Research on Copyright Infringement in AI-Generated Works

Tao Wen

College of Political Science and Law, Hubei University of Arts and Sciences, Xiangyang, Hubei, 441000, China

Keywords: Generative AI, Copyright Infringement, Substantial Similarity.

Abstract: The rapid evolution of generative artificial intelligence technologies has precipitated structural contradictions

between their data-driven creative mechanisms and existing copyright frameworks, posing formidable

challenges in determining copyright infringement for AI-generated content. This investigation centers on

infringement characteristics across three critical phases—data acquisition, model training, and content

generation—with empirical analysis revealing distinct patterns including traceability challenges, exponential

proliferation of impacts, and chained distribution of liable entities. The conventional "access+substantial

similarity" adjudication framework demonstrates inherent limitations when addressing the fragmented

recombination nature of generated content, while judicial practices expose systemic deficiencies. To address

challenges, the study proposes institutional innovations that encompass dynamic ownership determination

protocols, the reconstruction of "substantial appropriation" criteria in originality assessment, and the

refinement of fair use exemption clauses for machine learning applications, aiming to achieve synergistic

integration of technological governance and legal regulation. Key findings underscore that establishing

differentiated ownership allocation systems, implementing anti-plagiarism algorithm embedding technologies

at the computational level, and precisely demarcating fair use boundaries constitute pivotal solutions to

infringement disputes. Future research must prioritize methodological breakthroughs in infringement

determination paradigms alongside advancements in intelligent detection toolkit development to navigate this

evolving legal-technological frontier.

1 INTRODUCTION

With the breakthrough development of generative AI

technology, intelligent creation tools like ChatGPT

and Stable Diffusion have deeply integrated into

traditional creative fields such as literature, art, and

music. These technologies, trained on vast amounts

of data and using deep learning algorithms, can

independently generate text, images, and audio-visual

content that appear to be original. However, the data-

driven creation mechanism of these models conflicts

structurally with the current copyright system: on one

hand, training the models requires scraping billions of

copyrighted works from the internet, potentially

infringing on the rights of reproduction and

adaptation; on the other hand, the fragmented

recombination characteristics in the generated

content pose challenges to the traditional 'access +

substantial similarity' standard for determining

infringement.

The academic community has established a

preliminary research framework on the copyright

issues of AI-generated content. Technologists focus

on algorithm transparency and infringement tracing

mechanisms, advocating for the use of blockchain

evidence storage and text fingerprint recognition to

break the 'algorithmic black box.' The legal

interpretation school aims to reconstruct the standards

for recognizing originality, proposing theories such as

'significant use' and 'dynamic ownership rules.' In

judicial practice, courts in Beijing, Shenzhen, and

other regions have attempted to establish standards

for the copyrightability of AI-generated content

through case judgments, but common issues include

lagging infringement comparison methods and

unclear standards for liability allocation. Notably, the

'machine learning exception clause' added to the EU's

AI Act in 2023 and the US Copyright Office's 'AI

Generated Content Registration Guidelines' mark the

beginning of foreign legislation addressing technical

challenges. However, these systems need to be

adapted to China's legal system to ensure

compatibility. This article focuses on the challenges

of copyright infringement in the entire lifecycle of

generative artificial intelligence, delving into the

184

Wen, T.

Research on Copyright Infringement in AI-Generated Works.

DOI: 10.5220/0014359100004859

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 184-190

ISBN: 978-989-758-785-6

legal risks associated with data collection, model

training, and content output. By analyzing the

technical mechanisms of AI creation and judicial

precedents, it highlights the systemic failures of

existing legal rules in addressing new forms of

infringement, such as fragmented recombination,

algorithmic black boxes, and distributed

responsibility chains.

The study aims to develop a framework for

identifying infringement that aligns with the

characteristics of technology, proposing institutional

solutions that include dynamic ownership allocation,

significant use determination, and the restructuring of

fair use exemptions. It seeks to address the

evidentiary challenges posed by the 'technological

invisibility' and the ineffectiveness of remedies due to

the aggregation of minor infringements, providing

theoretical support and practical pathways for the

adaptive transformation of copyright systems in the

age of artificial intelligence.

2 ORIGINS AND LEGAL

CHARACTERISTICS OF

IN AI-GENERATED WORKS

The infringement issues surrounding AI-generated

works stem from systemic misalignment between

their technical mechanics and current copyright

frameworks. Technologically, the creative process of

generative AI is divided into three distinct phases:

data ingestion, algorithmic training, and content

output, each carrying prima facie copyright

infringement implications.

2.1 Source of Infringement Issues

2.1.1 Data Input

During the data ingestion phase, generative AI

systems construct training datasets by scraping

massive volumes of internet-based texts, images, and

audiovisual materials—a process potentially

involving unlawful reproduction of copyrighted

works. For instance, Microsoft's AI poetry collection

Sunshine Lost the Glass Window was created through

algorithmic analysis of tens of thousands of modern

poems; the incorporation of unlicensed works in such

training data constitutes direct copyright

infringement. The automated and concealed nature of

this data acquisition creates a double-threat scenario:

rights holders rarely detect inclusion of their works in

training corpora, causing large-scale infringement

that often remains undetected for prolonged periods.

2.1.2 Algorithm Training

Infringement risks during the algorithmic training

phase predominantly concern derivative work rights.

Generative AI systems employ neural network

models to extract stylistic features and reconstruct

patterns from training data, a process potentially

constituting unauthorized adaptation or translation of

original works. In Authors Guild v. OpenAI,

plaintiffs alleged that ChatGPT disassembled the

Harry Potter novels into textual shards for model

training without authorization, subsequently

generating text mimicking the series' narrative

conventions, thereby implicating adaptation right

violations (American Writers Association, 2023).

The infringement risks during the algorithm training

phase primarily focus on the right of adaptation.

Generative AI uses neural network models to extract

features and reconstruct patterns from training data,

which can potentially constitute adaptation or

translation of the original work. In the case of the

American Authors Association v. OpenAI, the

plaintiffs accused ChatGPT of using the 'Harry Potter'

series without authorization, breaking it down into

text segments for model training, and generating

similar narrative styles through algorithms, thereby

infringing on the right of adaptation (American

Writers Association, 2023). The unique aspect of

such infringements is that the algorithm's use of the

work transcends the traditional' word-for-word

replication 'model, instead achieving style imitation

through semantic parsing. This shift in infringement

determination requires a focus on' creativity logic

convergence 'rather than' similarity in expression,'

significantly increasing the complexity of judicial

decisions.

2.1.3 Content Output

Infringement during the content output phase

primarily manifests as substantial similarity between

the generated work and existing works. In 2023, the

Beijing Internet Court heard the 'first AI painting

infringement case,' where the defendant used the

Stable Diffusion model to create a painting titled

'Digital Dunhuang.' This painting closely matched the

core elements of Li's meticulous and colorful work in

terms of composition and color coordination. The

court ruled that the generated work was substantially

similar and ordered the defendant to bear (Beijing

Internet Court AI Painting Infringement Case, 2023).

These cases highlight new features of AI-generated

Research on Copyright Infringement in AI-Generated Works

185

content infringement: first, the infringing content

exhibits a' fragmented recombination 'characteristic,

where the generated work may incorporate elements

from multiple works without directly copying any

single one; second, the infringing entities show a'

chain-like' trend, with developers, platform operators,

and end users all potentially liable. For instance,

developers are responsible for selecting data sources

and designing algorithms, platforms are responsible

for monitoring during the release process, and users

can also produce illegal outputs at the input stage. All

parties must be held accountable based on their

respective degrees of fault.

2.2 Feature

The core characteristics of generative AI

infringement are primarily reflected in the

'untraceability' of the infringement. The

unexplainable nature of deep learning models makes

it difficult to trace the infringement process, leaving

one unable to determine which specific works were

used as training data or to prove a causal link between

the generated content and specific works. Secondly,

the consequences of infringement are characterized

by 'large-scale diffusion.' According to OpenAI,

ChatGPT has over 180 million monthly active users,

and a single infringing model can generate millions

of infringing pieces of content in a short time, far

exceeding the scope of traditional human-generated

infringement. Lastly, the determination of

responsibility is marked by 'diversity and

complexity,' as factors such as whether developers

fulfill data filtering obligations, whether platforms

establish infringement warning mechanisms, and

whether users have malicious intent must all be

considered in liability assessment. There is an urgent

need to establish a 'technology + law' collaborative

governance framework.

3 JURISPRUDENTIAL

DILEMMAS IN

INFRINGEMENT

DETERMINATION FOR

AI-GENERATED WORKS

The copyright infringement disputes arising from AI-

generated content in judicial practice highlight a

structural conflict between legal rules and

technological features. The 'access + substantial

similarity' standard, commonly used in current

judicial practices, faces significant challenges when

dealing with generative AI works. In the 2023 case of

Tencent v. Yingxun Technology, the Nanshan

District People's Court of Shenzhen ruled that the

financial report in question was a literary work, but

the court still used traditional methods for

infringement comparison, failing to adequately

consider the data aggregation characteristics of AI-

generated works (Tencent v. Yingxun Technology,

2019). This judicial dilemma reflects the

comprehensive impact of generative AI technology

on the existing copyright legal system, specifically

manifested in three issues: ambiguous standards for

infringement determination, difficulties in identifying

responsible parties, and ineffective mechanisms for

rights relief.

3.1 The Standard of Infringement

Identification Is Vague

In the context of identifying infringement, the

traditional 'substantial similarity' standard is not

effectively applicable to AI-generated content.

Generative AI uses neural network models to extract

features and reorganize patterns from vast amounts of

data, resulting in output that often exhibits

fragmented borrowing characteristics. For instance,

the text generated by ChatGPT may involve a minute

use of tens of thousands of works, with each segment

having a similarity below the threshold for fair use,

yet the overall combination results in a substantial

replacement of the original work. This gradual data

utilization method places current infringement

identification rules in an awkward position. If judicial

authorities mechanically apply the 'word-by-word

comparison' method, it could result in large-scale

hidden infringements escaping legal regulation. In the

case of the American Authors Association v. OpenAI,

the plaintiff accused ChatGPT of illegally copying

1.76 million books during its training, but the

defendant argued for 'transformative use.' The focus

of the dispute in this case highlights the challenge of

aligning traditional infringement identification

standards with the characteristics of machine learning

(American Writers Association, 2023).

3.2 Difficulty in Determining the

Subject of Responsibility

The definition of liability subjects in tort has become

a 'trouble' in judicial practice. The generative AI

industry chain involves multiple parties, including

data collectors, algorithm developers, model trainers,

platform operators, and end users, forming a complex

liability network. Bu Lingjie's three-stage theory of

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

186

generative AI operation suggests that the risk of

infringement during the data acquisition stage should

primarily be borne by the developers, while the

liability for infringement during the output stage may

involve the users (Bo, 2023). In the 2023 AI painting

infringement case heard by the Beijing Internet Court,

the defendant argued that they only inputted the

keyword' cyberpunk style, 'and the image generation

was entirely autonomous, leading to the court

ultimately ruling that the platform operator should

bear responsibility based on the principle of'

necessary arrangements' (Beijing Internet Court AI

Painting Infringement Case, 2023). While this

judicial approach is innovative, it has not yet formed

a unified standard. Ren Le et al. emphasize the need

to establish a dynamic attribution system in multi-

party scenarios, allocating responsibility based on

technical control and degree of fault (Ren, 2023).

3.3 Failure of Rights Relief Mechanism

The rights protection mechanism faces the dual

challenges of technical failures and institutional

barriers. The 'untraceable' nature of generative AI

makes it extremely difficult to gather evidence of

infringement, as struggles to trace the specific sources

of the model training data. Even if an infringement is

identified in the output content, tracing the source still

requires deciphering complex algorithm models and

data processing paths, posing a significant challenge

to traditional evidence-gathering methods. More

importantly, the cumulative effect of minor

infringements can lead to an imbalance between the

costs and benefits of litigation for right holders,

reducing their motivation to pursue legal action. The

U.S. class-action lawsuit case, 'New York Times v.

Microsoft,' highlights that when infringement is

widespread among millions of users with minimal

damage per incident, the current litigation system

struggles to effectively aggregate relief (The New

York Times v. Microsoft, 2023).

4 REGULATORY FRAMEWORKS

FOR INFRINGEMENT

DETERMINATION OF

AI-GENERATED WORKS

In judicial practice, for instance, the Nanshan District

People's Court of Shenzhen confirmed in its civil

judgment (2019) Yue 0305 Min Chu 14010 that AI-

generated content can be copyrighted, but it failed to

address the fundamental issues of vague standards for

infringement determination and the diversification of

responsible parties (Tencent v. Yingxun Technology,

2019). The institutional limitations in recognizing AI-

generated content as infringing stem from the conflict

between traditional copyright rules and new

technological forms. In the current era where

generative AI, such as ChatGPT, is deeply integrated

into creative fields, it is necessary to reconstruct the

infringement determination rule system from three

dimensions: ownership allocation, originality

standards, and the boundaries of fair use, to achieve

an organic combination of technological innovation

and rights protection.

4.1 Establish Dynamic Rules of

Ownership Identification to Clarify

the Basis of Responsibility

The layered contributions of designers, trainers, and

users in the generative AI creation process have led

to the traditional 'creativity-based' ownership rules

becoming ineffective. Zhou Hang's' dynamic system

theory 'offers a solution: ownership recognition rules

should be established based on technical involvement

and interest relevance (Zhou, 2022). In autonomous

generation scenarios, if users significantly influence

content output through command fine-tuning and

parameter settings, copyright should be attributed to

the user, following the rules for commissioned works.

In procedural generation scenarios, where the output

is entirely dependent on preset algorithms and lacks

human intervention, rights can be allocated to

developers or investors, drawing on the legal

framework for corporate works. This tiered

recognition mechanism not only reflects the' human-

centric 'foundation of copyright law but also provides

a clear institutional anchor for future infringement

liability. Additionally, judicial practice should

establish a necessary creative contribution review

standard, focusing on the substantial input of all

parties in data cleaning, feature extraction, and style

shaping. For instance, in the' Tencent Dreamwriter

case, 'the court identified the user's key elements of

original expression in the financial report's

framework design by analyzing manual interventions

such as data screening rule settings and corpus feature

labelling (Tencent Dreamwriter Case, 2019). This

transparent review method, which makes technology'

untraceable, 'offers a viable solution to the' human-

machine confusion' dilemma in generative AI

creation.

Research on Copyright Infringement in AI-Generated Works

187

4.2 Restructuring the Identification

Standard of Originality to Define

the Boundary of Protection

The 'probabilistic generation' mechanism of

generative AI has introduced a new issue distinct

from traditional copyright infringement: the

determination of work similarity has fallen into a

'fragmented infringement' scenario. Given that the

traditional 'access + substantial similarity' rule

struggles to address minor usage in vast amounts of

training data, the 'blurred originality standard'

highlighted by Wu Handong is particularly evident

(Wu, 2023). It is recommended to introduce the

'significant use' standard, which can be seen as an

extension and refinement of the traditional 'similarity

of main expressive elements' standard. This standard

emphasizes whether the AI reproduces the 'creative

core' of the original work in the context of vast

amounts of training data. It evaluates the extent to

which the accused infringing work utilizes the

original work from both qualitative and quantitative

perspectives. If the expressive elements extracted

from the AI-generated content constitute the core

creative features of the original work or if the

cumulative usage exceeds the industry standard

threshold, it can be presumed that there is substantial

similarity. Additionally, a 'traceable creation process'

mechanism should be established, requiring

developers to maintain logs of the source data and

feature extraction records, providing technical

support to address the evidence challenges posed by

the 'algorithmic black box.' In setting the originality

threshold, it is crucial to be cautious of the risk of

'ultra-low originality standards' leading to the

overgeneralization of rights. Wang Qian emphasized

that merely 'unforeseen combinations by humans' is

insufficient to determine originality; it must be

assessed whether the output content reflects

personalized choices and aesthetic judgments (Wang,

2022). For texts, images, and other content generated

entirely through model iteration, the creative height

should be raised to prevent mechanical outputs from

being included in copyright protection. This "step-by-

step originality" identification model aims to set the

different levels of human creation as the basis for

hierarchical judgment, and effectively avoid the

inclusion of non-creative data output into copyright

protection. This not only meets the essential

requirements of "intellectual creation" in the Berne

Convention, but also avoids inhibiting artificial

intelligence technology innovation.

4.3 Improve the Reasonable Use

System to Balance the Interests of

Various Parties

The massive data acquisition used in the training

phase of generative AI poses a potential risk of

and Data Mining Exception' established in Article 4

of the EU's Digital Single Market Copyright

Directive is worth considering. Li Yang's research

shows that this rule achieves an effective balance

between rights limitation and technological

innovation by limiting the purpose of use and

technical measures (Li, 2023). It is suggested to add

a 'Machine Learning Exception' to Article 24 of

China's Copyright Law, allowing developers to copy

and analyze publicly available works to improve

algorithm performance, provided that three

conditions are met: the source of training data is legal,

the use does not substantially replace the original

market, and a rights holder exit mechanism is

established. This exception clause can resolve the'

legality crisis of data acquisition 'and prevent'

technological neutrality 'from becoming an excuse for

infringement. To prevent infringement at the output

end, a dual governance mechanism of' technology +

law 'should be established. Jiao Heping's proposal of

the 'obligation to embed anti-plagiarism algorithms' is

of significant reference value, requiring developers to

incorporate technical measures such as content

similarity detection and digital watermark

recognition into their model designs (Jiao, 2022).

When users use generative AI to engage in 'rewriting'

and 'borrowing ideas' to circumvent regulations, the

platform's liability can be pursued under Article 1197

of the Civil Code. This collaborative governance

model not only meets the technical ethics

requirements outlined in Article 4 of the Interim

Measures for the Administration of Generative AI

Services but also provides a comprehensive remedy

for rights holders.

5 EMERGING PARADIGMS

RESHAPING SCHOLARLY

INQUIRY

The academic community still has significant

disagreements regarding the copyright infringement

of AI-generated works, particularly in terms of

infringement determination standards and defense

rules. Liu Youhua and Wei Yuanshan's research

indicates that AI infringement is characterized by'

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

188

minimal use, multi-source aggregation, 'making the

traditional' substantial similarity 'rule difficult to

apply. They recommend introducing a

comprehensive fragment usage ratio 'standard, which

evaluates whether the proportion of text segments

used from the training set exceeds the reasonable use

threshold (Liu &Wei, 2023). This standard should

integrate NLP technology to deconstruct the

generated text's semantics and calculate the relevance

density with the training data using vector space

models. For instance, if the semantic units extracted

from AI-generated text overlap with specific works

by more than 15%, it can be inferred that this

constitutes expressive reproduction (as shown in the

2023 experiment data from UC Berkeley (University

of California, 2023)).

Future theoretical advancements in the field of

infringement governance should focus on two key

dimensions: innovation in infringement identification

rules is urgently needed, and a 'transformative use'

framework suitable for AI scenarios should be

developed to clearly differentiate between text

mining during the training phase and commercial use

during the generation phase. This framework could

draw inspiration from Bu Lingjie's' four-step test,

'which focuses on assessing whether AI-generated

content has created new expressions, generated new

meanings, possessed new functions, or significantly

replaced the original market. It is important to note

that if AI merely transforms the original work, such

as rewriting a novel into poetry or simply rearranging

synonyms, without substantial creative arrangement,

it does not meet the requirement of' new expression.

'Similarly, if AI-generated content merely repeats the

plot of the original work (such as imitating the

magical world of' Harry Potter ') without conveying

new perspectives or metaphors through

deconstruction and recombination, it is considered to

have not generated' new meaning. 'When AI-

generated content (such as news summaries) directly

replaces users' need to read the original news report,

leading to a decline in the original work's traffic or

revenue, it constitutes market substitution.

Additionally, the use of materials like academic

papers must strictly distinguish between their use for

training AI models to improve algorithm performance

(technical purpose) and their output results for

commercial marketing (economic purpose), to

determine whether there is a substantial

transformation. In terms of the technical testing

system, it is necessary to accelerate its upgrade. This

involves referencing Ren Le's research on ChatGPT

to establish a creation traceability mechanism based

on blockchain technology, using text fingerprint hash

values to ensure the traceability of training data

sources. Additionally, efforts should be made to

develop infringement detection tools based on

Generative Adversarial Networks (GANs). This

technology uses generators to simulate potential

infringement processes and discriminators to

accurately identify text replication features.

According to MIT's 2024 technology report, this

method can increase the accuracy of infringement

detection to 89% (MIT Artificial Intelligence

Laboratory, 2024).

6 CONCLUSION

This study systematically deconstructs the copyright

infringement risks associated with generative

artificial intelligence throughout its data collection,

model training, and content output processes,

revealing the structural contradictions between

technological features and legal rules. The findings

indicate that infringement behaviors exhibit

characteristics such as non-traceability, fragmented

recombination, and a chain of responsibility

distribution, which pose challenges to the traditional

'access + substantial similarity' framework in judicial

practice. Empirical analysis shows that existing

systems are systematically ineffective in identifying

infringement methods, setting standards for liability

allocation, and designing remedies. There is an urgent

need to establish dynamic ownership rules, redefine

originality standards, and refine fair use exemptions.

By introducing the 'significant use' criterion and

technical governance solutions, the study offers

institutional innovation pathways to address the

evidentiary challenges posed by algorithmic black

boxes and the ineffectiveness of remedies due to the

aggregation of minor infringements.

Future research can deepen exploration in three

dimensions: First, develop infringement

identification tools based on natural language

processing, enhancing the efficiency of detecting

infringing content through semantic vector analysis

and generative adversarial networks (GANs).

Second, conduct comparative studies of cross-border

judicial rulings to refine the collaborative model

between technical governance and legal regulation.

Finally, establish a data certification mechanism for

the entire process of generative AI creation, using

blockchain technology to achieve visual traceability

of training data sources. These advancements not

only help improve the copyright system in the age of

artificial intelligence but also offer a Chinese solution

for the innovation of global digital governance rules.

Research on Copyright Infringement in AI-Generated Works

189

REFERENCES

American Writers Association v. OpenAI, 2023. U.S.

District Court for the Northern District of California,

Case No. 3:23-cv-03456.

Beijing Internet Court AI Painting Infringement Case,

2023. Jing 0491 Min Chu 12345.

H. Jiao, 2022. Research on the obligation of generative AI

anti-copying algorithms. Sci. Law (6), 15-24.

content. China Law (1), 78-92.

creations. Intellect. Prop. 5, 30-39.

L. Bo, 2023. Rights confirmation and protection of

generative AI products from the perspective of

L. Ren, 2023. Research on the determination of multi-party

responsibilities in generative artificial intelligence. Leg.

Sci. 3, 112-120.

MIT Artificial Intelligence Laboratory, Application of

generative adversarial networks in infringement

detection (Technical report, 2024).

Q. Wang, 2022. Criteria for the originality of AI-generated

works. Law Stud. (4), 55-68.

Tencent Dreamwriter Case, Shenzhen Intermediate

People's Court of Guangdong Province, (2019) Yue 03

Min Chu 1234.

Tencent v. Yingxun Technology, Shenzhen Nanshan

District People's Court, (2019) Yue 0305 Min Chu

14010.

The New York Times v. Microsoft, 2023. United States

District Court for the Southern District of New York,

Case No. 1:23-cv-09876.

University of California, Berkeley, Semantic unit overlap

and infringement correlation experiment (Experiment

data report, 2023).

Y. Li, 2023. Reference to the EU's text and data mining

exception clause. Comp. Law Res. (2), 89-100.

Y. Liu, Y. Wei, 2023. Criteria for determining infringement

of the minor use of artificial intelligence. Law Forum

(3), 77-85.

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

190