The Four-Factor Fair Use Test Applied to AI Training

The four factors are usually explained as a law-school checklist. That version tells you nothing about how a judge weighs them on these facts. Three US district courts applied the test to machine training in 2025 and produced three differently shaped answers, and the disagreements are more informative than the outcomes.

The statute, and what it is not

Fair use is codified at 17 U.S.C. § 107, which lists four factors a court “shall include” in its analysis. Two structural points get lost in summaries. First, the list is not exhaustive: the statute says the factors to be considered include these, so a court may weigh something else. Second, the factors are not scored and totalled. They are weighed, they interact, and in practice factors one and four carry most of the load, with the Supreme Court’s 2023 decision in Andy Warhol Foundation v Goldsmith having tightened factor one considerably by directing courts to the specific use at issue and its justification rather than to a general aesthetic of transformation.

Factor one: purpose and character

This is where the AI cases have been won and lost. The question after Warhol is whether the challenged use has a further purpose or different character from the original, and whether the copying was reasonably necessary to that purpose.

In Thomson Reuters Enterprise Centre GmbH v Ross Intelligence Inc., decided in the District of Delaware on 11 February 2025 by Judge Stephanos Bibas sitting by designation, the court granted partial summary judgment against the defendant on fair use. The critical facts were that Ross used Westlaw headnotes to build a legal search tool that competed directly with Westlaw, and that the tool was not generative — it retrieved judicial opinions rather than producing new text. Factor one favoured Thomson Reuters because the use was commercial and, on the court’s reading, not transformative: the headnotes were used for substantially the purpose for which they were made. The judge later certified the ruling for interlocutory appeal to the Third Circuit, which is the honest way to read it — a district court decision that its own author regarded as contestable.

Bartz v Anthropic, decided in the Northern District of California on 24 June 2025 by Judge William Alsup, went the other way on the training question, describing the use of books to train a large language model as highly transformative because the resulting model does something categorically different from what the books do. The distinction from Ross is visible in one line: Ross concerned a substitute for the plaintiff’s own product built from the plaintiff’s own editorial work, and Bartz concerned a general-purpose model trained on books it did not compete with.

Factor two: nature of the work

This factor asks whether the work is closer to the creative core of copyright or to the factual periphery, and whether it was published. It almost never decides an AI training case, and the decided rulings treat it briefly. Novels are creative and published, so the factor tilts slightly against the user and moves nothing. It is worth understanding precisely because arguments that lean on it are weak arguments — if a brief spends its energy here, it is because factors one and four are not going well.

The one place factor two does real work is with compilations and databases, where the protected element may be a thin layer of selection and arrangement over unprotectable facts. In Ross, a great deal of the analysis concerned whether individual headnotes were original enough to be protected at all — a copyrightability question that sits logically before fair use and disposed of part of the case on its own.

Factor three: amount and substantiality

Training copies entire works, which sounds fatal and is not. The established line of cases on machine analysis — the Google Books and HathiTrust decisions in the Second Circuit — held that copying a whole work can be reasonable where the purpose requires it and the whole work is not made available to the public. Courts applying the factor to model training have generally followed that reasoning: what matters is not how much went in but how much comes out.

Which is why regurgitation evidence is the pressure point in every one of these cases. A model that can be induced to emit long verbatim passages presents a different factor-three question, and a different factor-four question, from one that cannot. Deduplication, output filtering and memorisation testing are therefore not only engineering hygiene; they are the factual record on which this factor is argued.

Factor four: effect on the market

Factor four asks about harm to the market for the original work and for its derivatives, including licensing markets. Two competing framings are now visible in the case law, and the disagreement between them is the single most important unresolved question in US AI copyright.

• Substitution. The traditional framing: does the use displace sales of the original? On this framing a general-purpose model rarely substitutes for a specific novel, and factor four favours the developer.
• Market dilution. In Kadrey v Meta, decided on 25 June 2025, Judge Vince Chhabria granted summary judgment for the defendant as to the plaintiffs before him while stating at length that the more serious theory — that a model trained on an author’s work floods the market with competing works and depresses demand for human writing in general — could well be decisive, and that these plaintiffs had simply failed to build a record for it. That is an unusual opinion: a win for the defendant that reads as a roadmap for the next plaintiff.

Nothing has resolved which framing controls. It is a question that will take an appellate court, and until one answers it, a confident statement that training is or is not fair use is a prediction rather than a statement of law. The US Copyright Office reached its own view in the pre-publication part of its AI report dealing with training, which is worth reading as an agency’s reasoned position and not as binding authority; it and the other parts are at the Copyright Office’s AI initiative.

The factor that is not in the statute

The decided cases converge on something the four factors do not name: how the copies were obtained. Bartz separated training on lawfully purchased books, which the court treated favourably, from the retention of a library of pirated copies, which it did not — and that separation, not the training analysis, is what drove the case toward the very large settlement announced later in 2025.

The practical consequence for anyone assembling a corpus is that provenance is the part of the record you control. Where each file came from, whether access was lawful, whether a licence covered the use, and whether opt-out signals were honoured are all documentable at the time of collection and nearly impossible to reconstruct afterwards. That record is also what a European provider needs for a different reason — see the GPAI copyright policy obligation and the Article 4 text-and-data-mining opt-out, which turn on lawful access rather than on any balancing test.