A Study of Copyright Issues in the Input Phase of Generative

Artificial Intelligence Machine Learning

Weixi Song

School of Artificial Intelligence and Law, Southwest University of Political Science and Law, Chongqing, China

Keywords: Generative Artificial Intelligence, Input Phase of Machine Learning, Copyright.

Abstract: In recent years, the rapid progress of generative AI technology has had a profound impact and transformative

impact on multiple dimensions such as technological innovation, cultural evolution, and education model

transformation. Among the many complex issues raised by generative AI, the input phase of machine learning

is highly dependent on the input of massive amounts of data, which often cover the scope of content protected

by intellectual property rights, which has led to copyright disputes. This reality has directly led to the deep

contradictions and fierce conflicts between the traditional copyright industry and the emerging artificial

intelligence industry. Although the traditional solutions proposed by scholars can adjust the contradiction

between the two to a certain extent, there is still room for improvement. This paper explores a new mechanism

that conforms to the principle of balance of interests to solve the problem of infringement in the input stage

of machine learning.

1 INTRODUCTION

Machine learning is the core of AI and the

fundamental way to become intelligent. Generative

AI requires deep learning and big data analysis to

achieve the purpose of generating text, images, audio

and other types of content, and ultimately make it

have the scientific definition of "emergent

capabilities"(Wu and Lai,2024). Machine learning

can be subdivided into three phases, namely the input

phase, the learning phase and the output phase

(Xu,2022）, of which the input phase is an important

part of machine learning and the focus of this paper.

In the input stage, the generative AI needs to

capture and integrate a large amount of training data,

which includes both information protected by

law, aiming to provide rich material for the

subsequent self-learning process. Relying on the

training dataset constituted by these carefully

collected data, the generative AI carries out in-depth

model training so that the content it generates or

expresses can highly approximate the samples in the

training set in terms of effect, thus achieving the

desired level of intelligent output (Lai, 2024).

However, the data crawled in the input phase of

machine learning inevitably conflicts with the

it crawls to copyright-protected content.

If the problem of infringement at the input stage

cannot be solved, then machine learning will remain

under the cloud of infringement risk for a long time.

2 COPYRIGHT DISPUTES AND

INSTITUTIONAL DILEMMAS

AT THE INPUT STAGE

At the input stage, generative AI infringes on

crawling of data, which is often copyright-protected

content. In order to effectively train AI algorithms,

developers often obtain the large amount of data they

need through the technical protection measures of

deciphering databases(Liu and Wei,2019) and copy

and store the data obtained through such means into

their own servers or not, and directly feed the data to

the AI. This may create the problem of infringement

at the input stage of generative AI. In China, the

"Altman case" was the first legal decision on the issue

of generative AI infringement, and the final decision

resulted in a judgment declaring that the use of

generative AIs to provide generative services without

authorization was found to be infringing, and in

March 2024, Google Inc. was hit by a lawsuit for the

180

Song, W.

A Study of Copyright Issues in the Input Phase of Generative Artiﬁcial Intelligence Machine Learning.

DOI: 10.5220/0013975800004912

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

In Proceedings of the 1st International Conference on Innovative Education and Social Development (IESD 2025), pages 180-186

ISBN: 978-989-758-779-5

unauthorized use of copyrighted content to train its

chatbots. bots, for which Google Inc. was hit with a

financial penalty of up to €250 million by the French

market regulator.

The rapid development of the artificial

intelligence industry has caused a great impact on the

balance of interests system constructed by traditional

intelligence, the problem of infringement has become

more secretive, thus making the evidence mechanism

fail, and its massive use has likewise caused a great

impact on the rights licensing system (Yuan and

Xia,2024). However, at present, if artificial

intelligence input every work to obtain prior rights

license, almost equal to kill this new industry.

Generative artificial intelligence as an important field

of development, if it is caught in a large number of

infringement disputes, it is bound to cause a "chilling

effect", which is not conducive to the development

and progress of the new technology. How to protect

the interests of copyright holders while promoting the

development of generative artificial intelligence has

become an urgent problem to be solved.

From the existing explorations in the academic

community, some scholars have provided ideas the

problem of infringement .for solving generative

artificial intelligence the concept of temporary

reproduction by explaining "Temporary copying" is

an incidental and objective technical phenomenon,

which, in the field of artificial intelligence, manifests

itself as temporary storage in a temporary cache, and

does not produce a copy that can be recognized as an

act of copying, but is merely an addition to the

subsequent steps. In order to avoid expanding the

connotation of the right of reproduction, scholars and

judicial precedents have agreed that temporary

reproduction should not be recognized as the content

of the right of reproduction, which gives generative

artificial intelligence trainers a certain degree of

avoidance of infringement risk, but in practice, the

use of temporary reproduction means to avoid the risk

of copyright infringement is obviously not practical.

As a result, academics have proposed a fair use

system, hoping to rationalize the infringement

problem at the input stage by including machine

learning of generative AI as one of the twelve

elements stipulated in China's copyright law.

Similarly, some scholars have put forward other

solutions, arguing that the intellectual property rights

of copyright holders should be protected, and that

copyrights such as the right to machine learning and

the right to read should be added. At present, the

latter argument has received a certain degree of

criticism from the academic community, which

argues that human beings do not possess intellectual

property rights of the so-called "right to read", and

that analogous to the so-called right of machines to

read and the so-called right of machines to learn

would be even more important when it comes to

human reading behavior. The so-called "right to read"

and "right to learn" of machines have even less legal

basis. As a result, the fair use system has received

widespread attention from academics as an important

means to address the current infringement of

generative AI input.

3 OBSTACLES TO EXISTING

FAIR USE REGIMES TO

ADDRESS INFRINGEMENT AT

THE INPUT STAGE OF

GENERATIVE AI

3.1 Fair Use Regime is Feasible to

Address Generative AI

Infringements

Based on in the previous article, the input stage of

machine learning, due to the massive amount of

grabbing data caused by copyright disputes, want to

solve this problem, its key point is still focused on,

can the machine legally grab these data for their own

machine learning?

According to the provisions of China's copyright

law, many scholars have focused on the "fair use

system", which is undoubtedly the most suitable

solution for the development of generative artificial

intelligence, and its inclusion in the fair use system

can directly and effectively solve the problem of

infringement of copy right in the input stage, and can

produce full benefits for the development and

expansion of the artificial intelligence industry. The

development and expansion of the industry can

produce full benefits. However, the fair use system is

essentially a system that, for the purpose of

reasonable use and according to the principle of

proportionality, can produce a certain degree of

derogation to the rights of copyright owners. It is

strictly regulated and controlled by copyright law.

The construction of the criteria for the application

of the fair use rule in China's current copyright law

follows the "three-step test" established by the Berne

Convention.This standard requires that fair use must

meet the following three conditions: firstly, it must

meet specific and exceptional circumstances;

A Study of Copyright Issues in the Input Phase of Generative Artiﬁcial Intelligence Machine Learning

181

secondly, it must not impede the normal utilization of

the original work; and thirdly, it must not reasonably

impair the legitimate rights and interests of the

China's copyright law clearly stipulates twelve

situations that can be recognized as fair use.

The input phase of machine learning is currently

not included in the twelve types of fair use stipulated

by China, but the newly added paragraph 13, "Other

Laws and Regulations", provides for the possibility

that the input phase of machine learning can be

included in the fair use system. As a result, the fair

use system has some room for operation in theory. In

addition, in the wave of artificial intelligence, as

countries pursuing development together with us, the

EU and Japan have loosened the restrictions on

companies more space and freedom to develop, so

there is a certain possibility and reasonableness for

China to make regulations on the fair use of

generative AI by analogy.

3.2 Difficulties Encountered vs the Fair

Use System

Although the fair use system described above

provides us with ideas for solving the infringement

problem in the input phase of machine learning, the

fair use system still has some limitations and

problems that should not be ignored.

First, the use of the fair use defense in the input

phase of generative AI may have pushed the

boundaries of the fair use system's interest protection.

The fair use system itself is a restriction on the

exempt from certain infringement of copyright. It is a

sacrifice made to the rights of copyright holders for

the development of science and technology,

therefore, the law needs to strictly limit the content of

the provisions of the fair use system to ensure that the

rights and interests of copyright holders are not

infringed by unreasonable use(Wang and Chu,2024).

It is highly likely that the production of generative

artificial intelligence will act as a substitute for the

the revenue that the copyright owner obtains through

the work, which will break the balance of interests of

the fair use system original and will not be conducive

to encouraging creativity.

Second, generative AI may not pass the second

step of the fair use test. The key to integrating the

input stage into the fair use regime centers on the

successful realization of the last two steps of the

three-step test. With regard to the definition of the

"two shall nots", according to the official

interpretation of the Adjudicatory Committee of the

World Trade Organization, the core logic lies in the

methodological framework of economic analysis to

define them. Specifically, the first "shall not" has a

strict meaning, which requires that the original work

not be exploited in a way that conflicts with the

economic benefits through the exercise of legal

rights.

However, combined with the actual background

of the current artificial intelligence we can come to

this very unfortunate conclusion, that is, almost most

of the artificial intelligence on the market can not pass

the second step of the three-step test method.

Generative artificial intelligence as a highly

sophisticated industry, its nature requires a lot of

research and development and capital investment, so

it is almost difficult to see individual research and

development of generative artificial intelligence on

the market, often with technology companies as the

main body. And as an operating company, it will

inevitably generate competition for the original

copyrighted work in the market competition,

squeezing the space for its survival. According to the

theory of socially necessary labor time determining

the value of goods put forward by Max, artificial

intelligence will reduce the original work of socially

necessary labor time and thus reduce the value of its

goods the economic value of the work and the

interests of the work will be seriously impacted, and

thus, therefore, most of the generative artificial

intelligence is unable to pass this step. This also

creates an obvious dilemma for the fair use system,

how to advance the fair use system without violating

the existing legal basis system?

Again, generative AI also fails the third step of the

fair use test. In contrast, the second "shall not" is

relatively loosely defined. On the basis of not

violating the principle of balance of interests, it

allows the utilization of the original work to cause

moderate derogation to the economic interests of the

reasonable range. This is both a limitation on the

rights of the copyright holder and a license for the

rights of the user(Wu and Lai,2024).

However, it is also unfortunate that it does not

stipulate what constitutes "moderate derogation",

which leaves us with a problem of definition.

Although the concept of "proportionate derogation" is

IESD 2025 - International Conference on Innovative Education and Social Development

182

difficult to understand, we can explore this issue

through the principle of balancing of interests. The

principle of balancing interests permeates the entire

content of intellectual property law and has a role to

play in intellectual property rights that cannot be

ignored. Its external manifestation is the

measurement of interests. According to China's first

ALCG case, it can be seen that China is still more

inclined to protect the interests of copyright owners

to achieve the effect of balancing interests.

Currently, generative artificial intelligence has

been widely used in a variety of fields, and its future

prospects for large-scale application are also very

clear. This kind of technology has already grasped the

opportunity and initiative of the new era, and its

advantageous position in the dissemination of

information compared with individual copyright

holders should not be ignored. Generative artificial

intelligence not only has the ability to generate

content efficiently, but also relies on the wide

coverage of the Internet to realize the rapid diffusion

of information and access to it, thus substantially

changing the traditional pattern of information

dissemination and posing unprecedented challenges

to the rights and interests of copyright owners. The

position, and the application of the fair use system at

this time almost completely breaks the antagonism

between the interests of the copyright owner and the

generative artificial intelligence, and completely

reverses to the generative artificial intelligence.

Furthermore, the focus of the fair use system itself

is "fair use", and fair use itself requires the user to

meet the condition of "non-profit". The popular text-

generating artificial intelligence chatgpt, openAI

company launched several different price point

products and services, obviously with the fair use of

"non-profit" fundamental hedge.

According to existing judicial cases in China, the

number of courts that have ruled that the input phase

of machine learning is fair use is still in the

minority(Cong,2024). Judicial practice also confirms

that the application of the fair use system still has

certain difficulties in China.

To summarize, I believe that at the present time,

the fair use system does not have the reasonableness

of directly applying to the input stage of machine

learning, nor does it have the legal origin and custom,

and the blind expansion of the fair use system will

only run counter to the fundamental spirit of China's

establishment of the copyright law, and violates the

provisions of the basic principle of the principle of

balance of interests.

4 A SOLUTION TO THE

INFRINGEMENT PROBLEM

OF GENERATIVE ARTIFICIAL

INTELLIGENCE

Generative artificial intelligence of machine learning

input stage of the infringement problem of the

solution is not only can only purely apply the fair use

system, generative artificial intelligence infringement

problem of the solution needs to take the necessary

classification to do refinement, and fair use of the

system to carry out some transformation.

4.1 Adaptation of The Existing Fair

Use Regime - Inclusion of

Non-Commercial Purpose Use in

The Fair Use Regime

4.1.1 Do All Input Stages Violate the Two

"Shall Nots"?

Generative AI, as an emerging and highly intelligent

technological tool, has an extremely wide range of

applications, covering a variety of fields such as

personal learning, data retrieval, and assisted decision

making. It is worth noting that not all types of

generative AI cannot be evaluated by the three-step

test. In order to explore this issue more deeply, some

scholars in the academic community have already

proposed a division between two types of AI for

commercial and non-commercial purposes(Wu and

Lai,2024). The proposal of this categorization has

greatly assisted us in addressing more precisely the

infringement issues that may be involved in the input

stage of generative AI.

In the context of non-commercial use, the actual

harm to copyright holders from generative AI can be

minimal. It is mainly used in scientific research,

public governance and other fields that are important

for the well-being of society. Since non-commercial

use does not have the characteristic of wide

dissemination, its degree of infringement on the

interests of copyright owners is relatively low, and it

does not cause market substitution effect on the

original work, thus avoiding serious squeezing and

damage to the interests due to copyright owners.

A Study of Copyright Issues in the Input Phase of Generative Artiﬁcial Intelligence Machine Learning

183

In addition, non-commercial use can be applied by

analogy to item (a) of the twelve circumstances

stipulated in China's copyright law, which allows

individuals to use other people's works for the

purpose of study, research or enjoyment. Although in

this context, the subject of study using generative AI

may not be limited to individuals, but may also

include corporate entities such as companies, this

does not prevent the reasonableness of this mode of

use in terms of legislation or even in terms of

expanding the interpretation of the content of the first

subparagraph thereof.

From this I believe that it is reasonable and

necessary to include machine learning of generative

AI for non-commercial purposes in the category of

fair use. Once it is included, then the problem of

infringement at the input stage that we have been

exploring will be solved.

4.1.2 Benefits of including Machine

Learning for Non-Commercial

Purposes in The Fair Use Regime

According to discussions among scholars, artificial

intelligence is now divided into three stages, namely,

weak artificial intelligence, stronger artificial

intelligence and strong artificial intelligence.

Scholars widely agree that the current artificial

intelligence is still in the stage of strong artificial

intelligence, i.e., artificial intelligence is still only

used as a tool at this stage, and does not have the so-

called "human" thinking characteristics (Cong,2019).

This is because generative AI is a tool revolution

brought about by the evolution of human beings to

today's technological level. As a tool to assist

humans, generative AI must rely on machine learning

in order to escape the algorithmic bias that it may

induce(Lai, 2024).

Non-commercial purposes of generative artificial

intelligence application field is mainly scientific

research and public policy and other public welfare

science category, will be included in the category of

fair use, one is in line with the development needs of

the current society secondly, can be better application

of new technology to promote the well-being of

mankind, thirdly, is conducive to the generative

artificial intelligence reasonably legitimate get rid of

algorithmic bias.

According to the theory of business failure, if non-

commercial generative AI cannot be included in the

fair use system, it may lead to the loss of incentive for

its developers and users to conduct scientific and

public research without getting the benefits they

want.

Therefore, there should not be too much

restriction in generative AI for non-commercial

purposes, which in itself is a hindrance to

development and contrary to the concept of copyright

law making.

4.2 Reasonableness of Applying a

Statutory Licensing System for

Commercial Purposes

4.2.1 Use for Commercial Purposes Fails the

Three-Step Test

Generative AI for commercial purposes is itself

monetized by the content it outputs. For its

copyrightability issue this paper does not study this,

but it is undeniable that the share of generative AI has

been increasing rapidly since its introduction. The

profits of the traditional copyright industry, however,

have been hit by the rapid growth of the AI industry.

This suggests that our concerns are not unfounded

worries and that the AI industry is tangibly affecting

the functioning of the traditional copyright industry.

In the contemporary era when generative AI has

become an industry, users sign agreements with and

even pay certain fees to developers of generative AI,

and if such commercial behavior is regarded as fair

use, it will undoubtedly pose a serious challenge to

the copyright system and seriously threaten the

legitimate rights and interests of copyright holders.

Such behavior not only blurs the line between fair use

and commercial exploitation, but also may weaken

the core function of the copyright system to motivate

and protect creators, which in turn will have a

negative impact on the long-term development of

cultural innovation and knowledge dissemination（

Gao,2024）.

4.2.2 Reasonableness of The Statutory

Licensing System

Statutory licensing has been proposed as a measure to

resolve this dilemma by derogating from the rights of

the copyright but balancing the interests by means of

a fee paid by the user to the copyright. Such a solution

is more eclectic than the fair use system in that it

protects the middle ground between the two. It allows

the copyright holder to receive compensation for the

derogation of his rights, while for trainers, the

payment of compensation allows them to use the data

without prior permission from the copyright holder

and greatly facilitates their use of the data. This

IESD 2025 - International Conference on Innovative Education and Social Development

184

system is more in line with the principle of

harmonization of rights and obligations and the

balance of interests of intellectual property rights, and

more suitable for the development of China's national

conditions.

At present, China's copyright is still in a period of

development, and the public's awareness of

intellectual property protection is still weak. I have

always believed that only when the humanities lead

the development of science and technology can the

field of science and technology be stabilized and

brought to the forefront. The statutory licensing

system can increase the public's awareness of

intellectual property rights, thus promoting more

people to know the law, abide by the law and use the

law.

4.2.3 How Times Set the Amount of

Compensation

Traditional intellectual property disputes tend to

determine the amount of damages through two

aspects, one is the loss of the copyright owner, and

the other is the unjust enrichment of the defendant. In

the case of generative artificial intelligence, the

problem becomes more complex, because generative

artificial intelligence is often a massive amount of

data, involving a very wide range of copyright

owners, if the copyright owner loss as a measure, it

will undoubtedly increase the judicial pressure, and

very unrealistic. Therefore, I should advocate to take

the defendant's unjust enrichment to determine the

amount, should be in accordance with a certain

percentage of the benefits received by the generative

artificial intelligence company as the copyright

holder's compensation, this amount should not be set

too high, otherwise, according to the same market

failure theory, the development of the generative

artificial intelligence industry will be severely

challenged(Zhang,2024).

Similarly, since the data crawled is massive, it is

unlikely that AI companies will spend much time on

finding individual copyright holders, and since a

system of compensation is to be implemented, it is

necessary to set up a specialized agency to manage

copyrights. With reference to the American

Federation of Musicians

In keeping with the practice of the Hop and the

radio and record company AFM, they will entrust the

Foundation with the management of the money, free

concerts and jobs for more musicologists.

Although this last step is difficult to implement

concretely in China nowadays, the specialized

agencies can adopt internalized rules to rationally

allocate the use of compensation.

5 CONCLUSION

The rapid development of generative artificial

intelligence, for us to bring a new round of

technological revolution at the same time, but also

brought new copyright risks and challenges, in the

face of this new highly intelligent tools, we found that

the original principle of balance of interests as the

core of the construction of the copyright law has not

been able to adapt to the needs of the social

development of the present, generative artificial

intelligence training process is not only secretive but

also has a huge amount of sex, making the copyright

The training process of generative artificial

intelligence is not only secretive but also massive,

which makes the copyright suffer unprecedented

serious challenges.

Although some scholars strongly advocate that

the input stage of machine learning be included in the

fair use system through re-legislation or the logic of

expressive and non-expressive use, so that the input

stage of the machine can be rationalized and justified

from now on. However, the commercial use of

generative artificial intelligence significantly violates

the legal principle of the unity of rights and

obligations, and moreover violates the principle of

balance of interests in copyright law.

From this we divide the application of generative

AI into two segments, commercial use and non-

commercial use, non-commercial use incorporates

the fair use system is more reasonable and in line with

the principle of proportionality, while commercial use

through the combination of the statutory license

system and the compensation system of the system of

the two systems, so that it balances the interests of the

two sides, the rights of the copyright is impaired but

through the payment of compensation to make up for

the way of their Damage suffered.

REFERENCES

Cong Yike.2024. Limitations and Exceptions of Data Min-

ing in the Age of Artificial Intelligence. Technology En-

trepreneurship Monthly 2024,37(08):39-45.

ution of Artificial Intelligence Generated Con-

tent.China Publishing 2019,(01):11-14.

A Study of Copyright Issues in the Input Phase of Generative Artiﬁcial Intelligence Machine Learning

185

artificial intelligence training data. China Publishing

2024(15):12-18.

sponse Strategies in the Input Stage of Machine Learn-

ing. Technology and Law 2024,(10):161-165.

ment problem of machine learning and its solution.

Journal of East China University of Political Science

and Law ,2019,22(02):68-79.

erative Artificial Intelligence Machine Learning and Its

Institutional Response. Editor's Friend 2024,(11):96-104.

grams for machine learning. Southeast Academic 2022,

(02):237-245.

Yuan Zhenfu and Xia Zixuan.2024.A study on the copy-

right compensation system for work utilization in ma-

chine learning.science and technology and law.

2024,(07):28-36

Wang Qian,Chu Chu.2024.A preliminary study on the

boundary between artificial intelligence and copyright:

legal challenges and reflections under technological

progress. China Edit 2024,(08):56-62.

Zhang Ping.2024.The Institutional Difficulties of Copy-

right Legitimacy of Artificial Intelligence Generated

Content and Its Solution Path.Legal Science (Journal of

Northwest University of Politics and Law)

2024,42(03):18-31

IESD 2025 - International Conference on Innovative Education and Social Development

186