Generative AI Models and Copyright

By Kim Won, Attorney at Kim & Chang 

1. Introduction 

Rapid developments and achievements¹⁾ in recent years, along with corresponding expectations and concerns, have driven active research and discussion on AI, particularly generative AI (GAI).²⁾

Among the fields expected to be significantly affected by GAI are creative fields, such as literature, music, and the arts. In the past, it was argued that AI would not be able to fulfill the role of a creator in the arts, which rely on evoking human emotions. However, advances in AI indicate it could significantly (if not entirely) assume the role of creators in these fields.

There are a few opinions that AI is able to fulfill this role as a creator due to its ability to learn from existing works. This learning (without a separate licensing agreement) constitutes copyright infringement or unfair competition, thereby resulting in the unauthorized use of those works.³⁾  This has sparked numerous disputes across many countries.

This article will outline the learning and output principles of GAI models, and provide a concise overview of copyright law issues and recent U.S. cases.

2. Learning and Output Principles of GAI Models, Especially Large Language Models (LLMs)

GAI models can generate new content, such as text, images, and music, and their performance strongly correlates with the number of parameters, training data, and operations.

Large language models (LLMs) are a specialized form of GAI designed to understand and process human language (text) and generate new text. The LLM development and service process can be simplified into three stages: ① collecting training data, ② training the AI model, and ③ producing AI model results. The following will provide an overview of the learning and output principles of LLMs in this order.

A. Collecting training data for LLM development

Model training data is required for AI development. Securing sufficient, high-quality training data is a key requirement for AI development and competitiveness. Accordingly, many countries have enacted special regulations regarding data collection (Text and Data Mining, “TDM”).⁴⁾  Data collection used for training, validating, or testing an AI model is called a “data set.” The most basic unit that LLM processes text data is called a “token.” Just as humans recognize sentences as words or characters, LLM recognizes and processes text by dividing it into units called “tokens.”

B. Training phase for LLM development

The data collected for LLM development requires a training process, largely divided into pre-training and post-training. During the “pre-training” phase, the LLM acquires a broad spectrum of general knowledge, including the basic structure, grammar, patterns, and relationships in human language, thereby creating a “foundation model” capable of general language understanding and generation. This general-purpose foundation model can be unoptimized for specific tasks or produce inaccurate or undesirable responses, and the “post-training”⁵⁾  phase curtails these issues.

1) Pre-training phase for LLM development

The pre-training phase develops a sophisticated “step function” that statistically identifies the complex patterns, structures, and semantic relationships inherent in a large and diverse data set (corpus).

LLM is based on a “Deep Neural Network,” specifically the “Transformer” architecture, which combines a range of mathematical functions, including matrix multiplication and activation functions, across multiple layers. The transformer architecture, introduced in “Attention Is All You Need (NIPS, 2017)” by Google Brain and others, is designed to effectively identify contextual relationships among elements (tokens) within input data (token sequences) through a mechanism called “self-attention.”

The process of building a sophisticated “step function” can be described as finding “parameters” or “weights” that optimize its performance. For the function “f(x,y) = Ax + By,” the numerical values of the function’s parameters, such as A and B, are the parameters (weights).

LLM utilizes a “self-supervised learning” method in which the AI model generates its own learning objectives (predicting the next word) and corrects the answers from the training data, and then learns from these to adjust the values of each parameter, ranging from billions to trillions, without separate correct answers.⁶⁾

2) Post-training phase for LLM development

The post-training phase fine-tunes the performance of the basic model for a specific purpose, or aligns its behaviors with human values and preferences.

The representative method, “supervised fine-tuning,” further trains a pre-trained AI model to generate more accurate, useful, and desired responses to specific types of instructions or questions. Instead of the massive general data set used for pre-training, a relatively smaller, high-quality “instruction-response” data set, ranging from thousands to hundreds of thousands of entries, is used, created, or carefully selected by experts for the specific task.

Another representative method, known as “Reinforcement Learning from Human Feedback,” ensures the AI model's attitudes and values reflect those favored by people, allowing it to interact in ways users appreciate. During this phase, the AI model generates multiple responses to a single prompt, after which “preference data” is used to directly compare, rank, or score each response. By relying exclusively on human feedback on the AI model's output in response to the prompt as training data, this learning method strengthens the AI model's generative ability.

D. The output (generation) process of GAI

GAI model content output arises from probabilistically calculating the most likely next content for a given input (prompt), using the extensive statistical patterns and rules learned during training, and then combining these probabilistically to “generate” new outputs.

In simple terms, it is the computational process of inputting a complex “mathematical function (model)” to obtain a result.⁷⁾  The key of the output (generation) process is not retrieving pre-stored data, but applying a trained “step function” that performs complex internal calculations on the given input in real time, probabilistically predicting the next token and combining new elements based on that prediction. Generally, subtle differences in the computational process, such as temperature-based stochastic sampling, are designed to produce different results, enabling creative and flexible responses even when given the same prompt. 

3. Copyright Law Issues Related to LLM Development and Services

In LLM development and services, securing sufficient, high-quality learning data is essential. For this reason, many AI-related companies use “crawling,” a method that automatically gathers content to serve as AI learning data, often without the consent of website owners (though data may be obtained under separate contracts or platform terms and conditions). Several platform terms and conditions allow the use of member content for “new service development” or “AI development,” and there appears to be no disputes yet regarding the legality of this data collection method. Most ongoing disputes typically involve the unconsented collection and use of content from media outlets and other organizations through crawling.⁸⁾

While this is not unrelated to learning data collection method, the question of whether training AI models and providing outputs from these models constitute copyright infringement is also complex. As briefly discussed above, the transformer-based LLM divides large amounts of text data into “tokens,” units that a computer can process. A tokenizer refers to a highly complex and sophisticated mathematical function (model) that performs tokenization based on the entire data set, taking into account string frequency and probability, and then finds parameters that optimize the performance of the “step function.” This process allows the LLM to calculate the probability distribution of the next output sequence, given a specific input sequence, using the parameter values.

The LLM learning process does not store or utilize a literal copy of specific data; however, it divides the data into “tokens,” recognizes them, extracts universal principles and characteristics, and builds a statistical model with generalized predictive capabilities. Since the LLM content generation process involves “function computation,” in which input is applied to a complex “mathematical function model),” and the resulting output is produced, it is difficult to determine whether the learning and generation processes of GAI models constitute copyright infringement.

The following defenses illustrate the argument that copyright infringement is recognized when a copyrighted work is copied or modified during AI training.

First, copyright law protects the expression of a copyrighted work. Therefore, merely analyzing the language or visual patterns of a copyrighted work (without replicating the expression) for AI training purposes does not constitute copyright infringement. For literary works, the AI training process irreversibly transforms the copyrighted work into a vector containing the language patterns through tokenization and embedding, effectively removing creative expression from the copyrighted work.

Next, using or reproducing a copyrighted work during AI training or production constitutes fair use and does not constitute copyright infringement.⁹⁾  Fair use of copyrighted works is a defense often invoked by defendants in most U.S. cases alleging copyright infringement during AI model training. In these cases, the defendants argue that the AI model does not copy copyrighted creative expression but merely learns language patterns. In addition, the goal is not to “copy” the learned work, but to generate new data based on an understanding of language, logic, and the world, which is a classic example of transformative fair use.¹⁰⁾

In August 2024, a relevant decision was reached in the U.S. District Court for the Northern District of California in a copyright infringement lawsuit filed by author Andrea Bartz et al. against Anthropic, the developer of the AI model “Claude” (Bartz et al. v. Anthropic PBC, Case 3:24-cv-05417).

In this case, the defendant filed a motion for summary judgment, maintaining that using copyrighted works for AI model training constituted fair use. Judge William Alsup partly granted and partly denied the motion of the defendant on June 23, 2025. The defendant Anthropic downloaded over 7 million books from pirate sites, such as “Books3” and “LibGen” for its LLM studies and, after purchasing millions of printed books, destroyed their bindings, scanned them, and converted them into digital files. All digital books acquired in this manner were permanently archived in a “central research library,” where engineers separately copied portions of this library into “data mixes.” These mixes were then cleaned and tokenized to facilitate LLM model learning. 

Fair use was determined for each issue. The court did not view the defendant's copying as a single act of “LLM study;” instead, it distinguished multiple independent copies based on objective usage patterns and sources, and assessed fair use individually.

Specifically, the court separated the copies used for specific LLM training and classified the remaining copies into three categories forming a central library, which is then divided into (i) the copies used to convert purchased print books into digital form, and (ii) the pirated copies used to build a central library. The court recognized both ① and ②-(i) as fair use (the motion for summary judgment granted), while ②-(ii) was not found to be fair use (the motion for summary judgment denied). The motion for summary judgment was denied for those copies not involved in other LLM studies due to insufficient discovery.

The court held that the copies used to train specific LLMs qualify as fair use for the following reasons:

●  Purpose and nature of the use (advantageous): It is spectacularly transformative. It does not replace the copyrighted work; instead, it creates a tool that maps statistical relationships between texts to generate new outputs. The court ruled that even if it were assumed that LLM “compressed” and effectively “memorized” the copyrighted work, the transformative nature of the use remained intact because, unlike search AI (as in the Thomson Reuters case), the AI in question was “GAI,” meaning it generates new content.

●  Nature of the copyrighted work (disadvantageous): The fact that all books, regardless of their publication status or whether they are novels, were selected for their expressive characteristics prejudiced the fair use claim.

●  Amount and significance of the portion used (advantageous): While the entire book was copied, the amount was “reasonably necessary” for the transformative purpose of LLM training. Based on the precedent, the court determined that “the amount of work made available to the public and serving as a market substitute” outweighed the number of copies made, and ruled that there was no issue in this case because the infringing results were not disclosed.

● Effect on the market (advantageous): The use does not replace the traditional market demand for the original work. While a market for AI training licensing could emerge, the court ruled that copyright law does not protect control over such transformative uses.

1) After a long period of development marked by ups and downs, AI has rapidly advanced since Google Brain and other organizations published the paper “Attention Is All You Need” in 2017. Since OpenAI released ChatGPT, AI technology and models have advanced at an unpredictable rate. 

2) “Generative Artificial Intelligence (GAI)” refers to AI technology that generates new content such as text, images, video, audio, and code based on user requests, utilizing models created via large-scale data learning. Services provided to the public or private sector using such GAI are called “GAI services” (Korea Media Communications Commission, GAI Service User Protection Guidelines, 2025, p. 6). 

3) Even setting aside the unfair competition from unauthorized use of the results, there are various opinions that can be compared to the contention of a hundred schools of thought. These include the view that AI learning copyrighted works constitutes a violation of copyright law, denying it on the grounds that it constitutes fair use, the view that it does not constitute fair use but is copyright infringement, and the view that whether it constitutes fair use may depend on how the work is learned.
4) For instance, the EU has introduced a provision stating that “Reproduction and extraction of legally accessible works, etc., for text and data mining purposes is permitted.” (Article 4 of the EU DSM Directive). Germany has introduced a provision stating that “Reproduction of legally accessible works is permitted if the rights holder has not reserved the right.” (Article 44b, Paragraphs 2 and 3 of the German Copyright Act). Japan has introduced a provision stating that “Any method may be used within the limits deemed necessary, without restrictions on the subject, purpose, or method of use, as long as it does not unfairly infringe upon the interests of the copyright holder.”
5) It is sometimes called fine-tuning. Recently, as various learning techniques have advanced, the process after pre-training is now referred to by different terms, such as post-training and alignment training.
6) Each parameter value determines the specific form and connection strength of each function combined in multiple stages. The LLM calculates the error between its prediction and the actual data using the loss function (a function that quantitatively measures the difference between the predicted result of the AI model and the actual correct answer), and uses backpropagation (an algorithm that efficiently adjusts the parameter values by tracing and calculating the influence of each parameter inside the AI model on the error calculated through the loss function), along with an optimization algorithm to repeatedly and minutely update each parameter value, aimed at minimizing this error.
7) It is accomplished by “calculating” the probability distribution of the next token (word, etc.) using an inherent probability model (function) built through learning, and then "sampling" the next token from that distribution to append it sequentially. Instead of merely retrieving and outputting stored data, it generates new sequences through probabilistic predictions and selections at each stage, hence the term “generative AI.”
8) Here is a brief overview of the actions at issue in the ongoing lawsuits: The New York Times v. OpenAI, Microsoft lawsuit concerns OpenAI's unauthorized scraping (crawling) of news content from the New York Times and using it to train AI models; the Raw Story v. OpenAI lawsuit concerns OpenAI's scraping (crawling) of online news content from Raw Story (a media outlet), deleting the CMI, and then using it to train AI models; the Dow Jones & Company v. Perplexity AI lawsuit concerns Perplexity's use of copied (crawled) news content from the Wall Street Journal and New York Post for its RAG Index technology; and the Anderson v. Stability AI lawsuit concerns Stability AI’s use of billions of copyrighted images through internet scraping (crawling) and storing them for AI training.
9) Here is a brief overview of the actions at issue in the ongoing lawsuits: The New York Times v. OpenAI, Microsoft lawsuit concerns OpenAI's unauthorized scraping (crawling) of news content from the New York Times and using it to train AI models; the Raw Story v. OpenAI lawsuit concerns OpenAI's scraping (crawling) of online news content from Raw Story (a media outlet), deleting the CMI, and then using it to train AI models; the Dow Jones & Company v. Perplexity AI lawsuit concerns Perplexity's use of copied (crawled) news content from the Wall Street Journal and New York Post for its RAG Index technology; and the Anderson v. Stability AI lawsuit concerns Stability AI’s use of billions of copyrighted images through internet scraping (crawling) and storing them for AI training.
10) Tremblay v. OpenAI, Inc., No. 3:23-cv-03223 (N.D. Cal. Aug. 27, 2024); Authors Guild v. OpenAI, Inc., No. 1:23-cv-08292 (S.D.N.Y. Feb. 16, 2024); Concord Music Group, Inc. v. Anthropic PBC, No. 5:24-cv-03811 (N.D. Cal. Aug. 22, 2024), etc.