Fair Use Is About the Use, Not the Label

This chapter focuses on the first fair-use factor in the Copyright Office’s analysis of generative AI training: the purpose and character of the particular use. That focus matters because the same copyrighted work can be copied during dataset creation, broad model training, retrieval, or output, yet those acts can serve different purposes and expose different risks. A reader who names the use precisely can evaluate transformation, commerciality, lawful access, and output controls without treating any one label as a legal conclusion.

Fair use is an affirmative defense. In other words, a defendant generally invokes it after a copyright owner identifies an act that may fall within an exclusive right. Section 107 lists four nonexclusive factors, including the purpose and character of the use, but the factors are not a mechanical scorecard. The Copyright Office describes the inquiry as case-specific and dependent on the circumstances. The first factor is important, but it cannot carry the entire analysis by itself.

Start by naming the use

The first question is not “is training transformative?” It is “which use of which work is being assessed?” The Office’s training report distinguishes among initial collection, creation of a curated dataset, copying that dataset for training, an initial or later training phase, distribution of model weights, retrieval-augmented generation, deployment, and the production of outputs. A single company may participate in several of these acts, but a court need not assume that they all have the same purpose or character.

Consider a work that enters a model’s development in three ways. First, a crawler downloads it and a dataset provider filters and compiles it. Second, a trainer uses the resulting collection to build a broad model that can translate, classify, or answer questions. Third, a retrieval system later fetches the work because it is relevant to a user’s query and supplies it to the model. The same work is present across the chain, but the first act concerns acquisition and curation, the second concerns statistical learning, and the third concerns a particular response. Treating them as one “training use” would hide facts that may change the first-factor analysis.

That does not mean each intermediate step should be examined in isolation from the product’s purpose. The Office also warns that creating a dataset or training a model is rarely the ultimate purpose. A court may need to look through the immediate act of copying to the function the resulting system serves. The relevant comparison is not simply “copying versus no copying,” but the relationship between the source work’s purpose and the system’s eventual use.

This approach prevents two opposite errors. One error is to say that the first-factor analysis ends at the moment a developer creates a dataset, even though the dataset is designed to power a user-facing generative product. The other is to say that every early copy inherits the most commercial or substitutional purpose of every possible downstream use, even when a distinct actor uses the material for research or a constrained analytical system. The right level of analysis depends on the relationships among the acts, actors, and uses.

The question also changes when a model is further trained. A broad initial phase may teach general language or visual patterns. A later phase may improve instruction following, align behavior, or adapt the system to a particular domain. A third party may continue training already distributed weights. The label attached to a phase does not determine its purpose.