On February 26, 2026, the Ministry of Culture, Sports and Tourism (the “MCST”) and the Korea Copyright Commission announced the final version of the Guidelines on Fair Use Regarding Generative AI Training Based on Copyrighted Works (the “Guidelines”). Finalized following review by the “2025 AI-Copyright System Improvement Working Group” and after collecting opinions from relevant stakeholders, the Guidelines present the standards and considerations in determining whether the fair use provision (Article 35-5) under the Copyright Act applies to generative AI training.
|
1.
|
Key Details of Guidelines
Totaling 82 pages in length, the Guidelines set out the factors for determining fair use and introduce reference cases based on the Copyright Act and court precedents, and are organized into four sections:
|
|
Section
|
|
I. Generative AI
II. Possibility of Copyright Infringement by Generative AI Training
III. Commentary on Fair Use
IV. Measures to Prevent and Manage Potential Copyright Disputes Involving Generative AI
|
In particular, Section III., “Commentary on Fair Use” explains the four factors for determining fair use in the context of generative AI training, which are based on Article 35-5(2) of the Copyright Act:
|
(i)
|
The purpose and nature of the use;
|
|
(ii)
|
The type and use of the copyrighted work;
|
|
(iii)
|
The used portion’s share and importance within the entire copyrighted work; and
|
|
(iv)
|
The impact of the use on current market or its value, including the potential market or value of the copyrighted work.
|
The Guidelines set out specific considerations for each of these four factors and the circumstances that may be favorable (or unfavorable) to the recognition of fair use. In a press release dated February 26, 2026, the MCST explained that the Guidelines intended to clarify that AI training carried out for commercial use or through web crawling is not automatically excluded from fair use, and that the determination should be based on a consideration of favorable and unfavorable circumstances across the four factors.
Section III. also presents examples of hypothetical fair use cases based on the MCST’s commentary on each of the above four factors. However, in its press release and the Guidelines, the MCST makes clear that the Guidelines and examples are not authoritative interpretations by the MCST or the Korea Copyright Commission. Rather, the MCST emphasizes that the determination of fair use must be assessed on a case-by-case basis. The MCST also acknowledges that the details of the fair use principle may change with laws, regulations, court precedents, and/or technological developments, and that the Guidelines are not definitive until a court renders a final judgment in a specific case. Thus, continued monitoring on Korean and foreign court decisions on fair use is crucial.
|
2.
|
Key Changes and Implications
The final version of the Guidelines remains largely the same as the previous draft, although relevant examples, cases, and explanations have been added. Some key changes that may be of particular relevance to AI developers and rights holders are as follows:
|
|
(1)
|
New examples have been added about whether fair use may be recognized for training general-purpose AI foundation models.
The Guidelines set out a scenario in which a for-profit company trains a general-purpose AI foundation model using publicly available online postings and purchased books (except where access is restricted, the copyright holder has clearly refused the use of his/her copyrighted works for AI training, or the materials are posted illegally or without authorization). Based on this scenario, the Guidelines provide examples of circumstances that may weigh either in favor of or against recognition of fair use under the four-factor analysis.
The key exemplary considerations for each factor are summarized below:
|
|
Factors
|
Key Considerations
(Favorable and Unfavorable Circumstances to Recognition of Fair Use)
|
|
Purpose and Nature of Use
|
-
Favorable: (i) AI training is intended not to reproduce or replace the expressions in online postings and purchased books but to develop general conversational and generative capabilities, and (ii) technical measures and other safeguards are implemented to refuse requests that would generate outputs identical or similar to the copyrighted works used for training.
-
Unfavorable: AI model training, or services resulting from that training, generate sentences, compositions, or expressions that are substantially the same as or similar to those in a specific post or book, or they absorb the expressive essence of the original text.
|
|
Type of Copyrighted Work
|
-
Favorable: If the materials used for AI training are routine, non-professional expressions with relatively low creativity, such as short phrases from comments or reviews posted on social media intended mainly to convey facts or information.
-
Unfavorable: If the copyrighted work used in AI training possesses high literary or artistic creativity and a unique expressive style.
|
|
Quantity Used for Training
|
-
Favorable: If the AI model’s architecture renders the use of hundreds of millions of data entries unavoidable, and if the amount and scope of copyrighted works used are reasonably limited to what is necessary to achieve the goal of developing general language understanding and generation capabilities.
-
Unfavorable: If the use of the copyrighted works exceeds the scope reasonably necessary to achieve the above purpose in terms of quantity and quality.
|
|
Impact on Markets
|
-
Favorable: If (i) training results or services based on such results do not substantially replace the reading and sale of the original copyrighted work, and (ii) the copyright holder is difficult to identify, for example in the case of social media posts or personal blog entries with low creativity, or routine copyrighted works created for non-profit purposes.
-
Unfavorable: If AI training or the provision of services is likely to reduce sales of copyrighted works, cause economic harm, or deprive users of opportunities to license the original copyrighted works, creating a risk of market substitution or further deterioration.
|
|
(2)
|
With respect to the purpose and nature of use, “execution of an agreement” and “royalties,” which had been mentioned as factors for determining fair use in the draft Guidelines, were revised, and a case precedent downplaying the importance of for-profit/non-profit purpose of the use has been added.
Draft Guidelines had cited Supreme Court Decision 2021Da272001, rendered on July 11, 2024, which addressed whether anti-copying measures to avoid infringement of the copyright holder’s rights had been taken, to explain that (i) “in addition to the anti-illegal copying measures for the copyrighted works used, efforts to obtain a license to use the relevant copyrighted works (e.g., attempt to enter into a license agreement, negotiation of royalties) may be considered in determining fair use,” and (ii) even in the context of generative AI training, “the degree of AI developer’s efforts to enter into an agreement to purchase data or pay royalties may serve as a factor in determining fair use.”
However, the final Guidelines deleted these examples of “efforts to obtain a license” (including “attempt to enter into a license agreement” and “negotiation of royalties”). The Guidelines now state that “efforts to obtain a license to use a copyrighted work may be considered as a reference when reviewing the overall circumstances in case of a dispute.”
Moreover, while the draft Guidelines cited only a lower court decision holding that the scope of fair use allowed for for-profit purposes is considerably narrower than that for non-profit purposes (Seoul Central District Court Decision 2021Gahap512773 rendered on January 14, 2022), the final version of the Guidelines includes not only that decision but also another lower court decision that, considering the legislative history related to the purpose and nature of use, took the opposite stance and held that for-profit/non-profit purpose of the use has low importance when determining fair use (Patent Court Decision 2024Na10249 rendered on January 23, 2025).
|
As mentioned above, the Guidelines do not constitute authoritative interpretation by the MCST or the Korea Copyright Commission, and it’s expected that the MCST will continue to update the Guidelines to reflect new case precedents and technological developments. Given the potential for further update and amendment to reflect changes in the laws, court precedents, and evolving technologies, interested parties are advised to monitor standards and practices, as well as court decisions in Korea and foreign countries, particularly with respect to the factors used to assess fair use.
[Korean Version]
