AI patent filings keep climbing at a pace that would have seemed unlikely a few years ago. WIPO data shows generative AI patent families jumping from roughly 14,000 in 2023 to more than 37,800 in 2025, with the combined output of 2024 and 2025 already exceeding the previous decade. At the USPTO, generative AI applications published in the first nine months of 2026 already outstripped the full-year total for 2025. Similar growth appears in European filings. The volume is real, but so is the scrutiny. Both the USPTO and the EPO have sharpened their expectations for what an AI claim must actually show, and translation mistakes remain one of the quieter ways applicants lose scope or face outright rejection.
The core problem is familiar to anyone who has prosecuted software or machine-learning applications. Vague functional language or an imprecise rendering of a key term can trigger a subject-matter eligibility rejection under 35 U.S.C. § 101 at the USPTO or a finding that the claimed subject matter lacks technical character under Article 52 EPC. Even when eligibility is cleared, a translation that inadvertently narrows a claim or introduces inconsistency between the claims and the description can shrink the enforceable right or create an added-matter problem that is expensive or impossible to fix later.
At the USPTO the 2024 guidance update on subject-matter eligibility, together with later examiner reminders, makes the practical test clearer. Claims that merely apply a known machine-learning technique to a new data set without identifying a specific technical improvement are vulnerable. Examiners look for a concrete mechanism: how the architecture, training process, or inference step actually improves computer function or another technical field. Generic statements that the system “uses AI” or “trains a neural network” rarely suffice. The specification must support the claimed improvement with enough detail that a skilled person can understand what was invented and how it works. Recent practice also emphasizes that inventorship remains limited to natural persons; AI systems cannot be named as inventors.
The EPO applies a different but equally rigorous filter. Under the COMVIK approach, mathematical methods and computer programs as such are excluded. An AI invention clears the first hurdle if it involves technical means. The harder test is whether the AI features contribute to a technical solution to a technical problem. Faster execution of a known task is usually not enough. Examiners look for effects such as improved image quality, more accurate sensor data processing, reduced power consumption on edge devices, or better control of a physical process. The 2025 updates to the Guidelines reinforce that computational models and algorithms can contribute to technical character when they are applied in a technical field or adapted to a specific technical implementation. Sufficiency of disclosure under Article 83 EPC has also become more demanding: architecture, training methodology, and the characteristics of the training data that are essential to the technical effect generally need to be described so that a skilled person can reproduce the invention without undue burden.
These differences matter when the same invention is filed in multiple jurisdictions. A claim set drafted primarily for the USPTO’s “improvement” language may need tightening or reframing for the EPO’s technical-effect requirement. Conversely, an EPO-friendly claim that leans heavily on a particular hardware interaction can look overly narrow in the United States. The translation step is where many of these adjustments go wrong.
Real cases illustrate the risk. In the opposition against EP 2022349, which originated from a Chinese-language PCT application, inconsistencies in the English translation of terms for air inlets, porosity, and structural elements became central to added-matter arguments. Opponents successfully challenged certain corrections, and the patent faced significant pressure. Similar problems appear with everyday patent terms. Rendering “comprising” as the closed “consisting of” can open a design-around path that was never intended. Translating a Chinese term that implicitly allows one or more elements as a singular “one” can narrow the claim without anyone noticing until litigation. Inconsistent terminology between claims and description—something machine translation still produces regularly—invites clarity and written-description rejections under § 112 or corresponding European provisions.
Practitioners who handle multi-jurisdictional AI portfolios therefore treat translation as part of claim strategy rather than a downstream clerical task. The preferred approach starts with a carefully controlled English or source-language claim set that already anticipates both USPTO and EPO tests: specific technical problems, explicit mechanisms, measurable or at least explainable improvements, and consistent terminology. Specialized translators who understand both the technology and the claim-construction rules of the target offices then produce the foreign-language versions. Glossaries are maintained across the family so that the same concept receives the same rendering. Where national-phase entry or validation requires further amendment, the original disclosure is checked to ensure support remains intact.
The payoff is practical. Applications that clear eligibility and clarity on first or second action move faster and cost less. Scope that is neither accidentally narrowed nor overbroad survives challenge better. In a field where filing volumes are high and examination resources are stretched, precision at the claim-and-translation stage is one of the few levers applicants fully control.
For organizations that file AI inventions across the United States, Europe, and additional markets, the same discipline extends to every language of the portfolio. Artlangs Translation has spent more than twenty years handling exactly this kind of work. With a network of over 20,000 professional linguists covering more than 230 languages, the company regularly supports patent filings, technical specifications, and related documentation for AI and software inventions. Its teams also manage video localization, short-drama subtitle work, game localization, multilingual dubbing for short dramas and audiobooks, and large-scale data annotation and transcription. The combination of domain knowledge, consistent terminology management, and long experience with high-stakes legal and technical text helps applicants keep claim scope intact as they move from one patent office to the next.
