브레스저널 The Breath Journal

This article was translated automatically from the Korean original. Read the original in Korean

The gray zone of training data, can a guide cover it

곽동현·Published 2026-03-01 13:12 KST
The gap left between the publication of a fair use guide and 'use first, compensate later'
The width of the gate that data for AI training must pass through has not yet been set
The width of the gate that data for AI training must pass through has not yet been set / ⓒ Breath Journal

The Ministry of Culture, Sports and Tourism and the Korea Copyright Commission released a 'fair use guide' on February 26 covering the training of generative AI on copyrighted works. Fair use is an exception under the Copyright Act that allows works to be used without the permission of the copyright holder, and the provision the guide refers to is Article 35-5 of the Copyright Act. This is the first time the government has set out the relationship between AI training and copyright in document form.

The guide points to two main passages. That an AI reading in, storing and processing a work may amount to reproduction under the Copyright Act, so that prior permission from the rights holder is required in principle, and that a commercial purpose or a web crawling method does not by that fact alone rule out fair use.

Fair use is determined by weighing four factors together. The purpose and character of the use, the type and purpose of the work, the amount and importance of the portion used in relation to the work as a whole, and the effect on the current or future market for or value of that work. The examples in the guide are hypothetical rather than drawn from actual cases, and they are not authoritative interpretations by the Ministry or the Copyright Commission. The final ruling is made by the courts.

Reaching this point took nearly six months. The Ministry and the Copyright Commission set up a special subcommittee of the council on improving the AI-copyright system in September 2025, and from October 13 to November 2 surveyed AI developers and rights holders. After about three months of consultations with related ministries from November, they released a draft at a public briefing on December 4.

On February 25, the day before the guide was published, the second plenary meeting of the National Artificial Intelligence Strategy Committee approved the 'Republic of Korea Artificial Intelligence Action Plan,' which includes copyright tasks. It contains four core tasks. Spreading and refining the guide, reviewing ways to exempt developers of independent foundation models from criminal liability, promoting the use of gray-area works, and widening the release of public works.

The gray-area measures are the source of the controversy. Works with an established trading market, such as music, books and broadcasting, are to be opened up by stimulating transactions, while online public postings and the like, which have no market, are given an opt-out, that is, the right to refuse training. The plan is to pursue a Copyright Act amendment that would allow postings carrying no refusal notice to be used first for training under the 'use first, compensate later' principle.

The 'prior permission in principle' stated in the guide and the 'use first, compensate later' stated in the action plan set the order of use in exactly opposite ways. The two documents, issued a day apart, presented different starting points. This is also where creators objected. Sixteen organizations in the cultural content field, including the Korean Association of Broadcasting Performers' Rights, the Federation of Korean Music Performers and the Korea Broadcasters Association, issued a joint statement calling for a full review of the action plan, and the Korea Newspaper Association had earlier sent an opinion to the same effect to the Strategy Committee.

The item at issue is No. 32 of the action plan, 'Use of works for AI training and evaluation and activation of a fair distribution ecosystem,' which sits under the 'AI regulatory innovation' strategy. The design of creators' rights is placed inside an item that loosens regulation. Article 2 of the Framework Act on Artificial Intelligence includes systems that risk having a significant effect on the basic rights of authors and performers within 'high-impact artificial intelligence,' but what Article 13(2) and Article 35(1) impose on the government is a 'duty to make efforts.'

Who will calculate the level of compensation and by what standard, and what technical specification will be used to attach an opt-out notice, are not included in the guide. The exemption from criminal liability also remains at the 'review' stage, so the legislative form and scope of application lie outside the outline. Rather than filling this gap with law, the guide comes closer to covering it for the time being with an administrative document. Along with the publication, the Ministry said it would expand copyright counseling and dispute mediation support for rights holders, developers and users, and the existence of that measure also means disputes are anticipated.

Just as the Supreme Court has previously recognized state liability for damages in a mural defacement case (2012Da204587), the results of public actions surrounding creative works are reckoned in the end in court. However fine-grained the guide may be, a single lawsuit can change its interpretation.

For people who post writing on blogs and put pictures online, the one thing that changes right away is a notice. Until it is settled, in the form of a Copyright Act amendment, where and how to leave an indication that one does not want one's work used for training, the meaning of pressing the post button remains undetermined. The moment the amendment heads to the National Assembly is the next fork in this issue.

Reporter Kwak Dong-hyun · Breath.Tech

Related articles

댓글