South Korea sharpens the line on inventorship in AI fine-tuning

On June 19, 2026, the Korean Intellectual Property Office (KIPO) released a mid-year supplement to its Examination Guidelines for AI-Related Inventions, responding to the rapid rise of AI-generated and AI-assisted R&D outputs. The practical question is no longer simply whether AI may be used in research. It is how examiners will separate inventorship from tool use when the claimed advance rests on fine-tuning, parameter retraining, domain adaptation, or the selective shaping of model outputs.

The signal from the supplement is fairly clear. South Korea is not moving toward recognizing the model itself as an inventor, but it is also not treating every AI-enabled result as a routine use of software. For applicants, the real task now is to explain human technical contribution with more discipline: who framed the problem, who designed the fine-tuning path, what technical judgments were made during adaptation, and which improvements can actually be supported by the specification.

Inventorship turns on human contribution, not on whether AI was involved

Debates about AI inventorship are often framed too broadly. One side worries that AI will be pushed into the inventor slot. The other treats AI as nothing more than an ordinary software aid. KIPO’s latest supplement points in a narrower and more practical direction. As long as inventorship remains tied to natural persons, the examination focus is unlikely to be whether the model “created” something by itself. The real issue is whether the application shows what the human researchers actually contributed in setting the technical objective, choosing data, defining the tuning strategy, controlling parameters, selecting outputs, and validating the result.

That matters especially for a growing class of filings in which the foundation model is not proprietary, while the claimed inventive step is said to arise from later fine-tuning and deployment-oriented optimization. Many applicants still describe that stage in a thin sentence: the existing model was retrained and applied to a particular scenario. That style is becoming riskier. Once fine-tuning becomes the center of gravity, examiners will naturally ask whether the inventive contribution came from the general model, or from a human-designed adaptation scheme that changed how the technology actually works.

Fine-tuning will not be treated as a magic shortcut if the specification lacks technical judgment

The practical shift is not limited to how inventors are named. It also affects how specifications need to be written. In many AI applications, fine-tuning has been treated as a convenient background step: industry data was used, parameters were adjusted, performance improved. KIPO’s current direction is more demanding than that. If the claimed advantage depends on a specific data curation method, a loss function choice, an iteration strategy, a prompt refinement loop, a feedback filtering rule, a hyperparameter combination, or adaptation to deployment constraints, those elements are becoming harder to leave vague.

This does not mean every applicant must disclose its full training corpus or expose trade secrets. The issue is not disclosure for its own sake. The issue is whether the file reveals enough of the human technical decision-making to support the claimed advance. Why was the fine-tuning path not a mechanical application of known practice? What problem did the chosen adjustment solve with some stability? How did the real-world operating environment shape the model structure or training strategy? When those points appear only as conclusions, the application starts to look thin at exactly the place where it claims innovation.

KIPO’s approach is likely to divide filings into tool substitution and technical shaping

Applicants should expect a sharper distinction between two kinds of AI cases. The first is tool substitution: a known workflow is replaced by a generative model, a public foundation model, or a conventional fine-tuning process, without materially reshaping the underlying technical route. The second is technical shaping: the applicant reorganizes training logic, input and output constraints, data structure, model coordination, or deployment conditions around a concrete task, producing a demonstrable technical difference.

That line matters because many commercially important AI inventions are not about building a new model from scratch. They are about deeply adapting an existing one for healthcare, robotics, industrial vision, search, compliance, manufacturing, or risk control. If the application does not show that adaptation as a concrete technical shaping process, examiners may treat it as routine implementation, ordinary optimization, or simply connecting AI to an older system. At that point, inventive step becomes harder to defend, and the named inventors may also find it harder to justify who actually contributed the inventive concept.

The next adjustment is not just better drafting but better R&D records

Many companies will respond by revising patent drafting templates. That helps, but the more important work happens earlier. If inventorship is going to be tested more closely, internal records need to become more useful. Who proposed the fine-tuning target? Who decided on the data layers and labeling rules? Who set the evaluation threshold for iteration? Who determined the deployable configuration under real computing constraints? Those facts are often scattered across experiment logs, version histories, meeting notes, and product documents rather than managed as evidence that supports inventorship.

The practical value of KIPO’s supplement is that it raises the importance of those traces. Future AI filings, especially those built around fine-tuning, sector-specific adaptation, or AI-assisted research outcomes, will increasingly become contests over the ability to explain human technical contribution with precision. Applicants who can tell that story early and clearly will be in a stronger position not only for prosecution, but also for later disputes over inventorship, validity, and the real scope of the claimed advance.