The Mechanics of AI Copyright: Comparing Fair Use, Transformative Use, and the Evidence on Training Data

When a frontier AI model processes its first billion tokens of text, it does not save a single sentence to a hard drive. Instead, it adjusts mathematical weights across a vast neural network, creating a statistical map of human language. This technical reality sits at the center of the most consequential legal debate in modern technology: whether the act of training an artificial intelligence on copyrighted material is a protected form of learning or a massive scale of infringement.[2]

The mechanics of this process are crucial to understanding the legal arguments. During pre-training, a model ingests text, images, or code, breaking them down into discrete tokens. It analyzes the relationships between these tokens, updating its internal parameters to predict what comes next. Once the training phase is complete, the original data is discarded. The model retains only the mathematical patterns—the weights and biases—derived from that data.[2]

Because the model does not store a database of the original works, AI developers argue that training is fundamentally different from copying or pirating. They liken the process to a human student reading thousands of books in a library to learn how to write. The student does not memorize the books verbatim, but rather absorbs the vocabulary, syntax, and stylistic conventions. From a technical standpoint, the model is performing a highly advanced version of this same pattern recognition.

However, copyright holders argue that this analogy breaks down when applied to commercial machines operating at planetary scale. While a human might read a few hundred books in a lifetime, a frontier model ingests millions of copyrighted works in a matter of weeks, often without permission or compensation. The resulting system is then commercialized, sometimes directly competing with the creators whose labor was used to train it. This tension has forced the legal system to re-examine the boundaries of copyright law.

In the United States, the primary legal defense for AI training is the doctrine of fair use. Fair use is a flexible framework designed to promote freedom of expression by permitting the unlicensed use of copyright-protected works in certain circumstances. Courts evaluate fair use based on four factors: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect of the use upon the potential market for the original work.[3]

The first factor—the purpose and character of the use—is often the most heavily litigated. Specifically, courts look at whether the new use is transformative. A use is considered transformative if it adds something new, with a further purpose or different character, rather than merely superseding the original creation. AI developers argue that training is inherently transformative because it converts expressive works into unprotectable facts and statistical correlations.

Legal precedent provides some support for this view. In previous cases involving search engines and digital archives, courts have ruled that creating a searchable database of copyrighted works, or displaying low-resolution thumbnails, constitutes a transformative fair use because it serves a fundamentally different function than the original media. Proponents of AI argue that extracting statistical patterns to build a generative model is an even more profound transformation than creating a search index.

Yet, the fourth fair use factor—the effect on the market—presents a significant hurdle for generative AI. If an AI model can generate output that serves as a market substitute for the original works it was trained on, the fair use defense weakens considerably. For example, if a model trained on a specific illustrator's portfolio can generate new images in that exact style on demand, it directly threatens the illustrator's livelihood.

To navigate this, legal analysts are increasingly drawing a distinction between the input stage and the output stage of generative AI. The input stage involves the mass ingestion of data for training, which many legal scholars argue leans toward fair use because the intermediate copies are not distributed to the public. The output stage, however, is where the model generates new content. If the output is substantially similar to a specific training work, it may constitute infringement, regardless of how the model was trained.

The U.S. Copyright Office has also weighed in on the output side of the equation. In its formal guidance, the Office clarified that copyright protection extends only to works created by human beings. Material generated entirely by an artificial intelligence system, without sufficient human creative input, cannot be copyrighted. This means that while the training data may be protected, the raw outputs of the model are generally considered to be in the public domain.[1]

This creates a complex legal asymmetry. A company might use copyrighted data to train a model under the shield of fair use, but it cannot claim copyright ownership over the resulting AI-generated outputs unless a human significantly modifies or arranges them. The Copyright Office requires applicants to explicitly disclose when a work contains AI-generated material, ensuring that the boundaries of human authorship are maintained in the public record.[1]

As federal courts begin to issue rulings on these matters, the legal landscape remains highly fragmented. Some early decisions have dismissed claims that AI models are inherently derivative works of their training data, noting that the models themselves do not contain the copyrighted expressions. However, courts have allowed claims related to specific, infringing outputs to proceed, indicating that the fair use defense is not an absolute shield.

The technical architecture of the models complicates these legal proceedings. Because neural networks are black boxes, it is exceedingly difficult to prove that a specific piece of copyrighted data was used to generate a specific output. Researchers are developing new methods for machine unlearning and data provenance, but these technologies are still in their infancy. Until they mature, plaintiffs face a steep evidentiary burden in proving direct infringement.[2]

Ultimately, the resolution of the AI copyright debate will likely require a combination of judicial precedent, legislative action, and technical innovation. Whether through the establishment of mandatory licensing frameworks, the refinement of the fair use doctrine, or the development of models trained exclusively on public domain data, the legal system is actively working to balance the rights of creators with the rapid advancement of artificial intelligence.[3]