The next legal battle for generative AI may be: What is used, can it be made clear?
On August 25, the Japanese government officially adopted the "Generative AI Intellectual Property Protection and Transparency Principle (Principle Code)" related to the protection of generative AI intellectual property rights, requiring AI industry players to improve the transparency of training data, including disclosing information such as the summary, type and collection method of data used for model training. It is worth noting that this set of regulations is not currently a legally enforceable hard law, nor does it require companies to disclose complete databases. Instead, it improves the accountability of AI developers through soft law governance. Its applicable thinking is not limited to Japanese local companies, but will also take into account overseas companies that provide related AI services in Japan. Japan has not suddenly closed the door to AI development. AI can continue to learn, but can the training data always remain a black box that no one can see?
The reason why this policy is noteworthy is that Japan’s copyright law has long adopted a relatively relaxed attitude towards data analysis and machine learning.
Article 30-4 of Japan's Copyright Act allows, under certain conditions, the use of works for information analysis and other purposes "other than enjoyment of the thoughts or feelings expressed in the works", and has therefore long been regarded as a legal environment friendly to AI training. But this provision has never been an immunity gold medal for "AI can use all works at will". The law itself still has boundaries that should not unduly harm the interests of copyright owners. However, with the rapid development of generative AI, a bigger problem has gradually emerged: rights holders don’t even know whether their works have been used, so how can they judge whether their rights have been affected? When the law only discusses "whether the use is legal" but does not have enough information for the right holder to know "what exactly was used", it may be difficult to exercise substantive rights due to information asymmetry.
Therefore, the real importance of Japan’s policy this time is to advance AI copyright management from a simple “permission or prohibition” to a “transparency obligation.”
This is an important change in legal thinking. In the past, disputes between AI professionals and creators have often been simplified into two extremes: one side believes that training AI requires a large amount of authorization, otherwise it is infringement; the other side claims that machine learning requires a huge amount of data. If every piece of data must be licensed in advance, it will be difficult for the AI industry to develop. But there is actually a third system option between the two: the law can retain reasonable space for data analysis, while requiring large AI companies to bear a certain degree of transparency responsibility for data sources. In other words, "trainable" and "no need to explain at all" are not the same thing.
The value of a transparent system is not just about protecting copyright owners.
For the AI industry, clearly stating the source and method of obtaining data may reduce future legal risks. When AI models enter the enterprise, government, medical, education and professional service markets, users are increasingly concerned about whether the models involve unauthorized content, personal information, business secrets or other high-risk information. If developers are completely unable to explain the basic source of training materials, it will be difficult for downstream companies to conduct legal compliance and risk assessments. Therefore, training data transparency may even transform from "legal burden" to "market competitiveness" in the future. Whoever can prove that their data management is more complete and the source of authorization is clearer will be more likely to gain the trust of large corporate and government customers. Japan chose to promote it through Principle Code instead of immediately imposing comprehensive mandatory regulations, which also reflects its policy thinking: let the market form transparent standards first, and then observe whether further legislation is needed.
Japan’s choice is worthy of Taiwan’s reference.
Taiwan also hopes to develop independent AI and also has news, publishing, film and television, music and a large number of Chinese content industries. If future policy discussions still remain on the dichotomy of "whether AI training is a reasonable use", it will easily lead to a long-term confrontation between the technology industry and the content industry.
A more pragmatic direction is to first establish minimum data management rules, such as requiring basic model developers above a certain scale to disclose the main types of training data, source methods, whether authorized data are used, and provide rights holders with reasonable channels for inquiries or complaints. This does not mean that companies are required to disclose every piece of information, nor should they be forced to disclose model architecture or business secrets, but rather establish the most basic information symmetry between innovation and rights protection. A truly mature legal system in the era of generative AI can first ask developers to answer "How does your AI learn?" The message Japan is sending is clear: AI can be allowed to continue learning, but training data cannot be left in a black box forever.
