domain knowledge makes finding genuinely qualified post-editors expensive and uncommon.
The Five Most Dangerous MT Error Patterns in Legal Text¶
Knowing which error types cause the most damage changes where you focus your review. These five patterns account for the majority of legal MT errors that reach disputes:
1. Negation inversions. “The licensor shall not be liable for…” becomes “The licensor shall be liable for…” Complex negation structures - especially double negatives like “not without prior written consent” - are among the highest-error categories for neural MT in legal text. Always read operative clauses twice, specifically checking that negation structure is preserved exactly.
2. Modal verb substitution. “Shall” and “may” carry different legal weight - one is mandatory, one permissive. MT frequently substitutes modals, especially across languages where the source and target don’t have the same modal system. A “shall” that becomes “may” turns an obligation into an option; a “may” that becomes “shall” turns a permission into a requirement. Both directions create enforceable problems.
3. False legal equivalents. Terms like “consideration,” “discovery,” “injunction,” “tort,” or “negligence” don’t have one-to-one equivalents across legal systems. MT picks the closest surface-level match in the target language, which may have different scope, different prerequisites, or different legal consequences. The error is undetectable without knowledge of both legal systems.
4. Definition drift across clauses. Legal documents define key terms once and use the defined term consistently throughout. MT doesn’t maintain this consistency - it may render the same term differently in different clauses. By the end of a 40-page contract, “Intellectual Property Rights” might appear as three different translations in the target language, none of which precisely match the definition in Article 1. This creates interpretation disputes and ambiguity about what the parties actually agreed to.
5. Number and date errors. Standard for any MT post-editing, but higher-stakes in legal text. A wrong percentage in a penalty clause, a wrong date in a notice period, a wrong figure in an indemnification cap - these create quantifiable disputes. QA tools catch many of these mechanically, but they need to be configured for each document’s specific number patterns.
The Willie Ramirez case illustrates the cost of a single wrong word. A Spanish-speaking patient’s family described his condition as “intoxicado” - meaning poisoned or suffering an adverse reaction. An interpreter rendered it as “intoxicated.” Doctors treated for drug overdose instead of the brain hemorrhage he was actually suffering from. As Ulatus documents:
The resulting malpractice settlement totaled $71 million - a consequence of a single mistranslated word.
That’s an interpretation case, not a legal document. But the structure is identical: one word wrong, invisible to anyone without domain knowledge, consequences measured in eight figures.
A QA Process That Actually Works for Legal Documents¶
Generic MTPE QA processes don’t scale to legal risk. These are the elements that matter specifically for legal content:
Pre-editing suitability assessment. Before any MT engine touches the document, answer three questions: Is this a certified/sworn translation? (If yes, MTPE is the wrong workflow entirely.) Does the source contain highly idiomatic legal language or jurisdiction-specific concepts with no target-language equivalent? (If yes, evaluate whether viable MT output is achievable at all.) Is the language pair one where MT quality for legal content is established? (Legal MT quality varies drastically by pair - EN-DE and EN-FR typically produce better output than less-resourced pairs.) Skipping this assessment is the most common way to lose money on legal MTPE projects.
Terminology lockdown before translation. Build the termbase before the MT run. Extract all capitalized defined terms from the source document, establish approved target-language equivalents, and load them as enforced glossary entries before the MT engine processes the text. This eliminates terminology inconsistency errors at source rather than catching them during post-editing.
Mandatory 4-eye review for full post-editing. ISO 18587 requires a second qualified person to review full post-editing output. For legal content, “qualified” means domain-qualified: someone with documented background in legal translation for the specific legal field (contract law, IP law, corporate law) and both legal systems involved. A PM doing a read-through doesn’t count. A general translator without legal background doesn’t count. The reviewer needs to be able to recognize a false legal equivalent.
Automated QA before human review. Xbench and Verifika catch number inconsistencies, terminology mismatches against the termbase, cross-reference consistency, and formatting errors. Run these before the human review stage, configured for the specific document’s number patterns (percentages, monetary amounts, article references). They eliminate the mechanically-detectable errors so human review time focuses on the errors only a human can find.
Know when to decline. Some projects shouldn’t be accepted as MTPE at all. If the MT output is consistently poor quality - more than 40-50% of segments need substantive rewriting - the project should be reproposed as human translation. Charging MTPE rates for what is effectively translation from scratch is unfair to the post-editor and creates a cost structure that doesn’t hold up under review.
Post-Editor Qualifications: What ISO 18587 and ISO 20771 Require¶
Two standards govern legal MTPE qualifications. Understanding what each requires clarifies who you can legitimately assign to legal post-editing work.
ISO 18587:2017 covers MTPE generally and requires documented competence in six areas for any post-editor. The sixth - domain competence - is the most relevant for legal work. The standard defines it as specialist knowledge of the subject matter being translated. For legal content, this means documented background in legal translation, legal studies, or legal practice in the relevant field and both jurisdictions. Knowing both languages is necessary but not sufficient, and ISO 18587 makes that explicit.
ISO 20771, adopted in 2020, sets competence requirements specifically for legal translation. It covers understanding of both legal systems involved, the ability to research legal terminology, and the judgment to recognize when a concept has no precise equivalent in the target legal system. As the ATA Legal Translation Division notes, these requirements apply equally whether the translation starts from source text or from MT output.
In practice: a post-editor working on an employment contract needs documented experience with employment law documents in both jurisdictions - not a law degree, but demonstrable legal translation experience in that field. A post-editor working on German corporate documents needs to understand the structural differences between an AG and GmbH at the level required to translate governance documents accurately.
As transpose.ch explains regarding ISO 18587’s legal application:
For legal documents specifically, anyone who translates contracts must be familiar with both source and target contract law terminology. ISO 18587 expressly requires that these competencies are proven - for example through appropriate degrees or certificates.
If an agency needs to demonstrate post-editor qualifications in a client audit, “they’re experienced” is not a qualifying answer. The standard expects documentary evidence: degree certificates, legal translation portfolio in the specific field, client references from legal translation work. Collecting and maintaining these records is operational overhead, but it’s also the difference between claiming ISO 18587 compliance and having it.
Pricing Legal MTPE Correctly¶
Legal MTPE sits in a different pricing tier from general MTPE - and the gap exists for specific structural reasons that are worth understanding.
General full post-editing rates in 2025 run $0.08-0.15/word. Certified legal MTPE - with domain specialist post-editors, mandatory second review, termbase management, and MT suitability assessment built in as non-optional process steps - runs $0.15-0.25/word. That range overlaps with human legal translation rates of $0.18-0.30/word for specialized content.
The question clients always ask is: if legal MTPE costs almost as much as human translation, why use MT at all?
The honest answer: for some legal documents, you probably shouldn’t. For binding contracts, court filings, and anything with certified output requirements, the risk-adjusted cost of MTPE often exceeds human translation once you price in mandatory second review and the liability exposure for missed errors. The financial case for legal MTPE is strongest when volume is high, content is repetitive, and the stakes per error are lower:
- High-volume discovery documents where 80% of pages are informational and only flagged sections need detailed review
- Due diligence packages with repetitive structure - same clause types across 50+ contracts with TM leverage from earlier work
- Internal legal memos and reference materials where an error is caught before it drives decisions
- Multi-lingual regulatory submissions in standard formats with established terminology and termbase
Where the economics don’t work: anything with binding legal force, sworn/certified translations (excluded from MTPE by definition), patent claims, settlement agreements, and court orders.
On the translator side, the pricing psychology around MTPE creates a real problem in the legal space. A 2025 survey by GTS Translation found that 50% of translators don’t offer MTPE discounts at all - arguing that good post-editing requires as much skill as translation. For legal content specifically, that position is defensible. Finding a negation inversion in a 40-page contract requires exactly the same legal knowledge as translating that clause from scratch - and in some ways more, because the post-editor has to override a plausible-looking wrong answer rather than constructing the right one from scratch.
Price legal MTPE based on what the workflow actually costs. Domain-specialist rates are higher than general translation rates. Mandatory second review adds 15-20% to time cost. Termbase management and suitability assessment are non-billable overhead that still needs to be priced in. If the full cost delivers at 80-85% of human legal translation rates, price it there - not at 50% because “MT did most of the work.”
FAQ¶
Is MTPE suitable for legal document translation?¶
It depends entirely on the document type. MTPE is suitable for internal legal memos, discovery documents used for informational purposes, and large due diligence packages with repetitive structure. It’s not suitable for binding contracts, court filings, sworn translations, patent claims, or any document where a single translation error changes legal obligations. Always run a pre-editing suitability assessment before any MT engine processes legal text.
What types of legal documents should never be machine-translated?¶
Sworn and certified translations (which require a human translator’s certification to have legal validity in most jurisdictions), patent claims, court orders and judgments, settlement agreements, and any document a party will sign or file based on the translation’s accuracy. In Germany, France, and Argentina, courts may reject machine-translated documents as inadmissible unless certified by a recognized sworn translator.
What is ISO 20771 and how does it relate to legal MTPE?¶
ISO 20771, adopted in 2020, sets competence requirements specifically for legal translation: understanding of both legal systems involved, legal terminology research skills, and the judgment to identify when a source-language concept has no precise equivalent in the target legal system. These requirements apply equally to post-editors working from MT output. ISO 20771 and ISO 18587 together define the qualification baseline for legal MTPE work.
What is the difference between light and full post-editing for legal documents?¶
Light post-editing targets comprehensibility - fixing only clear errors so the text can be understood. Full post-editing targets publication quality comparable to human translation, with a mandatory second review. For legal documents, light post-editing is almost never appropriate for anything beyond internal reference materials. ISO 18587-compliant legal MTPE means full post-editing with a domain-qualified second reviewer.
How much does legal MTPE cost per word in 2025?¶
Certified legal MTPE runs $0.15-0.25/word. This overlaps significantly with human legal translation rates ($0.18-0.30/word for specialized content) because mandatory second review, termbase management, and domain-specialist post-editor rates are all required components, not optional add-ons.
What QA tools do post-editors use for legal translations?¶
Xbench and Verifika are the most widely used automated QA tools in legal translation workflows. Both check number consistency, terminology against the approved termbase, cross-references within the document, and formatting. They catch mechanically-detectable errors before human review but cannot identify false legal equivalents, jurisdictional nuance errors, or negation inversions - those require a domain-qualified human reviewer.
Can a non-lawyer post-edit legal translations?¶
General legal content - standard commercial contracts, corporate governance documents - can be post-edited by experienced legal translators with documented background in legal translation (not a law degree, but demonstrable experience in legal texts and both legal systems). Specialized legal content - patent law, financial regulatory submissions, criminal procedure - typically requires a specialist with domain experience in that specific area. Apply ISO 18587’s domain competence standard as your qualification benchmark.
What is the biggest risk in post-editing contract translations?¶
Invisible semantic errors: output that reads fluently but changes meaning. Negation inversions (shall not → shall), modal substitution (shall → may), and false legal equivalents are the three patterns that create the most legal disputes. These require active pattern-checking during post-editing - specifically checking negation and modal structures in every operative clause - not just a quality read-through.