When Market Effects Become the Boundary
The Copyright Office’s analysis of generative AI does not stop at the developer’s purpose. It asks what kind of work was copied, how much was taken, what the system makes accessible, and whether the resulting use harms a market protected by copyright. This chapter connects those questions to the fourth fair-use factor, so a reader can distinguish direct substitution from broader market dilution and from a licensing opportunity without treating every form of competition as infringement.
The central difficulty is that generative AI can create harm without producing a recognizable duplicate, while a technically large act of copying may expose little or none of the source expression to the public. The Office therefore treats the nature of the work, the amount copied, the purpose, public access, and market evidence as linked variables. None should be read as a universal rule in isolation.
The kind and amount of work matter
The second fair-use factor considers the nature of the copyrighted work. Copyright’s protection is closest to its core when a work contains creative expression. Novels, films, visual art, and music generally present a different factor-two case from factual or functional material such as a database of facts or computer code. Publication can modestly support fair use, while the use of unpublished material generally weighs against it because first publication is itself an important interest.
This distinction does not mean that a training dataset contains only one type of material. Language models may be trained on novels alongside scholarly articles, legal documents, news, and code. Image datasets may contain aesthetic works as well as ordinary photographs. The factor depends on the works relevant to the claim. Calling all of them “data” does not turn expressive choices into unprotected facts.
The third factor asks about the amount and substantiality of the portion used in relation to the copyrighted work as a whole. “Substantiality” includes quality as well as quantity: a small portion can matter if it captures the heart of a work. In ordinary training pipelines, developers often download and process all or substantially all of a work. That wholesale taking ordinarily weighs against fair use.
The amount is not evaluated by counting bytes alone. The question is whether the extent of copying was reasonable in relation to the purpose. Courts have sometimes found complete copying justified where it was necessary to create a search engine, a visual index, or a plagiarism-detection tool. In those settings, the copied material enabled a function that provided information about the works or accessed an unprotected feature, and the public could not use the system to obtain a competing substitute for the original.
The Copyright Office finds the analogy less straightforward for generative AI. A foundation model is not limited to telling users where a work contains a word or showing a small thumbnail. It may produce new expressive material, and a targeted training project may not require the same amount of every work as internet-scale pre-training. At the same time, the Office recognizes that large amounts of complete works may be practically necessary for some current forms of generalization. The evidence about necessity is contested, so “the model needed scale” cannot be treated as either a complete justification or a complete rebuttal.