Fair use is a doctrine of United States copyright law that allows the use of a protected work without the owner’s permission when a court, weighing four statutory factors, concludes that the use is fair. It is codified in Section 107 of the Copyright Act of 1976 and operates as an affirmative defense: the user admits copying and argues that the copying is lawful.
In the debate on Copyright & AI, fair use has become the central legal question for developers in the United States, because training a model requires copying large numbers of works. Unlike the closed lists of exceptions found in EU tech regulation, fair use is open-ended and decided case by case, which makes it flexible but hard to predict.
The four factors of fair use
Courts decide fair use by weighing four factors listed in the statute, none of which is decisive alone. The factors are not a checklist with points; judges balance them in light of the purposes of copyright, which is to promote the creation of new works.
| Factor | Key question | What tends to favor fair use |
|---|---|---|
| Purpose and character | Is the use transformative or commercial | New purpose, commentary, criticism, research |
| Nature of the work | Is the work factual or highly creative | Factual or already published works |
| Amount and substantiality | How much was taken, and was it the heart | Only what the purpose requires |
| Market effect | Does the use substitute for the original or a licensing market | No substitution, no lost licensing revenue |
The first and fourth factors usually carry the most weight. The concept of a transformative use, one that adds a new purpose or meaning rather than merely repackaging the original, was elevated by the Supreme Court in Campbell v. Acuff-Rose (1994). In Andy Warhol Foundation v. Goldsmith (2023) the Court narrowed it, holding that a new meaning is not enough when the new use shares the same commercial purpose as the original.
How fair use applies to AI training
The core question in Copyright & AI disputes is whether copying works to train a model is a transformative use or a substitute that harms authors. Developers argue that training extracts statistical patterns rather than expressive content, comparable to the book scanning that the Second Circuit accepted in Authors Guild v. Google (2015). Rights holders answer that models compete directly with the works they learned from.
Early decisions point in different directions. In Thomson Reuters v. Ross Intelligence (2025), a federal court rejected fair use for copying legal headnotes to build a competing, non-generative research tool. Later in 2025, two district courts in California found that training language models on books was highly transformative on the records before them, while one of them held that building a library from pirated copies was a separate act not covered by fair use. None of these rulings binds other courts, and appeals and settlements continue to shape the landscape.
| Issue | Argument for fair use | Argument against fair use |
|---|---|---|
| Purpose of copying | Learning patterns is a new purpose | Output serves the same market as inputs |
| Amount used | Whole works needed for statistics | Entire works copied at scale |
| Source of copies | Lawfully acquired material | Pirated datasets taint the use |
| Market harm | Outputs rarely reproduce works | Market dilution and lost licensing deals |
| Output similarity | Filters prevent regurgitation | Memorized passages can be extracted |
A distinct issue is the output side. Even if training is fair use, a specific output that reproduces substantial protected expression can still infringe, and fair use for that output is analyzed separately.
Fair use compared with other copyright exceptions
Fair use differs from the exception systems of most other countries in that it is an open standard rather than a closed list. The comparison below shows why the same AI training activity can be lawful in one jurisdiction and require a different legal basis in another.
| Mechanism | Where used | Structure | Relevance to AI training |
|---|---|---|---|
| Fair use | United States, Israel, South Korea, Singapore | Open, factor-based | Decided case by case |
| Fair dealing | United Kingdom, Canada, Australia | Closed list of purposes plus fairness | Narrow; commercial training rarely fits |
| EU exceptions | European Union | Closed list in directives | Relies on the text and data mining exception |
| Text and data mining exception | European Union | Specific rule with opt-out for commercial use | Lawful unless rights are reserved |
| Three-step test | Berne Convention, TRIPS | International limit on all exceptions | Constrains every national exception |
The practical consequence is that a developer operating in Europe cannot rely on fair use at all. It must comply with the TDM opt-out regime and the training data transparency duties of EU law, regardless of how a US court would view the same copying.
Assessing a use in practice
Because fair use is decided only in litigation, organizations assess litigation risk rather than obtaining a definitive answer in advance. The decision table below reflects patterns that courts have treated favorably or unfavorably, not guarantees.
| Use | Typical risk | Main reason |
|---|---|---|
| Quoting a short passage in a review | Low | Criticism with limited amount |
| Parody of a song | Low to moderate | Transformative commentary on the original |
| Search index of full books showing snippets | Low to moderate | New purpose, no substitution |
| Training on lawfully acquired text | Uncertain | Courts still divided on market harm |
| Training on pirated datasets | High | Acquisition itself may infringe |
| Generating outputs that mimic specific works | High | Direct substitution for the original |
Good practice is to document the purpose of copying, keep records of lawful acquisition, limit what is retained, and add output filters that reduce memorized reproduction. Licensing remains the safest route where a functioning market already exists.
Common misconceptions about fair use
- Giving credit does not create fair use. Attribution is courteous but legally irrelevant to the four factors.
- Non-commercial use is not automatically fair. Commerciality is one consideration within the first factor, not a safe harbor.
- There is no fixed word or second limit. Rules such as thirty seconds of music or ten percent of a book have no basis in the statute.
- Fair use is not a permission. It is a defense raised after a claim, so legal costs arise even when the defense succeeds.
- It does not travel abroad. A use that is fair in the United States can infringe in countries with closed exception lists.
In everyday language
Copyright normally means you need permission to copy someone else’s work. Fair use is a safety valve in American law. It lets people use parts of a work without asking, as long as the use is fair overall.
A judge looks at why you used the work, what kind of work it was, how much you took and whether you hurt the owner’s sales. A film critic quoting a few lines is usually fine. Copying a whole novel to sell it is not. The key takeaway is that fair use has no simple formula. It depends on the full picture.
Seen in the real world
A US company building a writing assistant plans to train its model on about two million books. Its lawyers review the plan against the four factors before training begins.
They recommend buying physical copies or licensed ebooks instead of downloading an online collection of pirated books, because recent rulings treated pirated acquisition as a separate problem. They also advise output filters that block long verbatim passages and a licensing deal with one academic publisher whose market is clearly at stake.
The outcome is not a guarantee, but the company reduces its exposure considerably. When an author group later files a claim, the company can show lawful sourcing, a transformative training purpose and measures against substitution.