A Study of Fair Use in Training AI Corpora

Zheyu Pei

Law School, Zhejiang Gongshang University, Hangzhou, Zhejiang, China

Keywords: AI Corpus Training, Fair Use, Copyright Law, Generative AI Program.

Abstract: In the digital age, artificial intelligence technology is developing rapidly. The emergence and extensive

application of generative artificial intelligence have drawn scholars' attention to the legal issues in the training

of artificial intelligence corpora. This paper summarizes and analyzes the qualitative problems of data mining

in AI corpus training by domestic and foreign scholars, and focuses on the legitimacy and necessity of fair

use. Through the re-search on the identification of fair use of AI corpus training in copyright law, this paper

points out that fair use system can better balance the rights and interests of creators and public interests and

promote the development of new technologies. In view of the limitations of the enumerated provisions of

China's Copyright Law, the study suggests that AI training should be explicitly included in the scope of fair

use, and that the EU should be used as a reference to establish special exception clauses, as well as to establish

a mechanism for compensating authors in commercial use scenarios, so as to achieve a dynamic balance

between promoting techno-logical innovation and protecting the interests of creators.

1 INTRODUCTION

In the current digital environment, artificial

intelligence has gradually come into the public's

view, and has caused a non-negligible impact on all

fields of human society. For example, the emergence

and application of generative AI bring transformative

potential to the field of education, reforming

educational tools as well as promoting the

development of critical thinking (Grant, 2023). The

emergence of generative AI has also boosted the

healthcare industry with the ability to gather large

amounts of information to facilitate the refinement of

treatment protocols (Moulaei, et al, 2024).

Similarly, the impact of AI makes a difference to

the field of intellectual property, as a new technology

worthy of research, its emergence brings brand new

opportunities for the development of intellectual

property rights globally, as well as new dilemmas for

the protection of intellectual property rights. In terms

of copyright, the use of generative AI can be focused

on two processes, the data input phase and the

generator output phase. Regarding the data input

stage, the AI program will absorb a large number of

copyrighted works for training. The legality of using

copyrighted materials for artificial intelligence model

training and whether such practices constitute

regulations and standardized guidelines. Academic

circles hold divergent perspectives on this issue: some

scholars contend that utilizing copyrighted works

without explicit authorization for model training does

not violate intellectual property rights, while others

maintain opposing viewpoints. This study, therefore,

focuses on the research of fair use in AI corpus

training, synthesizing existing academic perspectives.

Building upon this foundation, the paper proposes a

reformed framework for reasonable application,

aiming to facilitate the synchronous development

between China's fair use system and artificial

intelligence technologies.

2 QUALITATIVE DOCTRINE

FOR DATA MINING DURING AI

TRAINING

The process of collecting and processing large

amounts of training data and searching and analyzing

the information hidden in it through algorithms

during artificial intelligence training is known as data

mining. Scholars in academia hold divergent

positions regarding the legal characterization of data

mining. A segment of academics asserts the

justifiability of employing copyrighted works for AI

Pei, Z.

A Study of Fair Use in Training AI Corpora.

DOI: 10.5220/0014381400004859

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Politics, Law, and Social Science (ICPLSS 2025), pages 381-386

ISBN: 978-989-758-785-6

381

training under specific circumstances, while opposing

scholars categorize such utilization as prima facie

Scholars who support the infringement theory

argue that the training of large models of artificial

intelligence belongs to commercial use, and is not

within the fair use circumstances categorized by the

current copyright law. Therefore, it should be

recognized as a copyright infringement (Yi, 2024).

Such scholars adhere to the statutory principle, but

will appear to lack practicality in the light of evolving

AI technology.

On the other hand, the scholars who advocate the

justification of data mining have further posited

distinct doctrinal frameworks, developing novel

theoretical constructs within the parameters of

technological neutrality principles. Part of them have

put forward the theory of non-expressive use, which

refers to the use of an original work for the purpose

of treating it as factual information and making

functional use of it, rather than utilizing the original

expression in the original work or reflecting the

artistic value of the original work in the new work

(Jiao, 2022). Scholars holding this view argue that AI

training constitutes the analysis and synthesis of

patterns from large-scale data, rather than utilization

of the original work itself, thus being a non-

expressive use that does not amount to copyright

infringement; Some scholars adhere to the theory of

transient reproduction, claiming that copying under

dissemination, and only when the duplicated material

is utilized for distribution does it infringe the right of

reproduction (Tu, 2024). The purpose of the copying

behavior in AI corpus training is for model training,

and is only used to generalize and analyze the logical

patterns behind the content, which is a temporary use

of the data, and will not be disseminated to the outside

world, so this temporary copying behavior does not

constitute infringement under copyright law; Besides,

other scholars advocate the fair use doctrine,

maintaining that AI training adheres to the value

principle of technological neutrality. Based on the

public benefit and fundamental nature of AI models,

they assert the substantive utilization of copyrighted

works during machine learning possesses legitimate

purpose and does not adversely affect normal

exploitation of the original works. Thus, this kind of

usage is permissible to be categorized as fair use (Yi,

2024).

3 FAIR USE IN ARTIFICIAL

INTELLIGENCE CORPUS

TRAINING

The legal characterization of data mining in AI corpus

training remains contentious. This paper posits that

the fair use doctrine proves more congruent with

China's judicial reality and global trends, effectively

facilitating technological advancement while

stimulating socioeconomic development. In contrast,

the temporary copying doctrine focuses on the

transient and technical nature of the copying

behavior, such as temporary copying in caching or

browsing, and determines that the use of another's

work in AI training does not constitute infringement

in terms of the duration and technical characteristics

of the copying. The non-expressive use doctrine, on

the other hand, emphasizes the use of a work that does

not involve expressive content, such as data analytics

or text mining, and focuses on the fact that the

purpose of the use is not to convey information about

the work itself. The fair use doctrine upholds the

value principle of technological neutrality. Through

an interest-balancing framework, it comprehensively

evaluates factors including the purpose, nature, and

quantity of copyrighted work utilization during

training processes, along with market impact. By

determining the conversion ratio of public interest in

the usage behavior, it concludes such practices fall

outside the scope of infringing acts. This more

flexible approach to adjudication can accommodate

the rapidly evolving information technology and the

commercial interests behind it. Thus, through the

above comparative analysis, the author is more

inclined to analyze the use of others' works in AI

corpus training from the perspective of fair use.

3.1 Rationality

The fair use system is fundamentally about limiting

individual copyrights and protecting the public

interest, and its underlying logic is a balance of

interests, preventing the boundless expansion of

authors' copyrights while facilitating technological

development as a means of expanding the public

interest in society. In the absence of a fair use system,

authors' copyright over their works is too broad,

which is not only unfavorable for management, but

also unfavorable for technological advances in the

public domain, as authors' blocking of their works can

negatively affect the efficiency of overall

development. Chinese scholar Xu Xiaoben's view

that AI technology belongs to the universal

ICPLSS 2025 - International Conference on Politics, Law, and Social Science

382

technology that can benefit human society and

promote scientific and technological progress, and

therefore should be included in the fair use system

according to the pursuit of the public interest. And he

believes that the use of works in the training of AI

will not affect the normal use of the original works,

and that as long as it is complemented with other

mechanisms of copyright protection, it can also take

account of the original author's and the interests of the

legislation (Xu, 2024).

The fair use system was initially dominated by the

four-factor determination rule. The 19th-century case

Folsom v. Marsh is considered to be the first case on

fair use in the U.S. In this case, Judge Joseph Story

first proposed a four-factor analytical framework that

would become the foundational standard for

assessing fair use (Justia, 1841). This analytical

framework focuses on four key considerations: the

purpose and character of the use; the nature of the

original copyrighted work; the amount and

substantiality of the portion used; and the effect of the

use on the potential market or value of the work.

Then in the 20th century Campbell case, judges

introduced the "transformative use" criterion,

elevating the fair use doctrine from its traditional

four-factor assessment framework into a new phase

of development. (Justia,1994) Wang Qian, a scholar

in China, believes that "transformative use" refers to

the use of such works by adding new content and

ideas to the original work, so as to make the original

work show a new value. It is not a simple copy of the

content of the original work, but an approach of

realizing the conversion of the work's function or

purpose (Wang, 2021). Also, Chinese scholar Wu

Handong points out that within the criteria for

determining fair use, the rule of "transformative use"

emphasizes whether the new work’s utilization of the

original material demonstrates transformative

purpose and character, rather than rigidly adhering to

restrictions on the nature or quantity of the original

work’s usage (Wu, 2020). The perspectives of the

aforementioned two parties focus more on the

transform of subjective intentions when using the

original work, creating new value based on the factual

content of the original work, which is in perfect

alignment with the essence of artificial intelligence's

corpus training. Artificial intelligence corpus training

is not a direct use of the original work, but through

data analysis to summarize the patterns behind the

original work, so as to form a logical generation mode

and generate new content for the user. That’s the

reason why training method can constitute the so-

called "transformative use" provisions.

Furthermore, Chinese scholar Yi Jiming has

proposed a special version of "transformative use"

criterion. Within this framework, he emphasizes that

such usage not only entails a subjective shift in

purpose but also involves objective technological

innovation, thereby achieving a complete

transformation and value-added enhancement of the

original work's significance (Yi, 2024). This

evaluative approach can help trainers and users pay

more attention to whether the utilization of a work

contributes to the advancement of science and

technology as well as the realization of societal public

interests, thereby fostering the creation and value

enhancement of new technologies. According to that,

Artificial intelligence, a new technology which can

promote scientific and technological progress,

absolutely has rationality for its training process.

What’s more, the author contends that traditional

"transformative use" and the new version of it do not

need to form successive phases but should coexist to

accommodate the widespread implementation of

artificial intelligence. For instance, the criteria for

traditional one could govern the training of most

generative AI systems in producing images, texts, and

similar outputs, while the new standard might apply

to the development of AI models centered on

technological breakthroughs.

Of course, there are different views in the

academic community about whether AI corpus

training can constitute fair use. Chinese scholar Wang

Xuelei acknowledges that the utilization of works

during artificial intelligence training constitutes an

exploration and analysis of the patterns underlying

the works. However, she argues that the object of

such use is the data extracted from original works

rather than the works themselves, thereby eliminating

the necessity for copyright regulation in this context

(Wang, 2025). This study, however, contends that the

data utilized for analysis originates from copyrighted

works rather than existing in isolation. Such works

should not be categorically excluded during

analytical processes. Furthermore, from the

perspective of balancing interests, reducing works to

mere data vessels through disaggregation undermines

the rights and benefits of human authors. Meanwhile,

it must be emphasized that the fair use system

inherently functions as a product of balancing

competing interests. We should keep a balance

between the interest of individual copyright and

public interest of the whole society. Some scholars,

motivated by concerns for copyright holders'

interests, argue that categorizing works created by

human intellectual labor as training materials for

artificial intelligence under fair use provisions would

A Study of Fair Use in Training AI Corpora

383

lead to the systematic disregard of human creators'

rights and contributions (Zou etal, 2024). This paper

won’t deny the restriction of individual copyright

could be brought from the fair use system.

Nevertheless, excessive emphasis on protecting

individual authors' interests could impede

technological progress and generate adverse effects

on the advancement of public interests. Facing this

problem, people should actively seek or establish a

new system of benefit distribution that can provide

protection for human authors, rather than just

preventing the realization of the fair use in AI corpus

training.

Above all, through the analysis of diverse

scholarly viewpoints of fair use and some different

cases, this research concludes that data mining within

artificial intelligence training can be demonstrated as

a form of fair use with both possibility and rationality.

3.2 Necessity

From the points of modern technology, the training of

AI relies on a large scale of data, arranging from text,

picture, video in diverse field. Therefore, if the usage

of every single piece of data requires separating

authorization, the cost will be apparently increased

and the development of technology iteration will be

hindered.

From the points of commercial interest,

enterprises will face substantial compliance costs and

legal risks if too many authorizations are in need. In

contrast, the system of fair use can significantly

reduce upfront investment, encourage more

companies to invest in the AI track, and promote

overall innovation in the industry.

In addition, the application of fair use in AI corpus

training is not a regional issue, but a global one. With

ChatGPT, DeepSeek, and many other generative AI

programs being put to use around the world, a large

number of human authors' works are being

dismantled and learned by the training of AI. In the

face of this common problem, it has become an

international trend to categorize AI corpus training as

fair use. Of course, different countries do not adhere

to a unified framework of fair use determinations.

According to their different cultural backgrounds and

histories of jurisdictions, the appropriate legal system

has been established in accordance with the specific

practice of each country. For example, the United

States has taken a liberal stance based on

"transformative use", and case law in the United

States has brought text data storage and mining within

the scope of fair use, based on the doctrine of

"transformative use" (Wu, 2020). What’s more,

articles 3 and 4 of the Copyright in the Digital Single

Market Directive promulgated by the European

Commission establish exceptions for "text and data

mining for scientific research purposes" and

"exceptions or limitations for text and data mining"

respectively. These provisions constitute mandatory

exceptions, requiring all member states to implement

TDM exception rules within their domestic legal

frameworks as minimum standards (Bao, et al, 2025).

Both the U.S. and EU regulations have already

established the reasonableness of data mining by law,

and this change with the times further signals the

development of AI technology and the urgent need to

regulate the legal issues behind it.

Based on the above analysis, it can be concluded

that it is reasonable and necessary to classify the use

of others' works in AI corpus training as fair use. Of

course, from a human creator's perspective, having

his or her work used for AI corpus training puts him

or her at risk. The use of a human author's work by a

machine is out of the control of the original author

and does not generate revenue for the author, thus

creating an imbalance of interest between the AI and

the human author (Wu, 2020). To be specific, because

of the nature of AI training, a large amount of work is

put into it, and the original authors of the work are

most likely unaware that their work is being used to

train the AI, let alone make a profit from it. And the

process of AI corpus training is massive, fragmented,

and cut up, so in case of infringement disputes, the

original authors will face difficulties in pursuing their

responsibilities. Therefore, under the present

environment of the digital age and technological

development, we should adhere to the basic principle

of technological neutrality as well as the pursuit of

commercial interests and technological advancement

while taking into account the interests of human

authors, so as to reform the original system and find

a fulcrum for balancing interests. That’s why this

research believes that a balanced benefit-sharing

mechanism for creators should be established while

applying AI corpus training to reasonable use, taking

into account commercial interests and technological

development.

4 DEFICIENCIES IN THE LAW

AND RECOMMENDATIONS

4.1 Deficiencies

Although applying the fair use system to AI corpus

training is conducive to protecting the public interest

ICPLSS 2025 - International Conference on Politics, Law, and Social Science

384

and balancing conflicting interests, the current legal

provisions on fair use in China are rather thin.

Currently, there are only the 12 circumstances listed

in Article 24 of the Copyright Law and the "other

circumstances stipulated by laws and administrative

regulations", as well as the three-step standard of

"limited to certain special circumstances" introduced

by the Regulations for the Implementation of the

the determination of fair use is still based on specific

enumeration, which means that judge does not have

the right to create new exceptions of fair use. In that

case, judge does not have the right to directly expand

the interpretation of fair use to the AI corpus training

data mining as well.

In addition, AI corpus training cannot be

interpreted by the existing twelve circumstances in

the law. The use of works in AI corpus training is

usually large in scale and quantity, which is not in line

with the requirements of "appropriate citation".

Besides, the training of AI basically belongs to the

behavior of enterprises, and most of the training is

carried out for commercial purposes, which is not in

line with the purposes of "personal learning", "news

reporting", "education and research" or "national

public service".

The existing law does not specify whether the

emerging AI corpus training can be categorized as

fair use, which cannot meet the rapid development of

the existing AI technology and the status quo of the

massive use of AI technology. For the sake of the

relevant interests and protections of the creators, the

fair use system in China should be reformed.

4.2 Reinventing Fair Use System in the

Age of Artificial Intelligence

Regarding the difficulty of using the existing fair use

system for data mining under the current legal

environment, there are different views in academic

field. For example, Chinese scholar Huang Xijiang

suggests that even if there is no special provision for

fair use of "text and data mining", the fair use in AI

training can still be regulated by the general provision

the contrary, Chinese scholar Lin Xiuqin proposes

that we should learn from the U.S. model of "fair use

plus specific enumeration", which stipulates the

factors for fair use based on four-factor determination

rule, and then enumerates the common ways of fair

use (Lin, 2021). What’s more, Chinese scholar Wu

Handong contends that "text data mining" should be

added to the list as a new case of fair use (Wu, 2020).

Whether it is to explicitly list AI corpus training

as a special case of fair use or to recognize it through

judicial interpretation as a case stipulated in the

underpinning provision of Article 24 of the Copyright

Law, the relevant legal system should be in line with

the rapid development of the trend of AI technology

to make certain changes. The author believes that, in

contrast, the addition of AI corpus training to become

a specific case of fair use can be more intuitive and

quick to regulate the relevant legal issues, but it

should be differentiated between the use of AI corpus

training. For an instance, if it is used for scientific

research and non-commercial purposes, such as

teaching, then it will be directly applicable to the fair

use. And if it is used for commercial purposes, it

should be established with the author of the original

work to pay the appropriate compensation.

5 CONCLUSION

To summarize, the rapid progress of modern science

and technology has prompted the human society to

face the potential legal problems of AI technology.

The society have to commit that the contradiction

between the wildly use of generative AI for other

people's works and the limited provisions of the

existing law needs to be solved urgently. This paper

believes that the fair use theory can achieve a smooth

and fair state between the copyright of human

authors, the commercial interests of AI training

enterprises and the interests of the overall

development of society. Therefore, in order to

promote the development of the technology and

improve the efficiency of the judiciary in the face of

this problem, the legislator should clarify the specific

circumstances of fair use in the legal revision, and

incorporate the AI corpus training into the scope of

the fair use, and establish a unified standard of this

issue. At the same time, establishing a mechanism for

the original author of the work with the payment of

appropriate compensation is a good way to balance

the commercial interests of the AI training enterprises

with the human author's copyright, so that people can

promote the upward development of society.

REFERENCES

of generative artificial intelligence training data: An

analysis of the EU copyright exception rules and their

implications for China. Library Forum.

A Study of Fair Use in Training AI Corpora

385

Cooper, G. 2023. Examining science education in

ChatGPT: An exploratory study of generative artificial

intelligence. J Sci Educ Technol 32: 444-452.

adherence and change in the age of intelligence by

ChatGPT. Intellectual Property 8: 124-125.

utilization in the creation of artificial intelligence and

the path to resolving them. Contemporary Law 4: 130.

Justia U.S. Supreme Court. 1994. Campbell v. Acuff-Rose

Music, Inc. 510 U.S. 569. 2025.4.1.

https://supreme.justia.com/cases/federal/us/510/569/

Justia U.S. Supreme Court. 1841. Folsom v. Marsh. 9

F.Cas. 342 (C.C.D. Mass. 1841). 2025.4.1.

https://law.justia.com/cases/federal/district-

courts/massachusetts/madce/9fcas342/4104271/220/no

-4.html

in the age of artificial intelligence. Legal Studies 6(6):

182-183.

Moulaei, K., Yadegari, A., Baharestani, M., et al. 2024.

Generative artificial intelligence in healthcare: A

scoping review on benefits, challenges and

applications. International Journal of Medical

Informatics: 105474.

machine learning: Beyond the theory of non-expressive

use. Politics and Law 10: 166-168.

Wang, Q. 2021. A course in intellectual property law (7th

ed.). People's University of China Press: 330.

fair use system for artificial intelligence data mining.

Hebei Law Journal 3: 190-191.

generated by artificial intelligence. Zhongwai

Jurisprudence 3(3): 659-661.

model training under the perspective of technology

neutrality. Law Review 4: 94.

Yi, J. 2024. Research on the reasonable use of large model

corpus training. China Copyright 6: 7-8, 19.

Zou, H., Bi, M., Pu, J., Zhao, L., & Yan, L. 2024. Exploring

the compliance of artificial intelligence training data.

Modern Commerce and Trade Industry 19: 33.

ICPLSS 2025 - International Conference on Politics, Law, and Social Science

386