How are AI inventions examined in Korea?

Chanjoong Kim, Partner · Published 26 October 2026

Last verified 9 October 2026

Examination items at a glance

As of October 2026. The section numbers are those of Chapter I, “Artificial Intelligence Field”, of the Examination Practice Guidelines by Technology Field (“the Practice Guidelines”). [2]

Of the provisions in the table, Art. 42(3), subparagraph 1, Art. 42(4), subparagraph 1 and Art. 29(2) are the provisions stated in the list of cases in the Practice Guidelines. [2] The heading of section 2.2.2 of the Practice Guidelines, “Clear and Concise Description of the Claimed Invention”, is the requirement of Art. 42(4), subparagraph 2. [1][2] Section 3.1.1 of the Practice Guidelines states the subject matter eligibility requirement as “a creation of technical ideas utilizing a law of nature”. This is the definition in Art. 2, subparagraph 1. [1][2]

Article 2 of the Patent Act is not among the grounds for rejection in Art. 62. For an application that is not an invention, the notice of grounds for rejection cites the main clause of Art. 29(1). The Patent Examination Guidelines (특허·실용신안 심사기준), Part 3, Chapter 1, 4.2, state this. [1][5] A violation of Art. 29 or of Art. 42(3) or (4) is a ground for rejection (Art. 62, subparagraphs 1 and 4). [1]

The Practice Guidelines are not a statute. They are an examination practice guide that examiners follow. Matters that the Practice Guidelines do not explain follow the Patent Examination Guidelines or the Examination Guidelines for Computer-Implemented Inventions (컴퓨터 관련 발명 심사기준). [2]

There is no provision only for AI inventions, and Chapter I of the Practice Guidelines is the standard

The Korean Patent Act does not regulate AI inventions separately. The word “인공지능” (artificial intelligence) does not appear even once in the Patent Act in force from 11 November 2025. [1] AI inventions are also examined under the definition of an invention (Art. 2, subparagraph 1), the patentability requirements (Art. 29) and the description requirements for the specification (Art. 42). [1][2]

The standards specific to AI are in Chapter I of the Examination Practice Guidelines by Technology Field issued by MOIP. Chapter I applies to applications for inventions that require artificial intelligence technology based on machine learning. The Practice Guidelines call these “AI-related inventions” (인공지능 관련 발명). [2]

Chapter I was established in December 2020. It was revised in December 2021, May 2023 and January 2026. The division in charge is the Artificial Intelligence and Big Data Examination Division (인공지능빅데이터심사과). [2]

Chapter I has no Addenda or transitional measure that sets from which applications a revised standard applies. The only application note with a date is “Applicable to applications filed on or after July 1, 2014” in 2.2.2.2 (2) ②. This article states the standards in the edition published in March 2026. [2]

The Practice Guidelines divide AI inventions into two types. An AI learning model invention (AI 학습모델링 발명) is an invention whose feature is generating a trained model on the basis of training data and a learning model. An AI application invention (AI 응용 발명) is an invention whose feature is performing a specific function by applying training data or a trained model. [2]

The amended Patent Act scheduled to enter into force on 16 March 2027 (Act No. 21955) does not change the provisions used in this article. The wording of Art. 2, Art. 29, Art. 33, Art. 42, Art. 47, Art. 62, Art. 67-2 and Art. 208 is the same as in the current Act. [1]

The description of the invention must state the correlation between input and output

The description of the invention must state the concrete content of the artificial intelligence technology that the invention implements. The basis is the Patent Act, Art. 42(3), subparagraph 1. The description must be clear and detailed so that a person skilled in the art can easily carry out the invention. [1]

Section 2.1.1 of the Practice Guidelines gives examples of concrete means. They are the training data, the data preprocessing method, the learning model and the loss function. [2]

There is a case where the concrete means do not have to be stated explicitly. This is the case where they can be clearly understood by a person skilled in the art in view of the common general technical knowledge at the time of filing. The Practice Guidelines state that, in this case, it is not found for that reason that the invention cannot be easily carried out. [2]

Section 2.1.2 (2) of the Practice Guidelines requires the correlation. If the correlation between the input data and the output data of the trained model is not concretely stated, the enablement requirement is regarded as not met. [2]

The correlation is concretely stated where the following four conditions are met. [2]

1. The training data is specified.
2. Between the characteristics of the training data, there is a correlation for solving the technical problem of the invention.
3. The learning model to be trained or the training method is concretely stated.
4. A trained model for solving the technical problem is generated by that training data and that training method.

However, there is an exception. This is the case where a person skilled in the art can, with the common general technical knowledge at the time of filing, infer or understand that correlation from the embodiments. In this case, the enablement requirement can be regarded as met. [2]

The example that the Practice Guidelines give is a case where the input data is only listed. The claim controls the temperature of a house by using weather data and environment data as training data. The description of the invention only lists the environment data as input data and does not state its correlation with the output data. The Practice Guidelines regard the enablement requirement as not met in this case. [2]

Section 2.1.2 of the Practice Guidelines gives two more examples of violation. One is the case where a technical step or function is stated only abstractly. It is the case where there is no statement of how it is realized by hardware or software. The other is the case where the invention is expressed only by functional block diagrams or flowcharts. Both are limited to the case where a person skilled in the art cannot clearly understand the invention even in view of the state of the art at the time of filing. [2]

Section 2.1.3 of the Practice Guidelines adds standards by type. [2]

The same exception is attached to the standards for data preprocessing and reinforcement learning. It is the case where a person skilled in the art can, with the common general technical knowledge at the time of filing, clearly understand the invention from the embodiments. [2]

Claims can be written as a method, a device, a recording medium or a program, and some claim endings make the claim unclear

An AI-related invention can be stated in a claim as a method invention or a product invention (section 2.2.2.2 of the Practice Guidelines). [2] A product invention can be stated as a product (device) specified by its functions. The Practice Guidelines add five more formats. [2]

• A claim to a recording medium on which a computer program is recorded
• A claim to a “computer program stored on a recording medium” that is combined with hardware
• A claim to a recording medium on which a data structure is recorded
• A claim to a computer program that is stored on a recording medium and implements a learning model
• A claim to a product (device) that uses a learning model

The second format carries the note “Applicable to applications filed on or after July 1, 2014”. A claim written as a computer program that is not stored on a medium is not allowed. The Practice Guidelines give the reason that the claim is directed to a computer program per se. [2]

In a claim to a program that implements a learning model, stating only the learning model is not enough. The Practice Guidelines state that concrete means must also be stated in the claim. The concrete means are the training data, the data preprocessing, the loss function and the like. It is sufficient that at least one concrete means necessary for training is included. [2]

However, the description of the invention must state more. The description requirement can be met only when concrete means such as the loss function and the training data are stated in addition to the structure of the learning model. [2]

Section 2.2.2.3 of the Practice Guidelines gives examples of claims that are not clear. [2] The basis for clarity is the Patent Act, Art. 42(4), subparagraph 2. [1]

If the entity that performs the invention cannot be identified, there can be two grounds for rejection. Section 2.2.2.4 of the Practice Guidelines states that such a claim, in principle, also does not implement the concrete means of the subject matter eligibility requirement. However, the examiner does not always notify both grounds for rejection together. If one notice is likely to resolve the other ground for rejection as well, there is no need to notify both. [2]

The treatment of each claim ending is covered separately in the article “How are computer program claims written in Korea?”.

Subject matter eligibility is, in principle, assessed under the same standard as for software inventions

The standard for assessing subject matter eligibility of AI-related inventions is, in principle, the same as the standard for computer- and software-related inventions (section 3.1.1 of the Practice Guidelines). The Practice Guidelines state that the reason is that AI-related inventions are based on information processing by software. [2]

The standard is whether information processing by software is concretely realized by using hardware. In that case, an information processing device, its operating method, a medium on which the program is recorded and a computer program stored on a medium are inventions. A computer program per se cannot be an invention. [2]

Some cases are decided before this standard is reached (section 3.1.2 (2) of the Practice Guidelines). An invention that concretely performs control of a device is an invention. The same applies to an invention that concretely performs information processing based on the technical properties of an object. An invention that is an artificial decision or a human mental activity is not an invention. [2]

The assessment is made on the claim (section 3.1.4 (1) of the Practice Guidelines). If concrete realization by using hardware is found only in the description of the invention and the drawings and not in the claim, the claimed subject matter is not an invention. [2]

The name of the hardware alone may not be enough. Section 3.1.4 (2) of the Practice Guidelines deals with the case where “computer”, “processor” or “memory” is stated in the claim. It is the case where concrete means or a concrete method in which software and hardware cooperate is not stated. The Practice Guidelines state that, in this case, the claimed subject matter may not be an invention. [2]

The same section also gives a standard in the opposite direction. This is the case where the claim concretely states the calculation or processing of information that is specific to the purpose of use. In that case, it is acceptable that the only hardware stated is an ordinary “computer”. The Practice Guidelines state that the calculation or processing can be regarded as realized when the state of the art at the time of filing is taken into account. [2]

One more phrase is attached for AI-related inventions (section 3.1.4 (5) of the Practice Guidelines). The assessment depends on whether concrete means or a concrete method is stated in the claim. The Practice Guidelines describe those means and that method as ones “such that the same effect can be repeatedly achieved without reliance on the mental activity of a human being”. [2]

The example that the Practice Guidelines regard as an invention is a preference prediction device. It predicts content preference by using the preference information of similar users as the input of a machine learning algorithm. [2] The order of the subject matter eligibility assessment is covered separately in the article “Is software patentable in Korea?“.

Inventive step depends on whether the claim states how the artificial intelligence is implemented

The basis for inventive step is the Patent Act, Art. 29(2). [1] Section 3.2.1 (5) of the Practice Guidelines regards a mere combination of technologies used in several fields, and application to a specific field, as within the scope of ordinary creative activity. If the combination and the application involve no technical difficulty, inventive step is not acknowledged unless there are exceptional circumstances such as a remarkable technical effect. [2]

There is also a standard for arguments on effect. Section 3.2.1 (6) of the Practice Guidelines gives four effects as examples. “Rapid processing” and “processing large volumes of data” are such examples. So are “error reduction” and “accurate prediction”. The Practice Guidelines state that such effects are, in many cases, the natural effects of implementing an AI-related invention. [2]

Section 3.2.2 of the Practice Guidelines divides the examples that are the exercise of ordinary creativity by a person skilled in the art into four types. [2]

The third type has a proviso. The difference must have the effect that the result of the invention differs. If that effect is acknowledged as an advantageous effect beyond the ordinarily predictable range, inventive step can be acknowledged. [2]

Section 3.2.3 of the Practice Guidelines, in the opposite direction, states the types that are not the exercise of ordinary creativity. There are two common conditions. The technical features must be concretely specified in the claim. Those features must bring an advantageous effect beyond the predictable effect in comparison with the cited invention. [2]

All three types have a proviso. This is the case where the claim states no more than “performing data preprocessing”, “using a learning model” or “using a training result”. The Practice Guidelines state that this case may fall within the scope of ordinary creativity. [2]

The case where only the training data differs is dealt with separately. Section 3.2.3.4 of the Practice Guidelines states that it is difficult to acknowledge inventive step on a difference in training data alone. The examiner looks at whether information processing that is specific to the training data is specified. The examiner also looks at whether that difference produced an advantageous effect beyond the predictable effect. [2]

The case where the industrial field differs is also dealt with separately (section 3.2.3.3 of the Practice Guidelines). This is the case where the invention resolved a long-unsolved problem of a specific industrial field or overcame a technical difficulty. The same applies where an advantageous effect beyond the predictable effect of the change of industrial field was produced. The Practice Guidelines state that, in this case, care must be taken so that novelty and inventive step are not denied unconditionally for the reason that there is no difference in the technical features. [2]

There is a requirement for that effect. A person skilled in the art must be able to recognize it easily from objective evidence, concrete embodiments or the like stated in the description of the invention. [2]

Example: examination cases in the Practice Guidelines

Section 4 of Chapter I of the Practice Guidelines contains 15 examination cases. Six of them are shown below. [2]

A condition is attached to claim 2 of Case 4. It is the case where those properties and effects are not at a level that a person skilled in the art can derive as self-evident. [2]

Case 2 also states the scope of what can be cured by amendment. It states that adding an embodiment that proves the effect to the written amendment is an addition of new matter. It states that deleting from the claim the training data that has no correlation is an option that can be considered. [2]

Inventions that use generative AI are assessed in the same framework. Claim 1 of Case 11 obtains an answer with a fine-tuned large language model. The Practice Guidelines regarded this as a mere replacement, by an AI model, of obtaining an answer from an expert. They state that this is no more than a mere systemization of work that people did. [2]

Claim 2 of Case 11 uses information on the age group of the user. It converts the query with a prompt layout that corresponds to the age group and the selected AI model. The Practice Guidelines found that this feature is not disclosed in the cited invention and that the effect of leading to a customized answer arises. [2]

The Practice Guidelines set a limit on the use of the cases. Each case was prepared only for the patentability requirement reviewed in that case. The cited inventions and the well-known and commonly used art in the cases were adapted for explanation and may differ from actual examination. [2]

An application that named an AI as the inventor was not accepted

MOIP did not accept an application that named an AI as the inventor. At the time of the disposition, the office was named the Korean Intellectual Property Office (KIPO, 특허청). The Patent Act before the amendment of 1 October 2025 referred to this office as the “Korean Intellectual Property Office” (the “employee of the Korean Intellectual Property Office” in the proviso to Art. 33(1)). [1] Art. 33(1) of the Patent Act gives the right to obtain a patent to “a person who makes an invention or his or her successor”. [1] The written patent application must state the name and address of the inventor (Art. 42(1), subparagraph 4). [1]

MOIP invalidated the procedure of an international patent application that named an AI as the inventor. The application number is 10-2020-7007394, and “DABUS” (다부스) was stated in the inventor field. The explanation of MOIP is as follows. [3]

• February 2022: An amendment was requested for the reason that an AI is not a natural person and cannot be an inventor.
• September 2022: The procedure was invalidated because the applicant did not amend the inventor to a natural person.
• 20 December 2022: The applicant filed an administrative lawsuit against the invalidation of the procedure.
• 30 June 2023: The court of first instance upheld the invalidation of the procedure.
• 16 May 2024: The Seoul High Court, as the court of second instance, also upheld the invalidation of the procedure.

The above history is the explanation on the MOIP website page “인공지능 발명자 이슈” (AI inventor issue), last modified on 31 July 2026. This page reports the Korean litigation up to the second instance. [3]

How Korea differs

The table below places the wording of the European Patent Office (EPO) Guidelines for Examination, G-II, 3.3.1, and of the Korean Practice Guidelines side by side on two issues. It does not compare the examination results of the two offices. [2][4]

The EPO Guidelines regard the computational models and algorithms of AI and machine learning as being, in themselves, of an abstract mathematical nature. They state that such models and algorithms contribute to the technical character of the invention if they contribute to the technical solution of a technical problem. [4]

The EPO Guidelines also give a caution on terms. A term such as “neural network” may, depending on the context, refer only to an abstract model or algorithm. They state that such a term alone does not necessarily imply the use of technical means. [4]

What applicants should do

• In general, we suggest stating, for each item of input data, its correlation with the output in the specification at the time of filing. The Practice Guidelines give input data that is only listed as an example of a violation of the enablement requirement. [2]
• This content is difficult to add after filing. An amendment must be made within the scope of the matters stated in the specification or drawings originally attached to the written patent application (former part of the Patent Act, Art. 47(2)). [1] Case 2 of the Practice Guidelines states that adding an embodiment that proves the effect is an addition of new matter. [2]
• An amendment of a patent application in a foreign language has one more limit. It must also be made within the scope of the matters stated in the final Korean translation or the drawings originally attached, excluding captions in the drawings (latter part of Art. 47(2)). [1]
• For an international patent application filed in a foreign language, the documents that serve as the basis are different. The basis for the former part is the description of the invention, claims or drawings submitted by the international filing date (Art. 208(3)). The basis for the latter part is the final Korean translation or the drawings submitted by the international filing date, excluding captions in the drawings (Art. 208(4)). [1]
• After a notice of grounds for rejection is received, the times when an amendment can be made are fixed. They are the period for submission of a written argument and the time of requesting re-examination (Art. 47(1)). [1]
• Amendment of the claims is restricted when a notice of a ground for rejection caused by an amendment is received and when re-examination is requested. Only narrowing the claims, correction of clerical errors, clarification of an ambiguous description and the like are possible (Art. 47(3)). [1]
• There are cases where re-examination cannot be requested. They are the case where a decision to grant or reject has already been made following re-examination, and the case of a separational application (splitting-off). The same applies where an appeal against the decision of rejection (decision to reject a patent application) has been filed. However, the case where the decision of rejection was revoked in the appeal is excluded (proviso to Art. 67-2(1)). [1]
• We suggest stating concretely in the claim where the feature lies among the data preprocessing, the learning model and the use of the training result. “Using artificial intelligence” alone is a disadvantage in the assessment of inventive step. [2]
• We suggest stating the effect in the specification as an effect that comes from that feature, not as “the accuracy is high”.
• We suggest stating, in each step of the claim, the hardware that is the entity that performs it. The reason is that both subject matter eligibility and clarity are assessed on the claim.
• We suggest stating a natural person in the inventor field. MOIP requested an amendment for an application that named an AI as the inventor. [3]
• Please confirm the claim wording of an individual case with your Korean agent.

FAQ

Does the Korean Patent Act have a provision on AI inventions?

No. The word “인공지능” (artificial intelligence) does not appear in the Patent Act in force from 11 November 2025. The standards are in Chapter I, “Artificial Intelligence Field”, of the Examination Practice Guidelines by Technology Field of MOIP. [1][2]

Must the training data be stated in the specification?

Under the standard of the Practice Guidelines, the training data must be specified and stated. Section 2.1.2 (2) of the Practice Guidelines gives as a condition that the training data is specified. It is the first condition for the correlation between the input data and the output data being concretely stated. The exception is where a person skilled in the art can infer or understand the correlation from the embodiments. [2]

Can an invention that applies a publicly known model to a new field also be patented?

It depends on the conditions. The Practice Guidelines regard application to a specific field as within the scope of ordinary creative activity. If the application involves no technical difficulty, inventive step is not acknowledged unless there are exceptional circumstances such as a remarkable technical effect. However, the case where the invention resolved a long-unsolved problem or overcame a technical difficulty is viewed differently. The same applies where an advantageous effect beyond the predictable effect of the change of industrial field was produced. The Practice Guidelines state that, in this case, care must be taken so that novelty and inventive step are not denied unconditionally. [2]

Does a different standard apply to inventions that use generative AI or a large language model?

They are assessed in the same framework. Case 11 and Case 13 of the Practice Guidelines deal with claims that merely use generative AI. The Practice Guidelines give them as cases where it can be found that inventive step is not acknowledged. A claim that differs in the concrete feature that generates the prompt, or in the feature that post-processes the output data, is different. If a difference in effect arises, the Practice Guidelines give it as a case where it can be found that inventive step is acknowledged. [2]

Can an AI be named as the inventor?

MOIP did not accept it. MOIP requested an amendment for an international patent application that named an AI as the inventor, and invalidated the procedure when no amendment was made. According to the explanation of MOIP, the courts up to the Seoul High Court as the court of second instance upheld the invalidation of the procedure. [3]

How DEFINE IP Law Firm handles this

When drafting the specification of an AI invention, DEFINE IP Law Firm checks the description requirements first. We check whether the correlation between the input data and the output data is stated in the description of the invention. When a foreign application is brought into Korea, we check before entry whether the same content is in the original specification. For questions about the examination of AI patents in Korea, please write to contact@defineip.kr.

Sources

1. Patent Act (특허법), Act No. 21134, in force from 11 November 2025. Art. 2, subparagraph 1; Art. 29(1) and (2); Art. 33(1); Art. 42(1), (3) and (4); Art. 47(1) to (3); Art. 62; Art. 67-2(1); Art. 208(3) and (4). Compared with the version scheduled to enter into force (Act No. 21955, promulgated 15 September 2026, in force from 16 March 2027). The former wording of Art. 33(1) is from the version in force from 22 July 2025 (Act No. 20700). https://www.law.go.kr/법령/특허법 (verified 9 October 2026)
2. Ministry of Intellectual Property (지식재산처), Examination Practice Guidelines by Technology Field (기술분야별 심사실무가이드), March 2026, Chapter I, Artificial Intelligence Field (제1부 인공지능 분야), revised January 2026, 1., 2.1 to 2.2, 3.1 to 3.2, 4. https://www.kipo.go.kr/ko/kpoContentView.do?menuCd=SCD0200147 (verified 9 October 2026)
3. Ministry of Intellectual Property (지식재산처), AI inventor issue (인공지능 발명자 이슈), last modified 31 July 2026. https://www.kipo.go.kr/ko/kpoContentView.do?menuCd=SCD0201238 (verified 9 October 2026)
4. EPO, Guidelines for Examination, Part G, Chapter II, 3.3.1 Artificial intelligence and machine learning, 2026 edition (publication date shown: 31 March 2026). https://www.epo.org/en/legal/guidelines-epc/2026/g_ii_3_3_1.html (verified 9 October 2026)
5. Ministry of Intellectual Property (지식재산처), Patent Examination Guidelines (특허·실용신안 심사기준), MOIP Established Rule No. 7, in force from 13 March 2026, Part 3, Chapter 1, 4.2. https://www.law.go.kr/행정규칙/특허·실용신안심사기준 (verified 9 October 2026)