What Is 'Fair Use' and How Does It Apply to AI Training Data?

A Legal Balancing Test, Not a Simple Rule

Fair use is a doctrine within US copyright law that allows certain uses of copyrighted material without needing permission from the rights holder, under specific circumstances. It’s not a blanket exception, and it doesn’t work like a fixed checklist where meeting one condition automatically qualifies a use. Instead, fair use operates as a balancing test: courts weigh several factors together, and the outcome depends heavily on the specific facts of each situation. This flexibility is intentional — fair use was designed to adapt to new kinds of uses that lawmakers couldn’t have anticipated in advance, which is exactly why it’s become central to the debate over AI training data.

Because fair use is decided case by case rather than through a fixed formula, it produces genuine uncertainty. Two AI companies using seemingly similar training practices could, in principle, see different legal outcomes depending on details like what data they used, how they obtained it, and how their models are ultimately deployed.

The Factors Courts Weigh, and How They Map to AI

Fair use analysis typically considers four factors: the purpose and character of the use, including whether it’s commercial and whether it’s “transformative”; the nature of the copyrighted work itself; the amount and substantiality of the portion used; and the effect of the use on the potential market for the original work.

AI companies training on copyrighted material generally lean hardest on the first factor, arguing that training a model is transformative — the goal isn’t to republish or compete with the original book or article as a substitute for reading it, but to extract statistical patterns that help build a general-purpose system serving an entirely different function. This transformative-use argument has succeeded in other technology contexts in the past, such as certain search engine and indexing cases, which is part of why AI companies find it an appealing framework.

Rights holders challenge this on multiple fronts. They point out that AI training typically involves copying entire works, not small excerpts, which cuts against the “amount used” factor. They also argue that AI-generated outputs can end up substituting for the original works in the marketplace — an AI system that can write in an author’s style, or summarize a paywalled article’s content, could reduce demand for the original in ways that weigh against fair use under the “market effect” factor, often considered one of the most influential factors in the overall analysis.

Why This Debate Doesn’t Have One Universal Answer

Because fair use is fact-specific, the same general question — “is AI training fair use?” — can have different answers depending on which AI company, which training dataset, which specific copyrighted works, and which resulting AI outputs are being examined. A model trained on a narrow, licensed dataset used strictly for internal research might present a very different fair use analysis than a model trained on large amounts of scraped copyrighted journalism and later offered as a commercial product that can reproduce similar content. This is exactly why multiple separate lawsuits against different AI companies are working through courts rather than a single case settling the matter for the entire industry at once.

Bottom Line

Fair use is a flexible, multi-factor legal doctrine — not a simple yes/no rule — and while AI companies commonly argue that training on copyrighted material is transformative and therefore fair use, rights holders push back on grounds like wholesale copying and market harm. Courts are actively working through these arguments case by case, so there is no single settled answer covering all AI training practices.

Go deeper

• What Major Lawsuits Have Been Filed Over AI Art and Copyright?
• Can AI-Generated Images Be Copyrighted?
• Can You Use AI-Generated Images Commercially Without Legal Risk?