Copyright Protection Issues in Artificial Intelligence Data Training

2026-10-12

Author: Cheng Yingwen
Hu Kaizhong, Professor at Kaiyuan Law School, Shanghai Jiao Tong University

picture
On July 25, 2026, the "AI Era New Unfair Competition and Judicial Governance Seminar" jointly organized by Kaiyuan Law School of Shanghai Jiao Tong University and Intellectual Property Finance was successfully held in Shanghai. This conference gathered academic, judicial, and industry representatives to engage in in-depth dialogue on cutting-edge issues such as the identification of new forms of unfair competition in the era of artificial intelligence, data rights protection, platform responsibility, and intellectual property rights of AI generated products. It contributed diverse wisdom to building a competitive legal environment that adapts to the development of the AI industry. At the meeting, Professor Hu Kaizhong from Kaiyuan Law School of Shanghai Jiao Tong University gave a keynote speech on "Copyright Protection Issues in Artificial Intelligence Data Training", and Intellectual Property Finance organized his lecture content for readers.


After listening to the sharing of all the guests, I was deeply inspired. The discussion on related issues is quite cutting-edge, many of which involve intellectual property protection of artificial intelligence products. Next, I would like to share with you the issue of copyright protection in artificial intelligence data training.

The reason for discussing this issue is because the artificial intelligence industry is developing rapidly. In the process of industrial development, people pay more attention to the issue of model rights and less attention to the phenomenon of artificial intelligence enterprises using a large number of human works in data training. In this context, many copyright owners have found that some of the content generated by artificial intelligence models is highly similar to their works, leading to a lot of controversy. The understanding of related disputes between the industry and academia is not unified, and the disputes among copyright owners are also prominent. Therefore, studying this issue has practical significance.

From the copyright cases in the field of artificial intelligence in recent years, similar disputes are not uncommon. In the United States and Europe, there have been multiple cases of copyright owners suing artificial intelligence companies, claiming that these companies have used their works for data training without authorization, which has harmed the interests of copyright owners. In our country, there have also been related cases such as Xiaohongshu. There has also been a similar case in Shanghai, where the LoRA model used Medusa's artwork during data training, which led to a lawsuit filed by the copyright owner.

//
1、 Stakeholders involved in artificial intelligence data training
//
To solve the above problems, it is first necessary to clarify which stakeholders are involved in the artificial intelligence data training stage. Data training must be based on data, which involves the use of copyrighted works. After the work is integrated into the dataset, it is used for artificial intelligence training and provided to users through an artificial intelligence platform, which involves the interests of multiple parties.

In practice, taking the aforementioned Xiaohongshu case as an example, the artist's works were used for training and generated similar images. The author has also conducted experiments using artificial intelligence models, requiring the model to generate images such as Donald Duck, and the results generated images that are more similar to the original work. For example, in the Ultraman case, the Ultraman image was used for data training to generate images that were similar to the Ultraman image. In these cases, many copyright owners believe that artificial intelligence companies using their works for training, which involves copying, adapting, or even translating the original works, and using their works without permission, should bear infringement liability.

There are also many similar cases abroad, such as generating music similar to the original work after training. Therefore, data training has sparked significant controversy: whether using works for data training is a legal or illegal act? Copyright owners usually consider this behavior to constitute illegal use, as their works have been used without permission, and therefore hope that artificial intelligence companies will pay for it; Artificial intelligence companies often argue that this behavior constitutes fair use, citing that the work is only used for the data training process and is not publicly disclosed, but rather trained in a non-public state. Some scholars have also proposed that data training and content output should be evaluated separately, believing that during the data training stage, netizens cannot see the content generated by the model, so this behavior is considered reasonable use.

//
2、 Legitimacy assessment of using works for artificial intelligence data training behavior
//
（1） Does it constitute infringement

In fact, the above viewpoint has a cognitive bias. In intellectual property law, to determine whether infringement has occurred, it is usually necessary to examine four constituent elements: subjective fault, objective infringement behavior, damage consequences, and causal relationship. At the same time, it is necessary to determine whether there are any exemption situations, such as reasonable use, statutory permission, etc.

（2） Does it constitute reasonable use

In the judgment of fair use, if the data training and result output are completely separated, it does not meet the criteria for fair use judgment. Reasonable use requires examining both the behavior of using the work and the results of use, and cannot be viewed in isolation, otherwise it is inevitable to take it out of context. In this regard, the Bern Convention provides for a three-step test, while the United States copyright law provides for a four factor judgment method, both of which are essentially consistent. For ease of understanding, the following analysis will be based on the American four factor judgment method.

Firstly, the purpose and nature of using works in data training should be examined. Generally speaking, if the purpose of using a work is non-profit, such as for educational, research, or other purposes, it may be considered fair use; If it constitutes transformative use in data training, that is, generating innovative results through changes to the original work, it may also be considered fair use. For example, if a company develops plagiarism detection software through data training, or if an artificial intelligence model can be trained to recognize the genre of a work, such as distinguishing between prose, novel, and poetry, without affecting the interests of the original work, such use is considered transformative. Special attention should be paid to whether the method of obtaining the work is legitimate and has good faith. In practice, a large number of artificial intelligence companies extract works from the network during data training without obtaining copyright authorization, so the legitimacy of the obtained works is questionable, which is also the key issue.

Secondly, the nature of the work being used should be examined. Generally speaking, based on the criteria of fair use, novels and other works with strong originality should be protected to a higher degree, while factual and historical works may allow for relatively broad use.

Thirdly, the number of works used and their importance in the original work should be examined. In practice, a large number of works are used for data training and reflected in the generated results, and even some of the generated content is completely identical or substantially similar to the original work. Such behavior may exceed the limits of reasonable use. It should also be noted that data training has a significant impact on the use of works and copyright. Many works are used for data training, and the large amount of content generated by the model is substantially similar to the original work, or the original work appears in large quantities in the generated content after being split. Compared to traditional environments, the scale of model utilization of original works far exceeds that of the past. For example, when Lu Xun's "The Scream" is used for data training, it may be split into thousands or tens of thousands of copies, appearing in different generated content. Traditional works may be used thousands or even hundreds of millions of times in the era of artificial intelligence, causing more impact on copyright owners than in any previous era; If substantial substitution occurs, the impact on the copyright owner will be more severe.

It can be seen that data training poses a significant threat to the interests of copyright owners. From the perspective of usage purpose and transformative use, some works may meet the elements of fair use and can be treated as fair use; But there are also some behaviors, especially the widely existing data training behaviors, that do not meet the requirements for reasonable use. Although many artificial intelligence companies hold different opinions on this, according to copyright reports released by the United States and the European Union, it is believed that the data training behavior of a large number of artificial intelligence companies is difficult to fully meet the requirements for reasonable use.

//
3、 The Necessity of Copyright Protection in Artificial Intelligence Data Training
//
Therefore, it is necessary to protect copyright. Artificial intelligence companies conducting data training essentially need to utilize the works of copyright owners. Respecting copyright is not only a way to show respect for creative labor, but also beneficial for the development of cultural industries, social equity, and the protection of cultural diversity. Meanwhile, the works of copyright owners are the prerequisite and source for the development of the artificial intelligence industry. Without the works of copyright owners, artificial intelligence enterprises will have no way to carry out data training. Therefore, artificial intelligence companies must fully consider the interests of copyright owners.

Of course, there are also opinions that the above position is too strict. If we examine the recent reports released by the European Union, they are more strictly judged based on the "three-step test", which means that reasonable use must first be made in specific circumstances as stipulated by law. The relevant report from the European Union believes that there are currently no regulations in EU law regarding the use of works for data training, so this behavior is certainly not considered fair use, and the stance is stricter.

//
4、 The Path of Copyright Protection in Artificial Intelligence Data Training
//
In this situation, I believe that China should make a distinction when dealing with relevant issues: if the use of the work meets the elements of reasonable use, it should be treated as reasonable use; If it does not meet the requirements, artificial intelligence companies should pay fees. Specifically, it can be expanded from the following aspects.

In the standardization of the fair use system, conversion use, teaching and research use can be included in the scope of fair use, and judged based on the "three-step test method" or the four element method. For behaviors that do not comply with reasonable use regulations, two paths can be taken: voluntary permission and statutory permission. At present, OpenAI、 Google and other companies have already paid some copyright owners in the United States and the European Union. Recently, some artificial intelligence companies in China have also begun to consider paying copyright owners. If implemented, this will be a milestone event. At the same time, it should also be noted that many copyright companies in the United States have recently filed lawsuits against Chinese artificial intelligence companies due to their failure to pay fees, so this issue is worth paying attention to.

From the perspective of payment methods, voluntary licensing may be difficult to fully implement in practice. The author suggests that in specific circumstances, legal licenses can be supplemented to allow artificial intelligence companies to use their works for training in accordance with the law, and pay remuneration to copyright owners according to legal conditions after use. This can ensure that there are works available for data training, and also enable copyright owners to receive corresponding economic compensation, achieving a win-win situation. Of course, some artificial intelligence companies may have reservations about this. To promote the development of relevant systems, it is also necessary to improve the collective management system of copyright, establish an information disclosure system and exit mechanism.

Resolving the issue of copyright protection in artificial intelligence data training is indeed a difficult problem, with a long and arduous journey ahead. However, in the long run, AI companies conducting data training need to use the works of others and should respect their copyrights. Some artificial intelligence companies also hope to obtain legal protection for the copyright of their own works. Therefore, there is mutual benefit between protecting the interests of copyright owners and protecting the interests of artificial intelligence enterprises. If this issue can be handled properly, a win-win situation can be achieved, promoting the development of both the cultural industry and the artificial intelligence industry, which is also the goal we hope to achieve.
Return
Previous：The "Four Lines of Defense" that Enterprises Must Build from the 15th China Intellectual Property Annual Conference Next：The total number of patent authorizations from January to September 2026 is approximately 2.3 million!