Ruyu Yan

Faculty of Law, Macau University of Science and Technology, Macao, China

Keywords: Generative Artificial Intelligence (GenAI), Copyright Infringement, Data Acquisition, Fair Use.

Abstract: The rapid development of generative artificial intelligence (GenAI) increases the risk of copyright

infringement during data acquisition and use. This study examines infringement risks at GenAI's input stage,

focusing on the legal conflicts in data collection, processing, and output. It highlights substantial violations

of economic rights, such as reproduction and adaptation. Under China's Copyright Law, statutory licensing is

inapplicable due to non-compliant subject qualifications and behavioral discrepancies. Fair use defenses fail

because of commercial intent and excessive scope. Tests and analyses, including the three-step test, four-

factor analysis, and transformative use doctrine, consistently show non-exemption. To address liability

asymmetries, algorithmic opacity requires a fault presumption mechanism with a reversed burden of proof.

To counter enforcement deficiencies, the study proposes novel remedies like dynamic compensation models

and algorithmic injunctions. It concludes with institutional recommendations: enforcing enhanced robots.txt

compliance, creating open-licensed data repositories, and developing international compliance frameworks to

balance technological innovation with copyright protection.

1 INTRODUCTION

The accelerated evolution of generative artificial

intelligence (GenAI) exerts profound societal

impacts. While catalyzing transformative innovation

in the Fourth Industrial Revolution, its unregulated

deployment within incomplete legal frameworks has

triggered pervasive infringement litigation.

In 2022, software engineers instituted

proceedings against GitHub for unauthorized code

exploitation (DOE, 2022). March 2023 witnessed

artists filing claims against Stability AI for scraping

copyrighted images to train models and generate

derivatives (Andersen et al, 2023). Q4 2023 saw The

New York Times litigate against OpenAI and

Microsoft for training on millions of unlicensed

articles, seeking data expunction and statutory

damages (Li, 2023). Concurrently, generative art

features on NetEase's LOFTER and Xiaohongshu

platforms precipitated mass user attrition due to

unauthorized training practices. These developments

critically erode creator incentives and stifle

innovation ecosystems.

Scholarly consensus regarding copyright

infringement in GenAI input data acquisition remains

elusive. While predominant academic opinion

endorses fair use exemptions to foster AI

advancement, a significant minority advocates robust

creator rights protection to ensure innovation quality

(Xu & Yang, 2019 … Jiao, 2022). This paper

synthesizes these divergences to examine

infringement liabilities, fair use controversies, and

regulatory paradigms for GenAI systems.

2 ANALYSIS OF COPYRIGHT

INFRINGEMENT PATHWAYS

2.1 Data Acquisition as Functional

Exploitation

GenAI's purported "creativity" derives from

computational architectures of large language models

(LLMs) and corpus ingestion during training. LLM

construction and Transformer algorithm optimization

require massive datasets for pre-training/refinement.

Critically, training data quality dictates GenAI output

fidelity. Google's text models ingested >1.5 trillion

tokens during training, while ChatGPT-3 (released

June 2020) utilized multi-terabyte pre-training

corpora (Ye, 2025).

As acquired data materially constitutes generative

capability through model training, such acquisition,

Yan, R.

DOI: 10.5220/0014390300004859

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 577-582

ISBN: 978-989-758-785-6

577

though superficially informational, constitutes

functional exploitation. This is demonstrated by its

direct utilization in: (i) Model parameterization

processes;(ii) Enhancement of content-generation

efficacy.

2.2 Does Data Utilization Constitute

Infringement?

Article 52 of the Copyright Law of the People's

Republic of China enumerates eleven infringement

liabilities. Provisions I, II, VI, VII, and VIII expressly

stipulate that unauthorized use constitutes

infringement. Regarding GenAI training data

acquisition, scholars contend that securing mass-scale

licensing from numerous rights holders is

prohibitively costly and impractical for service

providers (Jiao, 2022).

Unauthorized data acquisition violates Article

VII, which prohibits "using copyrighted works

without remuneration." Such use constitutes

actionable infringement requiring civil liability. Nota

bene: While this analysis addresses input-stage

infringement, the concealed nature of such acts

creates evidential barriers: Infringement processes

lack traceability; Direct evidence is largely

inaccessible; Determination must rely on output-

stage "access + substantial similarity" tests.

2.3 Rights Infringed in Data

Acquisition

GenAI entails phase-specific copyright infringement

risks throughout its data processing lifecycle: During

input phase, unauthorized reproduction and storage of

works in training media may directly violate

reproduction rights; In processing phase,

deconstruction, reorganization, or adaptation of

original works through translation, annotation, or

compilation for model optimization may infringe

derivative rights, including translation, compilation,

and adaptation rights; when it comes to output phase,

dissemination of generated content bearing

substantial similarity to source works in expression or

core creative elements may trigger communication to

the public rights infringements.

Collectively, these full-process chain activities—

from data collection and processing to content

generation—create direct conflicts with copyright

law. Core legal controversies center on: Whether

unauthorized reproduction/derivation qualifies as fair

use; The applicable standard for determining legally

cognizable similarity between outputs and source

works.

2.4 Analysis of Infringement Liability

Exemptions

2.4.1 Statutory Licensing

China's Copyright Law establishes five statutory

licensing regimes. Besides Article 25 (textbook

compilation), the other four are Article 23 (periodical

reprinting), Article 35 (phonogram production),

Article 46 (broadcasting of published works), Article

50 (digital reproduction by public institutions).

A systematic review confirms none of the

licensing regimes mentioned above can be applied to

GenAI data acquisition.

In the first place, there is a discrepancy regarding

the eligibility of the subject entities. Statutory

licenses are strictly limited to specific entities, such

as textbook compilation organizations, periodical

publishing units, and producers of sound recordings.

However, the entities involved in artificial

intelligence (AI) research and development are

predominantly commercial companies, which do not

meet the qualifications of the legally stipulated

entities.

Secondly, the criteria for the required actions are

not met. Each statutory license mandates specific

modes of use, such as textbook compilation and

production of sound recordings. In contrast, the

method of obtaining data for AI involves complex

technological processes such as data scraping,

storage, and analysis via information networks. This

method significantly differs from the specific modes

of use stipulated in statutory licenses.

Finally, there is an insufficiency in procedural

requirements. Except for the clauses related to

textbooks, other statutory licenses retain the rights for

explicit permission from copyright holders and

without an effective mechanism for such declarations,

the acquisition of data for AI does not meet the

procedural requirements of statutory licensing (Zeng,

2019).

2.4.2 Fair Use

• Statutory Analysis

Article 24 of China's Copyright Law adopts an

exhaustive list with open-ended clause structure.

None of its twelve specific exceptions encompass

GenAI data acquisition. Scholar Jiao Heping

conducts a detailed analysis of several contested

points that are relatively relevant (Jiao, 2022).

He elaborates the reasons why AI-generated

creations do not comply with the first clause on

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

578

"individual learning and research": Firstly, from the

perspective of the subject, the "individual" in

"individual learning and research" typically refers to

a natural person. However, in the context of AI-

generated creations, the entity using the data is the AI

system rather than a natural person. Secondly,

concerning the purpose requirement, individual use

must be based on the non-commercial purposes of

"learning and research." Currently, AI-generated

creations are predominantly controlled and executed

by large commercial internet companies, which are

unlikely to meet the non-commercial purpose

requirement of personal use. Thus, this requirement is

not satisfied.

The reasons why AI-generated creations do not

comply with the second clause on "proper citation"

are also explained: Firstly, they do not meet the

purpose requirement, as the use of data in AI-

generated creations aims to produce new works rather

than to "introduce or comment on a specific work" or

"clarify a particular issue." Secondly, the use of data

works in AI-generated creations clearly exceeds the

requirements of propriety. Therefore, this

requirement is not fulfilled.

The attention is also drawn to why AI-generated

creations do not comply with the sixth clause on

"scientific research": Firstly, the type of fair use for

scientific research specified by copyright law

involves restrictions on copyright aiming at public

interest. Therefore, under this provision, research

institutions and activities should "only apply to state-

established educational and research public

institutions," which does not include commercial

internet companies. Secondly, this type of fair use

imposes limitations on the number of copies made,

while AI-generated creations often involve the full-

text replication of works, failing to meet the "limited

quantity" requirement. Thirdly, the entities using AI

data are not "restricted to use by researchers only." As

such, this requirement is not met.

Consequently, statutory interpretation alone

precludes fair use exemptions for generative AI,

establishing prima facie infringement liability.

However, Article 8 of the Supreme People's Court's

2011 Opinions on Promoting Socialist Cultural

Development through Intellectual Property

Adjudication introduced a hybrid standard expanding

fair use boundaries, stipulating that courts may

recognize fair use in exceptional circumstances

necessitated by technological innovation or

commercial development after evaluating: (i) purpose

and character of use; (ii) nature of copyrighted work;

(iii) substantiality of portion used; and (iv) market

impact—provided such use neither conflicts with

normal exploitation nor unreasonably prejudices

rightsholders' legitimate interests. This establishes a

multifactor framework requiring demonstrated

necessity, exceptional circumstances, satisfaction of

the four-factor analysis, and compliance with the

three-step test. For comprehensive rigor, this analysis

incorporates the Berne Convention's three-step test

(Art. 9(2)), U.S. Copyright Act's four-factor standard

(§107), and the contemporary transformative use

doctrine.

• Combining the " three-step test "

When applying the three-step test—which

assesses fair use through sequential criteria: (1)

limitation to certain special cases; (2) non-conflict

with normal exploitation; and (3) non-prejudice to

legitimate interests—GenAI data acquisition fails all

requirements (Xiong, 2018). First, it satisfies no

statutorily recognized "special case." Second, market

substitution effect analysis confirms GenAI's output

capabilities displace human creators in multiple

domains, violating criteria (2) and (3) by

unreasonably prejudicing economic interests. Finally,

unauthorized data ingestion lacks normative

legitimacy under fair use doctrine. Scholar CHANG

Ye consequently contends that unlicensed GenAI

training remains non-exempt under China's copyright

framework (Ye, 2025). While the author concurs with

this assessment, the three-step test's inherent

ambiguity necessitates judicial refinement before full

adoption in Chinese jurisprudence.

• Combining the "Four-Factor Test"

The "Four-Factor Test" originates from Section

107 of the United States Copyright Act, which

enumerates some typical scenarios of fair use and sets

flexible and open general provisions. Judges can

comprehensively consider the following four factors

to determine whether it constitutes fair use: ① the

purpose and character of the use; ② the nature of the

copyrighted work; ③ the amount and substantiality

of the portion used in relation to the copyrighted work

as a whole; ④ the effect of the use upon the potential

market for or value of the copyrighted work (Shen,

2020). It mainly explains "fair quotation". However,

the data acquisition of generative artificial

intelligence is for creation, and the research and

development companies are for commercial

purposes. Moreover, the quotations are basically full

copies, and the learning and imitation of the creator's

unique style will undoubtedly affect the market

position of the original author and intensify market

competition. Therefore, the author believes that the

"Four-Factor Test" cannot be used to defend fair use.

• Combining the "transformative use"

579

The "transformative use" standard was first

proposed by Judge Leval in the United States in the

judgment criteria for fair use of copyright and has

been permitted for application in American judicial

practice, with its connotation becoming increasingly

clear. In the application of the "transformative use"

standard, the key lies in determining whether the new

work adds new content with different characteristics,

using new expressions, meanings or information. The

more transformative the new work is, the more likely

it is to constitute fair use (Gu & Fang, 2023).

Subsequently, this standard was introduced into

China's judicial practice and academic discussions,

providing an important reference for the fair use

system and effectively compensating for the closed

nature of the Copyright Law. For instance, in cases

heard by courts in Beijing, Shanghai, Guangdong and

other places, the content of "transformative use" has

appeared in over 30 judgments (Han, 2023).

Although "transformative use" has been

frequently cited in judicial decisions, it is still limited

to specific contexts such as data generation for

educational and training purposes. Even when courts

refer to this standard, they often impose additional

thresholds such as "commercial purpose" and "market

impact". However, the author holds a contrary view

and does not recommend using "transformative use"

as a defense element for fair use. Essentially,

"transformative use" still involves deep learning,

imitation, and utilization of the original work. The

works output by generative AI after processing

through the "algorithm black box" do indeed present

"new expressions" due to the absorption of a large

amount of work data and the integration of various

styles, but in essence, they are still pieced together

from original works.

Netizens jokingly refer to AI as a "sewing

monster". If such pieced-together creations can be

exempted from legal liability, it would be akin to

feeding AI on the "bones" of original authors, which

not only chills and terrifies them but also greatly

dampens their creative enthusiasm, making it difficult

to foster a healthy and positive creative environment.

In conclusion, whether based on legal provisions

or various analytical methods, the analysis points to

one result: the infringement of obtaining input data

for generative AI cannot be exempted and should bear

the liability for infringement.

3 DILEMMAS OF ATTRIBUTION

AND PUNISHMENT

3.1 Dilemma of Attribution

The principle of attribution is the standard and

principle for determining the civil liability that

different types of tortious acts should bear, which

decides the elements of liability for a certain tortious

act, the burden of proof, the conditions for exemption,

the principles and methods of damages compensation,

etc (Wang, 2010).

However, there are significant difficulties in its

practical application. Even if it can be determined that

generative artificial intelligence has used

unauthorized works in the "feeding" of data for large

model training, there are still new problems brought

by new technologies in confirming the responsible

party. The output content of generative artificial

intelligence, on the one hand, relies on the training of

massive data, and on the other hand, is based on

interaction with users.

Therefore, the subjects of data "feeding" may

involve both users and generative artificial

intelligence, and the possible infringing party is

naturally not unique. The cause of this attribution

dilemma also lies in the complexity of the

explainability of algorithms from the input end to the

output end of generative artificial intelligence.

3.2 Dilemma of Punishment

Making the infringer bear responsibility is an

important means of punishing infringement.

According to Article 52 of China's Copyright Law, if

the data "feeding" of generative AI models does

indeed involve infringement, the infringer may be

required to assume responsibilities such as ceasing

the infringement, eliminating the influence, making

an apology, and compensating for losses (Li, 2003).

However, the issue of determining the amount of

compensation for losses has always been a difficult

problem in the field of intellectual property research.

What is even more challenging is that responsibilities

such as ceasing the infringement and eliminating the

influence may become unenforceable or difficult to

verify when applied to generative AI, a special object.

Currently, the training of generative AI in practice

is generally unidirectional and progressive, and

cannot be reversed. Many studies have shown that the

"contribution" of previous infringing "feeding"

training may continue to have an impact on the

subsequent content generation of generative AI, and

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

580

the infringement of the copyright holder's rights by

generative AI may persist (Wang, 2003).

Therefore, for generative AI, ceasing the

infringement and eliminating the influence may

essentially become unenforceable or difficult to

verify. This dilemma still has a lot of room for

research.

4 FRAMEWORK

CONSTRUCTION FOR

REASONABLE REGULATION

4.1 Establishing a Technical

Authorization Mechanism

Professor Chang Ye proposed regulating the use of

generative AI data by implementing the Robot

Exclusion Protocol and introducing mandatory

"machine forgetting" norms (Ye, 2025). The author

fully agrees. As a general rule for web crawlers, the

Robot Exclusion Protocol has been incorporated into

China's "Self-discipline Convention for Internet

Search Engine Services," but its current legal effect is

limited. It is suggested to "strengthen its binding force

through administrative regulations or legislation" to

enhance its actual enforcement. With a certain legal

foundation, specific measures can be more easily

implemented.

First, establish a technical rule for authorizing the

"feeding" of online works, setting whether it is

allowed to be used for generative AI training as a

necessary rule.

Second, require network service providers to

transfer the rule-setting authority to the users who

upload the works, and set the default to prohibit the

use of user works by generative AI, and prohibit

obtaining authorization through user agreements.

This move can substantially ensure that the works

uploaded by users are not crawled by AI, and better

protect the copyright rights of users.

In addition, I suggest that to improve the quality

of data obtained by AI and encourage users to grant

open licenses, a reward mechanism for users who

grant open licenses can be further improved. For

example, users who upload a certain number and

quality of works and have them adopted by AI can be

awarded certificates or monetary rewards. This will

further promote the creation of high-quality works

and the development of AI on the basis of protecting

the intellectual property rights of original authors.

4.2 Establishing a Prosecutorial

Management Mechanism

Professor Gao Yang proposed the establishment of an

open licensing mechanism for copyrighted works and

a dynamic review mechanism for infringing content

(Gao, 2024). The author strongly agrees. In the open

licensing mechanism, after the copyright

management department reviews the application of

the licensor, it publicly announces the information of

the data collection of copyrighted works and the

licensing conditions.

When potential licensees fulfill their obligations,

an open license is formed. This mechanism innovates

the traditional one-on-one negotiation model between

adopts a new form where copyright holders

voluntarily license to the public, set licensing fees and

payment methods, and licensees can obtain the data

collection upon accepting the conditions.

This move not only benefits AI enterprises in

obtaining training data, breaking down data barriers,

and helping small and medium-sized AI enterprises

access high-quality data, but also enhances the

utilization efficiency of the data collection of

copyrighted works, promoting mutual benefit and

win-win situations between the copyright industry

and AI enterprises.

5 CONCLUSION

This study has not exhaustively addressed

compensation mechanisms for input-stage

infringement risks in GenAI. A comprehensive

analysis of damages quantification standards and

liability forms requires further empirical

investigation into dispute resolution practices. GenAI

regulation constitutes a global regulatory challenge.

As GenAI evolves into a productivity tool,

transborder data flows become inevitable.

Consequently, governance of training data

acquisition necessitates international cooperation and

regulatory harmonization. A critical imperative

remains: establishing harmonized regulatory

frameworks to mitigate copyright infringement risks

during GenAI training data acquisition through

collective international efforts.

581

REFERENCES

Andersen et al. v. 2023. Stability AI et al. 3:23-cv-00201,

N.D. Cal.

behavior in the “Feeding” of Generative AI Data.

JOSAL, 2.

DOE 1 v. 2022. GitHub, Inc. 4:22-cv-06823, N.D. Cal.

acquisition and utilization in AI creation. CPJ, 4.

H. Li. 2003. Reversal of the burden of proof: theoretical

analysis and problem research." SILAB, 4, 87-94.

L. Wang. 2010. Characteristics of the accountability

principles system in China's tort liability law. Legal

Forum, 25, 7-10.

L. Wang. 2023. On some Issues of reversal of the burden of

proof." GDSS, 1, 150-158.

N.F. Gu, and Z. Fang. Reasonable boundaries and

Infringement Regulations for the use of works by

generative AI like ChatGPT." DLF, 7.

Q. Xiong. 2018. Clarification of judicial standards for fair

use of copyright. OJLS, 1, 144-157 .

S. Liu. 2024. Exploration of legal pathways for fair use of

in AI Creation. HBLJ, 37 .

W. Han. 2023. Research on risk avoidance pathways for

construction: from the perspective of introducing

transformative use rules in judicial trials of copyright

cases." LST, 3, 64-69

W. Shen. 2020. Study on transformative use from the

perspective of the public domain in copyright law."

SALJ, 5 29-30.

AI deep learning." SJTU Law Review, 3.

X. Zhao. 2024. Issues of fair use in generative AI and

machine learning." JNJ, 3.

training data. CPJ, 15.

Y. Li. 2023. The urgency of AI governance. CBJ, 3.

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

582