JPO and KIPO Raise the Bar for AI Patent Disclosure
Recent signals from the JPO and KIPO are best read as a shift in examination practice for AI inventions rather than a dramatic headline reform. No single new “AI patent statute” has suddenly appeared. But when the IP5 keeps deepening comparative materials on AI examination, the JPO expands its support structure for AI-related cases, and KIPO continues to formalize examiner exchange and AI-related examination frameworks, the practical message is hard to miss: an AI invention is less likely to survive on functional ambition alone and more likely to be tested on whether the specification explains a reproducible technical route.
For generative AI, large-model fine-tuning, data-processing and deployment claims, that matters immediately. Examiners are becoming more willing to ask how the claimed technical effect is actually achieved, what role the data or inference pipeline plays, which steps are indispensable, and whether the applicant has disclosed enough for a skilled person to carry the invention out without filling the core gap by guesswork. The pressure point is no longer just wording. It is evidentiary density.
This is convergence in examination practice, not a one-off AI statute
The timeline matters. In June 2024, the IP5 approved a more detailed comparison table on examination cases for AI-related inventions. The JPO also created a new “AI Advisors” support position so outside experts could help examiners stay current on fast-moving AI technologies. On the Korean side, examiner exchange with Japan has long been built around comparing examination practices, prior-art search methods and the operation of examination standards, while KIPO continues to frame AI and other fourth industrial revolution technologies as a distinct examination focus. Put together, these are not isolated announcements. They show an examination ecosystem that is building more structure, more shared materials and more confidence around how to interrogate AI specifications.
That does not mean every AI filing will face a brand-new statutory test. It means the old comfort zone is shrinking. Applicants can still argue technical character, implementation and inventive step in the ordinary way, but they should expect less tolerance for a specification that simply promises a model-driven result and leaves the decisive technical path obscure. In practice, the emphasis is moving from “can this be claimed at all?” toward “has this been disclosed with enough technical substance to justify the claim?”
Sufficiency of disclosure is being broken into concrete technical questions
In traditional filings, sufficiency often turns on structure, process steps, parameter ranges and worked examples. AI cases are different because applicants often compress the decisive technical contribution into broad phrases such as “model training,” “data processing” or “intelligent determination.” That drafting style is becoming harder to defend. A safer specification now anticipates the questions an examiner is likely to ask: what kind of input data is involved, how that data is screened, cleaned, labelled or transformed, what role the model or rule engine actually performs, how the output is tied to a verifiable technical effect, and what testing or evaluation supports that effect.
None of this automatically requires full source code, every training sample or the wholesale disclosure of a proprietary dataset. The issue is not maximal disclosure for its own sake. The issue is whether the invention’s asserted advantage actually depends on technical conditions that the specification has failed to explain. If novelty and effect are really driven by a particular data architecture, labelling scheme, fine-tuning route, retrieval pipeline or inference-control logic, then a bare statement that a skilled person can implement the invention will carry less weight than it used to.
Training-data disclosure is becoming issue-specific rather than optional
The practical question many applicants ask is whether they will now be forced to disclose training data in full. That is the wrong way to frame the issue. A blanket disclosure obligation is unlikely; what is becoming harder is to keep silent about data features that are causally linked to the claimed result. If the invention is said to improve accuracy, reduce hallucinations, increase robustness, limit bias or make a model workable in a defined industrial setting, examiners are increasingly likely to ask what data conditions made that effect possible. Was the result driven by public data, proprietary data or synthetic data? Were samples filtered or balanced in a particular way? Did the labelling scheme materially affect the output? How were training and validation conditions separated? Those points do not always require file-by-file disclosure, but they can no longer be treated as irrelevant background.
The same is true for large-model application cases. Claims that merely say a large language model improves document handling, optimises an industrial workflow or enhances decision support are usually too broad to carry real weight. What matters is how the inputs are constrained, how the knowledge base is built and refreshed, how retrieval and reranking operate, how confidence thresholds are set, how unsafe or non-compliant outputs are blocked, and how the model’s output is reconnected to a concrete technical process. Applicants who still rely on phrases such as “any model may be used,” “the training method is not limited,” or “the data source is not limited” should expect a rougher prosecution path in Japan and Korea.
What applicants should change now
For applicants filing across Japan, Korea and other major jurisdictions, the specification itself now deserves more attention than the headline claim strategy. First, separate the generic AI layer from the field-specific technical contribution. Examiners need to see whether the real advance sits in the model layer, the data layer, the interface layer or the underlying industrial process. Second, write the indispensable steps as part of the invention rather than hiding them in optional embodiments. Third, preserve support for the technical effect with material that can actually be used later in prosecution: comparative results, parameter ranges, error-control logic, resource-consumption changes, fallback handling and implementation boundaries. Fourth, when office actions arrive, do not defend the invention with marketing language about performance or convenience. Tie the alleged effect back to what the specification actually discloses.
There is also a broader drafting consequence. The old cross-border habit of starting from a deliberately thin, broad master specification and narrowing later is becoming riskier for AI inventions. If the original filing does not clearly spell out the relevant data conditions, training logic and implementation boundaries, later amendments may run straight into added-matter, support or plausibility problems. For AI patent work, writing for provability first and breadth second is no longer conservative drafting. It is increasingly the safer default.
