A single mistranslated Spanish word once moved $760 million of liability inside a $1.77 billion arbitration award. That happened back in 2012, in a document translated by actual human beings, not software. Fast forward to 2026, and AI handles most commercial translation, so the same kind of error travels faster and touches far more contracts than it ever did. Here’s the unsettling part: when AI translation gets idioms wrong in legal contracts, the mistake almost never looks like one. The sentence still reads perfectly. Only the legal effect underneath it has changed.
When “Time Is of the Essence” Becomes a Timing Bomb
The idiom problem hiding inside routine contract translation
“Time is of the essence” isn’t some poetic flourish tacked onto a contract for flavor. In common-law agreements, it’s a specific legal trigger: miss the deadline, and the other party gets to treat the whole agreement as breached. Translate that phrase literally, though, and it reads like a vague comment about punctuality rather than the enforceable clause it is. Idioms like this sit quietly inside routine boilerplate, and most AI translation tools have no way of flagging that they carry real legal weight.
Idiom mistranslation isn’t a vocabulary glitch a bigger model eventually fixes. It’s a category of risk baked into how idioms work: their meaning depends on legal context, jurisdiction, and precedent, not the literal words on the page.
Why this matters more in 2026 than it did five years ago
Crowdin’s 2026 AI Translation Report, a survey of 152 B2B professionals across the US and Canada, puts real numbers behind this: 95% of enterprise teams now use AI or machine translation in some capacity, and about 18% run every translation task through it by default. That means idiom-bearing contract language basically doesn’t pass through deliberate human choice anymore. It just runs through AI, as a matter of routine.
Here’s the telling part: the same survey found that 78.3% of those respondents consider legal and contractual content too sensitive to hand over to external AI providers. That’s basically an entire industry admitting, quietly, that the risk is real, even while adoption keeps climbing. What almost none of these warnings answer is the question that matters most to whoever’s about to sign: who’s on the hook when an idiom slips through, and what can you do about it before signing?
Why AI Gets Legal Idioms Wrong (It’s Not a Vocabulary Problem)
Literal translation vs. legal meaning
Legal language is technical and dense on purpose. Every phrase carries decades of precedent, and machine translation systems don’t understand any of that: they map language patterns statistically, with no way of knowing which patterns are load-bearing. To a model, an idiom looks like any other string of words, so it has no idea that “best efforts” in a contract is a binding legal standard, while in casual conversation the same phrase just means someone tried pretty hard.
A 2023 peer-reviewed study looked at how Google Translate handled legal text and found a lexical and terminology error rate of 43.4%, alongside a legal-register error rate of 30.2%. Scope matters here: that study only examined Arabic-to-English translation of five legislative texts (n=5), so those exact percentages shouldn’t be treated as a universal benchmark. What does carry over is where the errors clustered: around legal terms of art and formal register rather than everyday vocabulary. That’s exactly the category idioms fall into.
Current 2026 figures point the same direction. Overall AI translation accuracy runs somewhere between 82% and 96%, depending on engine, language pair, and subject domain, and legal and technical content cluster at the low end. Hybrid workflows, where AI drafts and a human reviews, are reported to cut legal-translation errors by up to 60% compared with AI-only pipelines. Nobody has published a matching 2025 or 2026 legal-register breakdown, but the mechanism the 2023 study documented, that legal register is where machine translation fails hardest, hasn’t gone anywhere.
Idioms that mean different things in different jurisdictions
Some idioms don’t even need to cross a language barrier; they mean different things between two English-speaking jurisdictions. Under New York law, “best efforts” can force a party to act against its own commercial interest, giving up an option simply because it works, even if it costs more. “Commercially reasonable efforts” carries no such obligation. And under English law, “best efforts” doesn’t mean the same thing as “best endeavours,” even though the phrases sound interchangeable to anyone who isn’t a lawyer. An AI system has no built-in sense that it needs to preserve that jurisdiction-specific legal weight, not just the dictionary meaning.
Force majeure: the same phrase, a different scope in every jurisdiction
Force majeure clauses are another minefield. The phrase is French, borrowed wholesale into English-language contracts, and its scope swings wildly by jurisdiction: some countries read it to cover only the events explicitly named, others read it broadly enough to include government action or a pandemic. An AI tool translating that clause has no reliable way of knowing which reading should govern, so the translated version can quietly expand or shrink a party’s obligations, and nobody notices until a dispute forces the question.
Why bigger language models still stumble on this
Making the model bigger doesn’t fix this, because the failure was never about vocabulary breadth. It’s about pragmatic inference: knowing a phrase means one thing in a supply contract and something else in a merger agreement, and that the right meaning depends on which country’s courts will eventually read it. That’s a legal judgment call, not a translation task, and no amount of training data teaches a model to make legal judgment calls reliably.
Real Contracts, Real Fallout: How Idiom Errors Have Played Out
The “Gesellschaft” clause that changed a company into a society
Here’s a plain illustration, hypothetical rather than litigated. The German legal term “Gesellschaft” correctly translates to “company,” yet machine translation tools routinely render it as “society” instead. Drop that into a cross-border joint venture agreement, and a party can later argue the entity described in the English version isn’t a commercial company at all. No court case is tied to this exact scenario, but the mechanism is real: one mistranslated term of art can blur the line between a for-profit entity and a non-profit association.
A historic case with a nine-figure price tag
The clearest documented example of a translation error moving real money is Occidental Petroleum v. Republic of Ecuador, an arbitration decided back in 2012 at the International Centre for Settlement of Investment Disputes (ICSID). The tribunal’s original $1.77 billion award turned in part on translating the Spanish term “solemnidades” as “legal requirements,” when its narrower, correct meaning was “requirements of solemn form,” something closer to a public deed. That misreading added roughly $760 million to the award, and in 2015, once the error was caught, an ICSID annulment committee cut that exact amount back out.
The case predates the AI translation boom; it was a human translation dispute. But it remains the clearest quantified example of what one mistranslated legal term can do to a nine-figure outcome, and it previews the exposure now moving through AI tools at far greater volume.
Who’s Actually Liable? The Legal Framework Nobody Explains
Is the contract voidable, unenforceable, or just ambiguous?
Most idiom mistranslations don’t void a contract outright. Courts generally would rather find ambiguity and interpret their way around it than declare an entire agreement unenforceable. That preference is exactly what makes these errors dangerous: the contract survives, but its meaning is now up for grabs, and fighting that fight costs money, time, and negotiating power, usually at the expense of whoever relied on the flawed translation.
Contra proferentem and prevailing-language clauses explained
When a court does have to sort out the ambiguity, contra proferentem often comes into play. The doctrine reads unclear language against whichever party drafted it or required its use, and it only kicks in after other interpretive tools fail. That creates a pretty direct incentive: whoever supplied or signed off on the AI-generated translation is the party most exposed if the wording turns out ambiguous.
Prevailing-language clauses head this off before it starts, by naming one language version as controlling. Lawyers have recommended them for years, citing “the inherent risk of mistranslation,” and 2025-2026 legal-tech commentary now extends that logic to AI-generated translations.
No published ruling has yet decided a case squarely on an AI translation error, but courts are already scrutinizing AI-generated mistakes in legal documents. In ARIHQ c. Santé Québec, decided by the Quebec Superior Court on April 22, 2026, the court annulled an arbitral award worth roughly CAD 1.2 million after the arbitrator handed legal reasoning off to generative AI, which produced fabricated case citations. That case turns on AI-hallucinated citations, not a mistranslated clause, so treat it as an adjacent signal rather than direct authority: a 2026 court is willing to undo whatever outcome rests on an AI-introduced error in a legal document. The advisory consensus on translation hasn’t changed either way: don’t let an AI-translated version be the one that governs.
Vendor liability caps vs. your exposure
AI vendor contracts routinely cap liability at whatever fees you paid over the prior twelve months, then disclaim responsibility for output errors, on the theory that AI outputs are probabilistic and can’t be warranted. So if your translation subscription cost $5,000 last year, that’s roughly your ceiling, no matter how large the dispute a single mistranslation triggers.
The 2026 insurance gap: why your E&O policy may not cover this
And the gap gets worse. Effective January 1, 2026, Verisk’s Insurance Services Office filed three generative AI exclusion endorsements for commercial general liability policies:
- CG 40 47 is the broad one, excluding both bodily injury/property damage coverage and personal/advertising injury coverage.
- CG 40 48 is a narrower version, limited to personal and advertising injury.
- CG 35 08 covers products and completed operations.
All three exclude claims “arising out of, or attributable to” the insured’s use of generative AI. Stack a vendor’s liability cap on top of that, and a business can end up carrying an AI translation failure on its own, no vendor indemnity, no insurance stepping in, which is exactly why the audit protocol and clause language below matter.
The Idiom Risk Audit: A Step-by-Step Protocol Before You Sign
Flagging idiomatic and culturally loaded phrases before translation
Before any contract goes through AI translation, run a manual pass and flag every phrase that carries legal weight beyond its literal words: effort standards, timing language, force majeure, indemnification triggers, and any jurisdiction-specific term of art. Then keep a standing glossary of those terms and their approved translations, so you’re not re-deciding the same question from scratch on every deal.
The back-translation test
Once a section is translated, translate it back into the original language, using a different tool or a human reviewer, and compare the result against the source. If the back-translation drifts from the original meaning, especially around effort standards or timing language, treat that as your signal to escalate. This one step alone catches a surprisingly large share of idiom errors before they reach a signature page.
Who needs to sign off (and when a certified human translator is non-negotiable)
Emerging 2025-2026 compliance guidance recommends keeping the translator, an independent reviewer, and a QA verifier as three separate people, each with documented credentials. That structure builds the audit trail regulators and courts increasingly expect wherever AI touches a legal workflow. ISO 17100 sets baseline competency requirements for translation providers, and current guidance treats centralized terminology governance, shared glossaries and translation memories enforcing consistent vocabulary, as the foundation of a defensible process. Any clause involving liability, indemnification, effort standards, or governing law should require certified human sign-off, no matter how confident the AI output looks.
Contract Clauses That Allocate AI Translation Risk
Sample prevailing-language clause language
Here’s a workable starting point: “This Agreement is executed in [Language A] and [Language B]. In the event of any conflict, ambiguity, or inconsistency between the two versions, the [Language A] version shall prevail and shall be the sole authoritative text for purposes of interpretation and enforcement.” Naming a controlling version doesn’t stop translation errors from happening, but it does determine which version a court reads once a dispute lands on its desk.
AI-use disclosure and liability allocation clauses
Add a disclosure clause requiring either party to flag when AI or machine translation produced any version of the agreement, plus an allocation clause spelling out who bears responsibility for resulting ambiguities. Something like: “Any translation produced using AI or machine translation tools is provided for convenience only and carries no warranty of legal accuracy; the party supplying such translation bears responsibility for any resulting ambiguity or dispute.” That clause does real work: it shifts the contra proferentem risk onto whoever chose to rely on the AI output.
What to negotiate with your translation vendor
Push back on standard vendor terms that cap liability at whatever fees you’ve paid and disclaim responsibility for output errors. Ask instead for a carve-out for legal content, higher liability caps tied to the document’s value rather than your subscription fee, and a requirement for human certification above a defined risk threshold.
When to Escalate to a Human Legal Translator
Low-risk vs. high-stakes clauses
Not every paragraph carries equal risk. Boilerplate definitions, recitals, and routine notice provisions can run through AI translation without a second thought. But effort standards, force majeure, indemnification, governing law, and timing or liability-cap language should always route to a certified human translator, since those are exactly the spots where idioms carry outsized legal consequences.
Building an AI-plus-human review workflow that scales
That same 2026 survey found 20.4% of organizations using AI translation reported at least one quality incident since rolling it out, even as 65.8% reported better consistency and 73% reported faster releases. So build a workflow where AI handles first-pass volume, a flagged list of high-risk clause types routes automatically to human review, and nothing involving liability or jurisdiction-specific effort standards ships without certified sign-off. That isn’t a compromise. It’s the only version of “fast” that doesn’t quietly gamble away enforceability.
Don’t Let an Idiom Decide Your Next Dispute
Every idiom sitting inside your next cross-border contract is a small decision waiting to be made badly, by a tool that has no idea it’s even making one. So where does this actually land? Contra proferentem pushes exposure toward whoever supplied the wording. Prevailing-language clauses decide which text a court will even bother reading. And vendor liability caps, stacked on top of the January 2026 insurance exclusions, make sure very little of the cost travels anywhere else. Every one of these mechanisms points the same direction: the party that didn’t check first is the party that pays.
Don’t wait for a mistranslated idiom to surface mid-dispute. Download our free Idiom Risk Audit checklist and run it against your next cross-border contract before you send it out for signature.
