AI 훈련 사전 보상 주목…그러나 기준은 여전히 ​​공백 - IT조선

As the National Artificial Intelligence Strategy Committee collects opinions from various sectors on Korea’s AI Action Plan through January 4, the government is known to be pushing a “pre-use compensation” scheme for copyrighted works used in AI training. This marks a step beyond the previous debate over whether permission is required to use copyrighted data for AI model training. The problem is that it is likely to take a long time to establish concrete compensation standards. In fact, the pay-TV industry has failed to reach an agreement on program usage fee calculation standards for more than a decade and is still mired in conflict. According to industry sources, the government is reviewing a plan to collect a portion of AI companies’ revenues into a fund and distribute it to the creative sector. The idea is to first allow AI to use copyrighted works for training and then return a share of the resulting profits. The structure is similar to existing schemes such as the Film Development Fund and the Broadcasting Development Fund, which return a certain percentage of revenue to the creative ecosystem. The background to discussions on pre-use compensation is the widening AI model gap. Compared to global big tech companies such as OpenAI and Google, as well as Chinese AI firms, Korea’s AI industry has been relatively cautious about data training. This is largely due to concerns over copyright disputes. There is also the practical issue that even if copyrighted works are used without permission for AI training, it is difficult for rights holders to detect or prove such use. Even if policy direction is finalized in favor of pre-use compensation, challenges remain. It is difficult to predict how long it will take to establish standards for calculating compensation for the use of copyrighted works. The pay-TV industry is a representative example. The industry has failed to establish standards for calculating content (program) usage fees for more than 10 years. So far, the only concrete case is last year, when IPTV operators submitted their own calculation standards to the Ministry of Science and ICT and applied them. Program Providers (PPs) strongly oppose this, arguing that IPTV operators are unilaterally determining compensation. Although 18 years have passed since the launch of IPTV in 2008, the conflict remains unresolved. Discussions on copyright compensation for AI are likely to take even longer. This is because the types and scope of data used for AI training are extremely diverse. The recent increase in omnimodal AI models—those that learn multiple types of data at once, such as text and audio, rather than combining separate modules—also adds complexity. Experts agree that if an AI model learns combined data consisting of text, images, and audio in a single process, standards must be established for how to calculate the proportion of each type of copyrighted work. This is necessary in order to distribute revenue accordingly. There are also few overseas policy precedents that Korea can reference. Major countries have yet to clearly define AI training on copyrighted works as an exempt activity. In the United States alone, 66 AI copyright infringement lawsuits were pending as of December last year. The European Union and Japan are among the few jurisdictions that allow AI training on copyrighted works. The EU has adopted an opt-out system that allows AI to use copyrighted works for training unless creators explicitly indicate, “Do not use my work for AI training.” Japan permits the use of copyrighted works as long as the purpose is not for human enjoyment. However, creative communities in both the EU and Japan are pushing back against these legal frameworks. In both cases, only the training stage is institutionally permitted. The Korean government has announced a grace period of more than one year ahead of the enforcement of the AI Basic Act on January 22. This is intended to ease industry concerns about regulatory burdens and to avoid becoming the world’s first country to implement full-scale AI regulation. Lee Jin-soo, Director at the Ministry of Science and ICT, said at an AI Basic Act briefing session on the 24th last year, “We are also keeping open the possibility of extending the grace period in consideration of overseas trends, including in the EU, and the pace of technological development.” This is interpreted as an intention to wait and see, given that no country has yet established clear standards for compensation calculation. An industry expert said, “AI training on data and the generation of outputs using that trained data should be viewed separately. If an image-generation AI produces a specific image, it is extremely difficult to prove which creator’s work or artistic style it was based on.” He added, “If AI outputs infringe copyright, legal disputes will inevitably increase. To reduce legal risks for the industry, it would be better to first establish standards for determining copyright infring