Patent Drafting Gets Faster With AI, but Weaknesses Can Surface Years Later – KoreaTechDesk | Korean Startup and Technology News

At first glance, a patent draft produced by generative AI in minutes can feel complete enough to file. But patents are not judged at the moment of creation; they are tested later, when they become commercially important and someone starts actively looking for ways to challenge or invalidate them. As AI reduces the cost and effort of drafting, Korean deep-tech startups face a quieter but more serious risk: the growing gap between fast-generated language and the technical rigor needed for long-term patent strength.

Generative AI Is Making Patent Drafting Easier to Attempt

Generative AI is moving deeper into professional patent workflows. Dedicated systems now assist with invention disclosure, prior-art research, claim preparation, specification drafting, prosecution support, and document review.

A March 2026 Practical Law evaluation published by Reuters examined 10 commercially available generative AI patent-drafting tools. The assessment found meaningful efficiency gains, particularly in human-in-the-loop systems where practitioners iteratively review and shape the output.

And yet, while the same evaluation estimated that automated drafting tools can typically generate around 70% to 80% of a draft, the remaining portion still requires legal judgment, client-specific strategy, and quality control. That is because fully automated systems still generally do not produce accurate, filing-ready applications without significant practitioner intervention.

Now, for startup founders, it means that being able to generate something that looks like a patent application does not mean it accurately captures the invention or anticipates how the patent will be tested later.

Samuel Judge, Partner at Page White Farrer, is a European Patent Attorney and Chartered Patent Attorney whose practice spans chemical, pharmaceutical, mechanical, and computer-implemented technologies.

As discussion on Europe patent strategy for Korean deep-tech startups continues with KoreaTechDesk, Judge raised a concern he has already encountered among startups.

“I have encountered startups who have fallen into the trap of believing that mass-market generative AI systems can draft a patent application.”

And he was extremely direct about what he had seen in those outputs.

“The standard of the output from such systems is extremely poor.”

In this case Judge was referring specifically to mass-market generative AI systems, not patent-specialized AI tools used within professional workflows.

Patent Authorities Are Allowing AI but Keeping Humans Responsible

The regulatory direction emerging in 2026 does not reject AI-assisted patent drafting. Instead, patent authorities are making responsibility increasingly explicit.

The European Patent Office added a dedicated statement on artificial intelligence to its 2026 Guidelines for Examination. The EPO recognizes that AI can support quality and efficiency, but it states that applicants remain responsible for the content of patent applications and submissions regardless of AI assistance.

Korea’s Ministry of Intellectual Property issued its own guidance in June 2026 on proper patent applications in the AI era. The ministry warned applicants against relying on generative AI output without checking its truthfulness and technical feasibility.

The guidance identifies a particularly sensitive problem for deep-tech companies. Generative AI can produce technical descriptions, claimed effects, candidate compounds, or experimental results that appear plausible even when they have not actually been established.

For pharmaceutical and advanced-material inventions, the ministry warned that unverified AI-generated candidate materials or effects can create problems with demonstrating that an invention can actually be carried out. It also cautioned against presenting AI-generated experimental results as genuine test results.

This creates a different risk profile than an ordinary hallucination in an email or internal presentation.

A false sentence in routine business writing can usually be deleted. A technically inaccurate statement embedded inside a patent application may remain relevant during examination and later scrutiny of the resulting patent.

A Korean Battery Experiment Shows How Plausible Errors Can Enter a Draft

A 2025 study published by Chung-Ang University’s Institute of Legal Research provides a useful Korean example of the problem.

Researcher Shin Sang Hoon used ChatGPT to generate a patent specification based on a recent academic paper in the secondary-battery field. According to the study, the generated patent specification introduced a multilayer structure that was absent from the research paper used as its technical source.

The author identified the additional structure as a likely AI hallucination and then examined the resulting specification as if it were being evaluated during actual patent examination.

Now, this study highlights a key risk for engineering-led startups: a model can introduce technical material that appears consistent with an invention even when it did not originate from the underlying R&D.

For startup management, verification therefore needs to go beyond checking grammar and formatting. Engineers and patent specialists need to know which technical statements came from the inventors, which were inferred, and which may have been introduced by the system itself.

The Real Quality Test May Begin When Someone Wants the Patent to Fail

The risk becomes more interesting when patent drafting is viewed through the eyes of an opponent rather than the applicant.

Judge has represented clients in European patent proceedings and described one of the first issues he investigates when examining another party’s patent.

“When opposing, and when writing third-party observations against a competitor’s application, the first ground that I investigate is whether the case contains any added subject-matter.”

His next point reveals why apparently small drafting or prosecution decisions can become commercially important later.

“If the application was drafted poorly at the outset, or if a mistake was made during prosecution, then I may have an easy-to-raise and very powerful objection at my disposal.”

Even after a European patent is granted, it can still be challenged through an opposition process. These challenges may argue that the invention is not patentable, that the application does not describe the invention clearly enough, or that the granted claims include information that was not originally disclosed.

And depending on the outcome, the patent may be upheld as granted, amended through narrowing or other changes, or completely revoked.

The European Patent Office’s 2026 Quality Action Plan shows that the issue remains important enough to influence institutional learning. The EPO said it will use opposition decisions involving added subject matter to improve examination, particularly in technically complex cases involving combinations of features.

This creates a crucial distinction for startups adopting AI drafting tools: the first test may only be about whether the system can produce a coherent patent application at all, but the more important test comes later when a competitor, opponent, examiner, or tribunal actively probes the document for its weakest technical and legal assumptions.

And these two stages reward fundamentally different capabilities.

AI Can Produce Language Faster Than It Can Acquire Deep-Tech Judgment

As inventions increasingly span multiple technical disciplines, the challenge becomes even more pronounced. Judge illustrated this with a semiconductor case that crossed chemistry, device engineering, and software.

“Chemical processes enabling fabrication of the device were central to the invention.”

But the patent strategy could not stop at the fabrication process.

“It was also important to include coverage for operating the device, which involved machine learning aspects.”

That combination required expertise beyond a single technical field.

“I collaborated with one of our firm’s AI specialists to refine the final claim set,”

Judge recalled.

The example highlights a growing challenge in deep-tech patent strategy: even though a patent is drafted as a single document, the underlying invention often spans multiple disciplines and does not fit neatly within just one technical field.

A semiconductor technology can involve materials chemistry, fabrication processes, device architecture, control systems, and machine learning. Similar cross-disciplinary combinations can appear in biotechnology, diagnostics, robotics, advanced manufacturing, and AI-enabled materials development.

Korea’s Ministry of Intellectual Property has already expanded its AI patent examination guidance to address newer categories such as physical AI, on-device AI, generative AI, and inventions using AI in chemical contexts. The EPO’s 2026 quality program also identifies a growing number of applications combining AI or computer-implemented features with other technical elements.

In that environment, the difficult part of drafting may increasingly be deciding what technical relationships need to be captured, rather than producing the sentences that describe them.

Specialized Patent AI Is Moving Into Korean Corporate Workflows

Korea’s industry is already showing what a more controlled model of AI adoption may look like.

In August 2026, Electronic Times reported that LG AI Research and Korean industrial IP AI company WERT Intelligence had begun applying a patent-specialized AI model to actual IP workflows. Their project aims to connect activities including invention disclosure, patent searching, drawing analysis, specification drafting, corporate IP teams, and external patent practitioners within a shared workflow.

LG AI Research IP Strategy Leader Yoo Kyung-jae described the shift as moving beyond individual employees simply using generative AI. According to the interview, organizations now need to design systems around patent-document accuracy and responsibility for results.

The initiative also shows why a specialized deployment is different from simply asking a general chatbot to draft a patent application.

It depends on patent-specific infrastructure, controlled datasets, professional review, and close collaboration between corporate IP teams and patent practitioners. LG AI Research also emphasized security as a core requirement, since patent workflows often involve invention details that have not yet been disclosed publicly.

Most importantly, the productivity goal is not about replacing people in the process. By reducing repetitive drafting work, it allows human specialists to focus more on tasks that require legal and technical judgment.

This may ultimately be the more practical model for Korean startups exploring AI-based patent tools.

Human-in-the-Loop Patent Drafting May Be More Important Than the Tool Brand

Research on large language models suggests that advanced systems can already generate technically coherent patent claims when operating under controlled conditions. At the same time, commercial patent-focused tools are improving quickly, and evaluations such as those reported in Reuters Practical Law point to measurable efficiency gains in drafting and review workflows.

Still, the central question remains the same: where does generation end and legal or technical judgment begin?

A model can assist with structuring invention disclosures, searching prior art, drafting initial language, flagging inconsistencies, and speeding up document production. However, it cannot independently ensure that a startup has identified the commercially critical aspects of an invention, maintained technical accuracy across domains, managed cross-disciplinary dependencies, or anticipated the kinds of vulnerabilities a future challenger might exploit.

For founders, this changes how AI efficiency should be measured. While saving drafting hours is useful, those gains can be misleading if they come from skipping technical interrogation, inventor interviews, expert review, or quality control, because the cost is often pushed downstream to a later stage where fixing issues becomes harder.

The Patent Should Be Reviewed for the Day Someone Wants to Break It

Generative AI is likely to become a normal part of patent work because the efficiency gains are too significant to ignore. The more important competitive question will be how companies build verification processes and specialist judgment around that automation.

A startup may only need a few hours to produce an impressive first draft. But that patent may ultimately need to withstand years of technical development, commercial negotiation, examination, and adversarial scrutiny.

And that shift is raising the bar for how the work is managed.

Instead of asking only whether AI can draft a patent, founders need to confront a harder question: if this patent becomes valuable enough that a competitor eventually looks for ways to weaken it, who made sure the assumptions inside today’s faster draft were actually strong enough to survive that level of scrutiny?

Key Takeaway

• Generative AI can accelerate patent drafting, but drafting speed does not establish patent durability. Commercial tools already reduce first-draft workload, while current evaluations still find substantial need for specialist review and quality control.
• The European Patent Office keeps responsibility with applicants even when AI assists drafting. Its 2026 Guidelines recognize AI efficiency while making clear that parties remain accountable for application content and compliance.
• Korea is warning founders to verify AI-generated technical claims. The Ministry of Intellectual Property specifically cautions against unverified technical effects, candidate materials, and experimental results appearing in AI-assisted patent applications.
• Patent vulnerabilities may not surface until a patent is actively scrutinized or challenged, when prior art, added subject matter, or disclosure gaps are deliberately examined.
• Modern deep-tech patents increasingly require multidisciplinary judgment. Semiconductor and similar cases often combine materials science, fabrication processes, device engineering, and machine learning, requiring cross-domain expertise.
• Korean industry is moving toward specialized, human-in-the-loop patent AI. Enterprise workflows are integrating patent-focused systems while maintaining human oversight, security controls, and professional responsibility.
• AI should accelerate patent work without weakening accountability. The key measure is not drafting speed, but the strength of verification, technical validation, and expert review behind the final application.

---

– Stay Ahead in Korea’s Startup Scene –
Get real-time insights, funding updates, and policy shifts shaping Korea’s innovation ecosystem.
➡️ Follow KoreaTechDesk on LinkedIn, X (Twitter), Threads, Bluesky, Telegram, Facebook, and WhatsApp Channel.

---

🤝 Looking to connect with verified Korean companies building globally?
Explore curated company profiles and request direct introductions through beSUCCESS Connect.