AI and Copyright — Is Generative AI 'Learning' or 'Plagiarizing'?

The debate surrounding generative AI has been clashing from the very beginning with two metaphors. One side argues: AI reading hundreds of millions of texts and images, learning patterns, and creating new ones is no different from a person reading books and learning a painting style—it is 'learning'. The other side argues: scraping copyrighted works wholesale without permission and making money on top of that is merely the industrialization of 'plagiarism', copying someone else's work.

This dispute is not a matter of emotion, but a blank space where the law has not yet settled an answer. Here, copyright refers to the right of creators who made works like writing, art, and music to control their reproduction and use.

First, we clarify the scope of this article. The law we place at the center is the US Copyright Act, and where necessary, we compare it with the EU, UK, Germany, and South Korea. Because copyright laws differ by country, when mentioning rulings, we also record which country and court made the judgment. The progress status of the lawsuits is based on what was checked in October 2026.

There is one more thing to clarify. Most of the materials cited in this article are not peer-reviewed natural science papers, but legal research. They are law school journal articles (law reviews), SSRN working papers released before publication, and academic book chapters. This means they are not agreed-upon facts, but the arguments and analyses of scholars. The exception is a part of 'memorization' that appears later, which is actually measured engineering research. Therefore, this article does not conclude with a single verdict, but shows various perspectives and unresolved parts as they are.

Splitting the Problem in Two

If you look at AI and copyright as a single lump, you lose your way. This is because there are two questions. One academic book chapter summarizes it like this: ① Can outputs created by AI receive copyright protection? In other words, who is recognized as the creator of that output, which is the issue of authorship. ② Is training AI models with copyrighted works an infringement? This is the issue of training infringement (DOI: 10.2139/ssrn.6542719).

The second question is again divided into the input stage and the output stage. The input stage is the act of copying copyrighted works as training data. The output stage is the act of the model outputting protected expressions again (DOI: 10.2139/ssrn.6542719). Other analyses use the same distinction. Generative AI infringement is seen as happening in two places. One is from pulling protected copyrighted works into training data. The other is from AI outputs copying, transforming, or substantially imitating original works (DOI: 10.69971/tipr.4.1.2026.104).

The answers to the two questions can be different. We will look at them one by one.

Question ① — Is AI-Generated Content Copyrightable?

The core of this question is surprisingly not AI, but humans. The current copyright system hangs the condition of protection on "human creative contribution." This characterization is called anthropocentric. A comparative law study surveys Germany, the EU, the United States, the United Kingdom, Japan, and China, and then summarizes as follows. Fully autonomous AI outputs lack a human author in principle. Therefore, they are placed outside of traditional copyright protection (DOI: 10.2139/ssrn.6313438).

Approaches differ by country. According to this study, U.S. jurisprudence consistently applies the "human authorship" requirement. German and EU law require "personal intellectual creation." The UK has separate legislative solutions for computer-generated works. Chinese case law has developed a more functional approach. It is a method of recognizing authorship when there is meaningful human organization and direction (DOI: 10.2139/ssrn.6313438).

Why must it be human? A paper discussing EU law points out that this question is not new. Even when photography appeared in the past, copyright law had to redefine "what is a meaningful human creative input." The current AI debate is a repetition of that history. This paper considers that EU law has consistently placed human authors at the center. The criteria are "free and creative choices," "the stamp of personality," and "expression perceptible by humans." Therefore, it says that integrating AI as a tool is acceptable. On the other hand, it views that outputs lacking human creative choices, such as synthesized voices or digital avatars, can be pushed outside of copyright (DOI: 10.2139/ssrn.6138549).

A more provocative proposal emerges here. An American law review paper views that generative AI churns out high-quality expressions at "zero marginal cost." This means that the cost of making one more piece is virtually nonexistent. Because of that, the utilitarian foundation that "copyright encourages creation to increase the benefit of society as a whole" is shaken. This paper proposes a "human-required creativity" standard. The point is not to protect results that the latest model of the time can produce with only trivial human input. It does not matter whether a person made it or a machine made it (DOI: 10.2139/ssrn.6340338). This paper views that the core of the issue is not whether to recognize AI as an author, but the threshold of human creativity.

The field is not unified. A study analyzing 87 judgments from India, the EU, the US, and Japan views that the trend of recognizing "hybrid authorship" is increasing. It is a method of acknowledging the contributions of both humans and algorithms together. At the same time, it reports that a significant portion of the legal systems examined do not have clear regulations on works made with minimal human intervention (DOI: 10.2139/ssrn.6120486). They converge in the direction of protecting them step-by-step according to the degree of meaningful human contribution. However, exactly where that line is varies by jurisdiction and is still undetermined (DOI: 10.2139/ssrn.6120486, DOI: 10.2139/ssrn.6313438).

On the ownership side, different models are also proposed. There is also a paper analyzing the UK High Court's DABUS patent ruling (Thaler v The Comptroller-General, 2020). This paper argues for the "AI owner" approach. It is a method where the person who owns and controls the AI system claims the rights to its output. It gives parties greater certainty and lower transaction costs than other proposals. However, since this is a proposal that brings arguments from the patent context into copyright, it is not established law (DOI: 10.2966/scrip.190122.5).

Question 2 — Is Training AI on Copyrighted Works Infringement?

Large-scale models massively copy and train on protected works such as books, images, code, and music. This is the question of whether this copying constitutes infringement.

The first fork in the road is whether that copying should be seen as 'copying'. One paper argues that what happens in AI machine learning is merely a kind of 'reading'. It contends that this is a special form of use of works to which copyright does not extend. It views this as similar to how copyright is not involved when a person reads a book (DOI: 10.69724/2786-8834-2025-4-1-72-89). On the other side is the perspective that AI training is infringement because it inevitably entails copying and electronic storage.

When we break down the technological steps, it leans to one side. One study comparing the UK, EU, and the US follows the technological steps of training. These are scraping (scraping data), tokenization (breaking text into small units), and parameterization (converting into numerical values inside the model). This study concludes that AI training inherently entails copying. However, whether that copying is legal varies by jurisdiction. The UK is the strictest, setting the narrowest exceptions. The EU permits training through text and data mining (TDM, a technology that analyzes data in bulk) rules while providing an opt-out. Opt-out is a method where rights holders can opt out by expressing their intent to refuse. This study summarizes that the US grants the broadest protection under fair use (DOI: 10.30574/ijsra.2026.18.1.0166). Even when starting from the same fact that 'copying occurs', the legal conclusions diverge completely from country to country.

America's Four Factors of Fair Use — Is Training 'Transformative'?

The heart of the American debate is fair use. Fair use is an exception in U.S. copyright law that allows copyrighted works to be used without the copyright owner's permission under certain conditions. 17 U.S.C. § 107 considers four factors: ① the purpose and character of the use (especially whether it is transformative or commercial), ② the nature of the copyrighted work (whether it is highly creative or factual), ③ the amount and substantiality of the portion used (how much, and whether core parts were used), and ④ the effect upon the market. The AI training debate is fiercely waged on top of these factors.

The core issue is the first factor's 'transformative use'. It asks whether the work is used for a new purpose or with a new character rather than simply repeating the original. The ruling in Bartz v. Anthropic drew attention as a watershed moment by recognizing AI training as a transformative use. This ruling did not view AI training as simple copying. It viewed it as a creative process that extracts language patterns and statistical structures to create new outcomes. It also judged that the training data market cannot be monopolized solely by copyright holders (DOI: 10.15539/khlj.60.3.3). This ruling was issued by the U.S. District Court for the Northern District of California in June 2025. The judgment concerns the portion trained on legally acquired books. The same court viewed that the act of downloading illegal copies to build a library was not fair use. This case ended with a class settlement regarding the illegal copy library claims, and the court gave final approval to the settlement in July 2026.

A study analyzing this ruling in the Korean context attaches the caveat that it is too early to view it as established law. Past cases have sometimes denied fair use in the exact opposite direction. Korean courts have generally held a conservative attitude that places more weight on protecting copyright owners. Therefore, it is said that there was narrow room to accept new uses like AI training (DOI: 10.15539/khlj.60.3.3).

The Bartz ruling itself faces fierce criticism. An SSRN paper argues as follows. The court's view of training as "exceedingly transformative" misconstrued the first factor. The court defined the 'character of the use' at the level of intermediate-stage mechanisms such as training and weight transformation. This paper argues that it should have anchored on the actual purpose toward the market that triggers and monetizes that copying (DOI: 10.2139/ssrn.6169529).

Another paper points out the same 'framing problem'. Foundation models are not single-function tools but general-purpose systems used for various purposes. Thus, litigants can describe 'use' at arbitrary levels of abstraction, ranging from training mechanisms to social innovation. That choice dictates the assessment of transformativeness. Therefore, it proposes inferring purpose from observable evidence such as the developer's monetization structure and design choices (DOI: 10.2139/ssrn.6324500).

Market harm, the fourth factor, is also controversial. One paper considers that AI training can be transformative, and copying occurs 'upstream' of competitive outputs. At the same time, it points out that the harm of generative AI targets not the market for specific works, but 'future human creation' itself. If so, that harm, no matter how real, may lie outside the scope of the fourth factor. This is because this factor is designed to recognize 'market substitution harm'. In that case, the doctrine of unjust enrichment may be more appropriate. Unjust enrichment is a legal doctrine that requires the return of benefits unfairly obtained using someone else's property (DOI: 10.2139/ssrn.6111926).

There is also a more fundamental counterargument. An essay criticizes that the fair use debate has become "too technical and too far removed from common economic sense." AI pays for every other input like engineers, chips, cloud, electricity, and capital. Yet, it attempts to use for free only the human expressions that make the model valuable. This essay also seriously engages with defensive arguments that put innovation first. Even so, it considers that the core question is not 'is AI useful?' The core is whether fair use should allow commercial technology to treat creative works as free industrial raw materials (DOI: 10.2139/ssrn.6740239).

There is also a perspective that casts doubt on transformativeness itself. One paper considers that the transformative use doctrine is still inadequately defined and inconsistent. It is insufficient to handle the functional and non-expressive nature of AI training. This paper cites as evidence two lawsuits filed in the U.S. District Court for the Southern District of New York. One is New York Times v. OpenAI. Filed in 2023, the court allowed major copyright infringement claims to proceed in March 2025. As of October 2026, the first instance trial is underway, at the stage where motions for summary judgment regarding fair use have been submitted. A summary judgment is a ruling made based on documents alone before a formal trial. The other is Dow Jones v. Perplexity. Filed in 2024, the court dismissed all motions regarding jurisdiction and venue in August 2025. This case is also undergoing its first instance trial. The paper summarizes that this litigation landscape shows the urgency of legislative clarification (DOI: 10.46553/bri.2025.5054).

When a Model Memorizes and Spits Back What It Learned — Memorization

The point where the metaphor of "learning versus plagiarism" collapses most dangerously is memorization. This refers to the phenomenon where a model memorizes training data just as it is. An ideal model should only learn the general patterns of the data and forget the individual original texts. However, research shows that actual LLMs (Large Language Models) memorize parts of their training data as they are, and can pull them back out with specific prompts (input sentences).

This is not a legal scholar's metaphor, but an engineering-measured phenomenon. An empirical study published in AAAI to evaluate the risk of personal information leaks proposed a precise entity-level definition of memorization. It also showed a method to efficiently extract sensitive entities from autoregressive models. In the experiments, LLMs strongly memorized at the entity level. They reproduced training data with only partial clues and appeared to understand associations between entities (DOI: 10.1609/aaai.v38i17.29948). LLMs fine-tuned for classification also memorize a significant portion of their training data. Research indicates that if they are exposed at inference time, this becomes a severe threat (DOI: 10.3233/faia251323). In other domains like protein language models, detectable memorization was measured, though the signal was weak (DOI: 10.64898/2026.06.08.730685).

This fact shakes the legal metaphor once again. A paper dealing with amendments to Russian civil law proposes a balance point like this: a distinction where internal memorization itself is not viewed as infringement, and only content generation that reproduces a substantial part of copyrighted works is viewed as an infringement of exclusive rights. It adds that generation that mimics a specific author's writing style as-is can also be an infringement (DOI: 10.2139/ssrn.6690979). In this proposal, the boundary between "reading" and "copying" is drawn at the output, not the input. This is a different approach from the Munich District Court's judgment (the GEMA case below) that viewed memorization inside the model as copying.

As for solutions, there are proposals that directly target memorization. A paper comparing Germany, the UK, and the US divides the market into two. One is the market where AI accesses training data, and the other is the market of outputs reproduced by AI. Reflecting the harm in the latter, it proposes a capability-based levy linked to the measurable memorization rate of the deployed model. Non-monetary harms arising from using works without citing sources are handled differently. It is a method of making the ability to indicate sources mandatory at the model design level (DOI: 10.2139/ssrn.6814498). The method linked to the memorization rate is said to be less exposed to evading detection by moving across territories than methods based on training data. It also gives developers a greater incentive to reduce memorization (DOI: 10.2139/ssrn.6814498).

Different Answers in Different Countries — US Fair Use, EU TDM, and Beyond

From this point on, the fact that "there is no single right answer" is the answer itself. The exact same action swings between legal and illegal depending on which country you are standing in.

The EU permits training through the text and data mining (TDM) exception under Articles 3 and 4 of the Digital Single Market (DSM) Directive. Instead, it provides an opt-out structure allowing rightsholders to object. Whether this opt-out actually works is the issue. One paper comparing the US and EU suggests that two conditions are necessary for an opt-out to be effective. Rightsholders must be aware of the usage fact and the refusal option, and the refusal procedure must be possible with reasonable effort. And it concludes that the current AI training opt-out fails to meet these conditions and is not functioning properly (DOI: 10.2139/ssrn.6163107).

There is also a higher ceiling called international treaties. One paper considers that whether it is US fair use or EU TDM, they fall within the treaty constraints of the "three-step test" set by the Berne Convention and others. The demands of this test are three-fold. Exceptions must be restricted to special cases. They must not conflict with the normal exploitation of the work. They must not unreasonably prejudice the legitimate interests of the rightsholder. However, an IIC paper questions whether this test even applies to TDM in the first place. It analyzes that even if it does apply, the TDM exception can pass the test. The conditions are an opt-out opportunity, clear policy goals such as supporting scientific research, and fair compensation for commercial development (DOI: 10.1007/s40319-026-01680-2). The earlier paper states that these constraints have been overlooked in the meantime and proposes reforms conditioned on transparency, auditing, and compensation (DOI: 10.2139/ssrn.6447160).

Ongoing lawsuits add flesh to this abstract debate. Below are the courts and statuses of the cases cited by the papers.

• One chapter covering EU law considers that two rulings diverged on whether a model is a "copy" of a copyrighted work. The High Court of England and Wales dismissed secondary infringement claims in November 2025 in Getty Images v. Stability AI. The reasoning was that the image generation model Stable Diffusion does not store or reproduce copyrighted works, so it is not an infringing copy. Getty was granted leave to appeal this part in December 2025, and the outcome of the appeal is unconfirmed. The Munich Regional Court in Germany held in GEMA v. OpenAI in November 2025 regarding the matter of 9 lyrics that memorization inside a model constitutes reproduction. OpenAI appealed, so this ruling is not yet final. The two cases also have different issues. The UK issue is the interpretation of an "infringing copy," while the German issue is whether the memorization of lyrics is reproduction. This chapter also points out a structural problem. Protected content is encoded in model parameters, making "surgical removal" technically nearly impossible (DOI: 10.2139/ssrn.6356599).
• A study analyzing US cases with a focus on Taiwanese law emphasizes the wall of evidence. According to this study, the US District Court for the Northern District of California dismissed part of the claims in Andersen v. Stability AI, finding factual allegations regarding compressed copies insufficient. This case is at the first-instance stage awaiting summary judgment hearings as of October 2026. On the other hand, in New York Times v. OpenAI, the plaintiffs submitted ChatGPT outputs resembling the original works. The study considers that this suggests the possibility of reproduction and derivative work right infringement (DOI: 10.33093/ajlp.2025.19).
• Another paper comparing the US and EU takes Thaler v. Perlmutter, Andersen v. Stability AI, and the EU's TDM opt-out as reference points to point out three persistent gaps. First, the threshold for human creative control is indeterminate. Second, responsibility is misaligned across a multi-layered value chain. Third, it is difficult to prove reproduction amid opaque training processes (DOI: 10.2139/ssrn.6159507). The progress of the cases themselves is as follows. Thaler v. Perlmutter is a case where a copyright registration application listing AI as the author was rejected. The US Court of Appeals for the D.C. Circuit ruled in March 2025 that the author must be human. The Supreme Court denied certiorari in March 2026.

An encyclopedia entry comparing the UK, US, and Germany considers that current law is particularly powerless against imitation at the level of meaning and style. There is no consensus on authorship. There is no clear licensing model for training data. Protection of moral rights regarding a creator's voice, style, and persona is also fragmented (DOI: 10.3390/encyclopedia6010009).

When Training on Scientific Literature

There is a corner of this debate that hits close to home for us. It is the problem of training LLMs on scientific papers. One paper argues that American fair use is actually ill-suited for scientific literature-based LLMs. The reasons are fourfold. The accuracy, transparency, and interpretability that science values conflict with the "transformative use" requirement. Scientific expression is for precision, not creative originality. Fair use emphasizes "using only a small amount from a single work." This contradicts the need to comprehensively train on information-dense scientific texts. It also argues that commercial use restrictions block a model that is beneficial to researchers, publishers, developers, and the public alike. Therefore, this paper proposes redrawing the exception for scientific literature and eliminating commercial use restrictions (DOI: 10.3389/frai.2026.1781692).

Summary So Far — What the Law Has Not Decided Yet

The law has not settled on an answer yet. Three things are clear.

First, the problem is not one, but two. The authorship of outputs and whether training constitutes infringement must be answered separately, and the latter breaks down again into input and output (DOI: 10.2139/ssrn.6542719).

Second, there is a loose convergence regarding the authorship of AI outputs. The direction is that fully autonomous outputs are not protected, and protection is granted only when there is meaningful human contribution. However, exactly where that threshold lies varies by jurisdiction and remains uncertain (DOI: 10.2139/ssrn.6313438, DOI: 10.2139/ssrn.6120486).

Third, there is almost no convergence on training infringement. Even starting from the same technical fact that "reproduction occurs," opinions diverge. The U.S. leans toward broad fair use, the EU toward opt-out TDM, and the UK toward a stricter stance (DOI: 10.30574/ijsra.2026.18.1.0166). One chapter covering EU law notes that Getty (UK) and GEMA (Germany) diverged on whether the model is a copy. Both are at the first-instance level and pending appeal (DOI: 10.2139/ssrn.6356599).

Therefore, the neat dichotomy of "training versus plagiarism" is close to a wrong question rather than an answer. Reproduction at the input stage is viewed from one perspective as "reading" (DOI: 10.69724/2786-8834-2025-4-1-72-89) and from another as the exploitation of free industrial raw material (DOI: 10.2139/ssrn.6740239). Memorization at the output stage redraws that boundary, creating a line where internal memorization may pass, but the reproduction of protected expression is infringement (DOI: 10.2139/ssrn.6690979). The law is shaping these lines case by case and bill by bill.

This process is what the law always goes through in front of new technology. Photography did (DOI: 10.2139/ssrn.6138549), and computer programs and sound recordings did. Before that process finishes, neither "AI training is legal" nor "AI is a plagiarism machine" is something the law has said yet. What can be said right now is that it depends on the jurisdiction, depends on the input and output, and case law is currently being formed.

This article is for informational purposes regarding law and policy, and most of the cited evidence consists of arguments and analyses from legal research—law reviews, SSRN working papers, and academic book chapters—rather than peer-reviewed scientific papers. Because many copyright issues are unresolved and vary by jurisdiction, this does not replace legal advice for a specific matter.

Source Papers

The Two Questions at Issue (Authorship / Training Infringement · Input/Output)

• Artificial Intelligence and Copyright (Training vs. Output, Authorship in China, UK, and US) — DOI: 10.2139/ssrn.6542719
• AI Copyright Infringement: Navigating the Legal Risks of AI-Generated Content (The two infringements of input and output) — DOI: 10.69971/tipr.4.1.2026.104

Question ① — Authorship of AI-Generated Works (The Human Requirement)

• Generative AI and Copyright Law: Human Authorship, Originality, Comparative (Autonomous output = public domain) — DOI: 10.2139/ssrn.6313438
• Human Agency at the Core (EU, Historical parallel with photography) — DOI: 10.2139/ssrn.6138549
• Human-Required Originality: Copyright Eligibility in a Post-AI World — DOI: 10.2139/ssrn.6340338
• Intellectual Property Rights in AI-Generated Creative Works (87 rulings, jurisdictional discrepancies) — DOI: 10.2139/ssrn.6120486
• Copyright in AI-generated works: Lessons from patent law (Thaler/DABUS, AI ownership) — DOI: 10.2966/scrip.190122.5

Question ② — Training Infringement (Copying vs. 'Reading')

• The Legal Status of AI Training Data (UK strict / EU TDM / US fair use) — DOI: 10.30574/ijsra.2026.18.1.0166
• Machine learning is not a Copyright use of a Work (The 'reading' argument, NYT v. OpenAI) — DOI: 10.69724/2786-8834-2025-4-1-72-89

The Four Factors of Fair Use · The Transformative Use Debate

• A Study on Fair Use in AI Learning: Bartz v. Anthropic (Korean context) — DOI: 10.15539/khlj.60.3.3
• Bartz v. Anthropic and the Misframing of Fair Use Factor One — DOI: 10.2139/ssrn.6169529
• Revealing Purpose and Character in the Age of AI (Framing Factor One) — DOI: 10.2139/ssrn.6324500
• When Market Disruption Is Not Market Harm (Factor Four, unjust enrichment) — DOI: 10.2139/ssrn.6111926
• Will Fair Use Save AI from Common Sense? (Free industrial raw material) — DOI: 10.2139/ssrn.6740239
• Copy, Paste, and Generate (NYT and Perplexity lawsuits) — DOI: 10.46553/bri.2025.5054

Memorization (A Machine That Memorizes and Spits Out — Some Measured by Engineering Research)

• Quantifying Entity-Level Memorization in LLMs (Extractability) — DOI: 10.1609/aaai.v38i17.29948
• Predicting Memorization Within LLMs Fine-Tuned for Classification — DOI: 10.3233/faia251323
• Pseudoperplexity Probes Memorization in Protein Language Models — DOI: 10.64898/2026.06.08.730685
• Copyright and AI: The Importance of Memorisation and Attribution (Compensation linked to memorization rates) — DOI: 10.2139/ssrn.6814498
• Copyright Borders in the Era of AI (Internal memorization is OK, reproduction is infringement) — DOI: 10.2139/ssrn.6690979

Jurisdictional Divides · TDM · The Three-Step Test · Lawsuits

• Text and Data Mining, Generative AI, and the Three-Step Test — DOI: 10.1007/s40319-026-01680-2
• The Three-Step Test in International Copyright (US and EU) — DOI: 10.2139/ssrn.6447160
• Opt-Outs in Copyright Law (Dysfunction of opt-out mechanisms) — DOI: 10.2139/ssrn.6163107
• AI in EU Law: Training, Memorization (Conflicting rulings in Getty vs. GEMA) — DOI: 10.2139/ssrn.6356599
• Legal Challenges and Evidentiary Barriers (Andersen and NYT evidence) — DOI: 10.33093/ajlp.2025.19
• AI-Generated Content and Copyright Law: Legal Gaps (Thaler and Andersen) — DOI: 10.2139/ssrn.6159507
• AI, Authorship, Copyright, and Human Originality (UK, US, and Germany, style and moral rights) — DOI: 10.3390/encyclopedia6010009

The Special Front of Scientific Literature

• Limitations of copyright frameworks for LLMs trained on scientific literature — DOI: 10.3389/frai.2026.1781692