Every year, more law firms handle documents that were never written in English. A supplier contract drafted in Mandarin. A witness statement in Spanish. Bank records in Italian. As business goes global, the paperwork that lands on a lawyer’s desk increasingly arrives in a language the case team cannot read at a glance, and that gap is where cases quietly go wrong.
The stakes are not theoretical. In one of the most expensive examples on record, an international arbitration between the oil company Occidental Petroleum and Ecuador ended in an award of roughly 2.3 billion dollars, and Ecuador’s representatives argued that poorly rendered court documents had distorted the tribunal’s understanding of local law (as documented in this review of multi-million-dollar cases). A single misread clause, in the wrong document, can move a number that large.
Most firms will never touch a billion-dollar dispute. But the same failure points show up in ordinary commercial and cross-border consumer contracts, immigration files, and discovery. Here are five of the most common ways foreign-language documents sink a legal case, and what a careful firm does differently.
1. The document gets thrown out before anyone reads it
The first risk has nothing to do with the meaning of the text. It is procedural. Courts in the United States generally will not accept a foreign-language document as evidence unless it arrives with a certified English version. Under Rule 901 of the Federal Rules of Evidence, the party offering a document has to show the court that it is what they say it is, and for foreign-language material that usually means a certified version authenticated by someone who can vouch for its accuracy, a hurdle litigators run into more often than they expect.
Courts enforce this routinely. In Sicom SPA v. TRS, Inc., an Italian company tried to introduce bank statements in Italian to support a breach-of-contract claim, and the court held the documents inadmissible because they came with no certified English version. The evidence existed. It simply never reached the judge. A firm that discovers this the week before trial has lost time it cannot buy back.
2. One wrong word rewrites the contract
The second risk is about meaning. Legal language is precise, and a single term can carry a specific consequence. The word “consideration” means one thing in common-law contract drafting and something entirely different in everyday speech. Render it loosely into or out of another language and the obligation it describes can weaken or invert. The other side then gets room to argue the reading that favors them.
This is not a rare edge case. The Shanghai Maritime Court has attributed close to 5 percent of the contract disputes it handles to substandard translation, and that figure covers navigation law alone. In one 2011 dispute, “drydocking” was rendered as “tank washing” and a clause about fuel “for domestic service” became fuel “for domestic flights.” Two small slips, two very different sets of obligations, and an expensive fight over which version governed.
3. Machine output that looks right and is not
To move faster, many firms now run foreign-language documents through an AI model first. The instinct is reasonable. The problem is that a single AI model can produce fluent, confident text that is simply wrong, an error type the industry calls a hallucination. Industry data synthesized from Intento’s State of Translation Automation and the WMT24 research shared at the ACL conference indicates that individual top-tier large language models fabricate or distort content somewhere between 10 and 18 percent of the time on translation tasks.
Here is the part that matters for legal work: different AI models fail in different places. One mishandles a date. Another quietly softens a mandatory clause into a permissive one. A third misses the formal register a court filing requires. Because each model sounds equally confident, a lawyer reading only one output has no signal that anything is off. Trusting a single AI on a high-stakes document is a bet that its particular blind spot does not land on the clause that matters.
4. The hidden cost of checking everything by hand
So firms check the machine output manually, which quietly reintroduces the delay the AI was supposed to remove. This verification burden is the real tax on AI-assisted legal work, and it lands on the people least able to absorb it: paralegals and junior associates comparing outputs across five browser tabs.
The approach now emerging to solve this is consensus. Instead of trusting one model, newer platforms run the same text through many models at once and surface only the rendering most of them agree on, discarding the outliers where hallucinations hide. MachineTranslation.com, an AI translation platform developed by the language company Tomedes, runs documents through 22 different AI models at the same time and returns what the majority converge on. According to the platform’s internal benchmarks, that majority-agreement method holds critical errors to under 2 percent and cuts overall error risk by roughly 90 percent against relying on any single mode.
“Individual models disagree far more often than most people realize, and on a long legal document that disagreement compounds,” says Rachelle Garcia, AI Lead at Tomedes. “The safe thing to hand a court is not the output of one model. It is the rendering that many independent models agree on.”
For a firm, the point is not speed for its own sake. It is that the cross-checking a paralegal used to do by hand now happens before anyone reads the result.
5. No human who actually signed off
Reaching agreement across many models handles accuracy at scale. It does not, on its own, deliver the thing a court and a client ultimately want: a named human being who reviewed the document and stands behind it. That is the fifth and final risk, submitting something no qualified person actually verified.
The stronger platforms now close this gap by keeping human verification inside the same workflow. AI agreement produces the draft, then a professional linguist reviews and signs off before the document goes to authorities, with no separate vendor, contract, or delay. For court filings, sworn records, and regulated submissions, that human accountability is what turns a fast rendering into a defensible one. It is the same discipline litigators already apply to preserving and documenting an evidence trail everywhere else in a case.
The takeaway: three habits that remove most of the risk
None of these five risks require a billion-dollar case to hurt you. They surface in routine cross-border contracts, discovery, and filings, and they share one root cause: treating a foreign-language document as a formality rather than a liability. A firm that builds three simple habits removes most of the exposure before it starts. Get a certified version for anything that may touch a court. Never trust a single machine output on high-stakes text. Keep a qualified human in the loop before anything is filed.
For more practical guidance on handling complex matters, browse the Law section. The language on the page is evidence. Treat it that way from the first day of the matter, not the last.
