Most artificial intelligence developers need to establish internal knowledge databases (training corpora) containing massive works to train artificial intelligence, but this massive, systematic, and large-scale commercial use has not yet been included in the scope of rights restrictions stipulated by any international treaty or national copyright law. Article 7 of the Interim Measures for the Administration of Generative Artificial Intelligence Services jointly issued by the Cyberspace Administration of China and seven other departments stipulates that when conducting training, providers of generative artificial intelligence services shall not infringe upon the intellectual property rights enjoyed by others in accordance with the law, if it involves intellectual property rights. However, the regulation does not specify under what circumstances infringement can be recognized. How to qualitatively and standardly train artificial intelligence using works has sparked the most intense debate worldwide. The main reason for this is that the development of artificial intelligence technology is too rapid, and existing copyright laws and institutional design theories are not yet prepared.

The existing viewpoint often cites theories of non expressive or transformative use originating from the West as a basis for the legitimacy of artificial intelligence training. However, even Western cases indicate that large-scale internal replication that has a substantial impact on the subscription market does not constitute fair use. Judge Leval, the proposer of the concept of transformative use, believes that the defendant company's library's copying of journal articles is neither transformative nor non-commercial, but rather substitutive. The defendant company only mechanically copied the relevant articles. The copy of the original document is not intended to be used as a part of a larger entity for certain new purposes. The main purpose of the defendant company's copying is to replace the original and allow for copying, even in bulk. Researchers can make copies of an article for future reading or reference, eliminating the need to share the original journal with others. This type of copying behavior does not make any substantial additions or modifications to the original work, it only achieves an increase in the number of copies. The so-called 'transformative use' requires the creation of new content that is different from the original, rather than simply making copies for a beneficial purpose. The second instance also supports this idea, as the court found that the defendant company cannot obtain a reasonable use exemption just because it copied the archives of an article (or book). Similarly, a newspaper cannot claim that, as its business is "news reporting," it can self copy news books and place them on journalists' bookshelves for them to learn from, nor can a school claim that in order to help teachers improve their teaching abilities, it can make a large number of copies of educational books and have teachers read the copies. No matter how much benefit copying and reading such books may bring to improving teaching ability, it cannot constitute reasonable use for teaching purposes.


This article argues that artificial intelligence training is deeply embedded in the context of China's industrial upgrading, technological competition, and institutional adjustment, and therefore has a strong policy orientation. As pointed out in the "Outline for Building a Strong Intellectual Property Country (2021-2035)", intellectual property protection should "highlight key areas and major needs, and promote the deep integration and development of intellectual property with various aspects such as economy, technology, culture, and society". Against this backdrop, the "non expressive use," "transformative use," and "opt out" currently being discussed in the West are difficult to provide a permanent solution to the copyright issues arising from using works to train artificial intelligence. At different stages and periods of the application of artificial intelligence technology, there should be policies and institutional designs that are suitable for the social and economic conditions and national conditions at that time. The next more feasible direction is not to forcefully apply the problem of artificial intelligence training to these previously emerged concepts and mechanisms, and thus universally recognize that using works to train artificial intelligence is a reasonable use. On the basis of fully grasping China's actual national conditions, it is necessary to clarify the crux of the problem of artificial intelligence training, promote the improvement of relevant mechanisms, and safeguard the rights and interests of rights holders. In this process, some temporary institutional designs with transitional nature may be needed, and more explanatory Chinese discourse may be extracted from them to achieve a spiral rise in the legal protection of artificial intelligence training.



V. Conclusion
While maintaining the boundaries of copyright holders' rights, necessary and clear legal guarantees should also be provided for the development of new technologies. What is truly needed in the era of generative artificial intelligence is to provide clearer and more forward-looking rules to delineate predictable action boundaries for all parties involved. Only in this way can technological development not come at the cost of excessive erosion of rights, and rights protection can also obtain a sustained institutional foundation in response to changes in reality.

*The above article is from the 4th issue of China Rule of Law in 2026

Source: China Rule of Law Online