A single conjunction cost a company its entire patent claim. Not a missing paragraph, not a botched technical drawing — one three-letter word, swapped by a machine translation engine that had no idea what it had just done. If that sounds like an outlier, it isn’t. It’s becoming one of the most quietly expensive risks in international IP strategy, and most legal and R&D teams don’t see it coming until the damage is already filed, published, and permanent. 📄
Patent language is arguably the least forgiving text type a machine translation engine will ever encounter. Every word in a claim carries legal weight, and courts across multiple jurisdictions have shown, repeatedly, that they will hold a translated document to the same standard as the original — even when the discrepancy was never intended by anyone involved.
The Word That Quietly Rewrote a Patent’s Scope
In a dispute before the Court of Rome, TEMA Technologies and Materials challenged a patent held by Schluter Systems over what looked, on paper, like a trivial linguistic detail. The original claim described a foam core layer as “moisture-resistant and moisture-proof.” Somewhere in the Italian translation, “and” became “or.” That single substitution broadened the claim’s apparent scope, opened the patent to challenge, and forced the patent owner into a defensive legal posture it never should have had to take. 😬
A similarly small error unraveled a European patent entirely. In litigation over EP1397304B1, a Finnish patent originally used a term meaning “circular.” The English translation rendered it as “round.” The Opposition and Appeal Division ruled the two terms were not legally equivalent, the claim scope was found to have shifted, and the patent was ultimately revoked.
Then there’s the case that reached the U.S. Federal Circuit itself: IBSA Institut Biochimique v. Teva Pharmaceuticals. An Italian priority application used the term “semiliquido.” The filed English translation rendered it as “half-liquid” instead of “semiliquid.” Even after IBSA later submitted a corrected, certified translation showing the accurate term, the court found the original claim indefinite under 35 U.S.C. § 112, and the patent was invalidated. The correction came too late to matter.
What These Cases Have in Common
| Case | Mistranslated Term | Jurisdiction | Outcome |
|---|---|---|---|
| TEMA v. Schluter Systems | “and” → “or” | Italy | Claim scope broadened, patent challenged |
| EP1397304B1 opposition | “circular” → “round” | Europe (EPO) | Claim scope altered, patent revoked |
| IBSA v. Teva Pharmaceuticals | “semiliquid” → “half-liquid” | United States | Claim found indefinite, patent invalidated |
| EP1835015 appeal | Unclear MT-generated paragraph | Europe (EPO) | Application required re-examination via certified translation |
None of these were caused by a translator inventing new meaning. In several documented cases, the failure traced back to machine translation output that was never adequately reviewed by someone with both linguistic and patent-drafting expertise. The European Patent Office’s own Patent Translate system carries an explicit disclaimer that its output is not legally binding and is meant only to convey the gist of a document — a caveat that gets lost the moment a translated draft is treated as filing-ready. ⚙️
Why Patent Language Breaks Machine Translation Engines
Generic AI translation tools are trained primarily on everyday language patterns. Patent drafting is the opposite of everyday language — it’s deliberately dense, deliberately repetitive, and deliberately precise, because every clause is written to withstand adversarial scrutiny years later. A neural engine optimizing for fluency will often “smooth out” phrasing that a patent attorney wrote to be intentionally rigid, and that smoothing is exactly where scope gets lost or gained.
Terminological inconsistency compounds the problem. A term that must stay identical across dozens of claims can be rendered two or three different ways by an AI engine depending on sentence context, and in patent law, inconsistency is often read as intentional differentiation. One claim narrowed by a single swapped term can cascade into a defense that undermines the entire filing.
The volume of filings only raises the stakes further. According to WIPO’s most recent global filing report, patent applications worldwide reached roughly 3.7 million in 2024, growing faster year-over-year than at any point since 2018. More companies are filing across more jurisdictions than ever, and a larger share of those filings depend on a translated version of the original claim becoming the legally authoritative text in the target country.
The Damage Rarely Stops at Rejection
In several EPO member states — including Italy, Poland, and Turkey — national courts rely on the local-language translation when assessing infringement, not the original filing language. That means a translation error doesn’t just risk procedural delay; it can directly determine whether an alleged infringer walks away free. In one Netherlands proceeding involving EP2029941, a first-instance court limited the scope of patent protection based explicitly on the wording of the Dutch translation, regardless of what the original claim intended.
Procedural fallout is its own cost center. In an EPO Technical Board of Appeal case involving EP1835015, a machine-translated paragraph was found so unclear that no unambiguous technical meaning could be extracted from it at all. The application had to be re-examined using a certified translation, adding delay, legal fees, and uncertainty to a filing that should have been routine. 💸
Where AI Still Belongs in the Process
None of this means machine translation has no place in patent workflows — it clearly speeds up first-pass drafts and helps teams triage which foreign filings need urgent attention. The failure point isn’t the technology itself; it’s treating unreviewed AI output as filing-ready legal text. Every case above involved a translation that looked complete and professional on the surface, which is precisely what made the error so easy to miss internally.
A workflow that actually holds up under litigation typically layers three things: an initial AI-assisted draft for speed, a linguist with genuine technical fluency in the invention’s field, and a final review focused specifically on claim-term consistency across the entire document — not just fluency or readability. Teams that skip that third layer are the ones showing up in the case law above.
What This Means for Teams Filing Internationally Right Now 🔍
If your organization is preparing multi-jurisdiction filings this year, the question worth asking isn’t whether AI tools were used somewhere in the pipeline — nearly every workflow touches one at some point now. The real question is whether a qualified human reviewer with patent-specific expertise checked every claim term against the original before it became the legally binding text in a foreign court. The companies in the cases above didn’t lose their rights because they used technology. They lost them because nobody caught the gap between what the technology produced and what the law required.
Sources:
(source: https://www.wipo.int)
(source: https://www.epo.org)
(source: https://www.aeonlaw.com)
(source: https://www.gorodissky.com)