“Training AI models on copyrighted works without permission is fair use.”
This is a question of US law (other countries handle it differently, through text-and-data-mining exceptions), but it determines much of the economics of AI. Fair use turns on four factors, and courts so far have split along the most important ones: whether training is "transformative", and whether AI outputs harm the market for the originals.
Two 2025 rulings in the Northern District of California found training on lawfully acquired books to be fair use, while treating pirated copies differently; one resulted in a $1.5 billion settlement over pirated books. Other courts have ruled against non-generative uses, and dozens of cases, including the New York Times against OpenAI, are pending. Note: Crux's content was drafted with help from an AI developed by a company that was a defendant in one of these cases.
The US doctrine (17 U.S.C. §107) allowing some unlicensed use, judged on four factors: purpose and character (including transformativeness), nature of the work, amount used, and effect on the market.
Copying works into datasets and computing model parameters from them. Separate questions arise about how the copies were obtained and about outputs that reproduce works.
Training is highly transformative. The model learns statistical patterns of language, not the works themselves, and uses them for a different purpose. Courts have repeatedly protected large-scale copying for non-expressive analysis.
In Bartz v. Anthropic (June 2025), Judge Alsup held that using lawfully acquired books to train language models was "exceedingly transformative" and fair use, comparing it to a reader learning to write. In Authors Guild v. Google (2d Cir. 2015), scanning millions of books for search was fair use.
Bartz v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal. June 23, 2025); Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015).
Limitation
The same Bartz ruling found that downloading pirated books was not excused; the case then settled for $1.5 billion. Google Books showed snippets, not generated substitutes.
Requiring permission for every work would make training infeasible except for the largest companies able to sign licensing deals, entrenching incumbents and making models worse on underrepresented material.
Japan's copyright law (Art. 30-4) allows using works for data analysis, including AI training, and the EU's 2019 copyright directive allows text-and-data mining unless rights holders opt out.
Japan Copyright Act Art. 30-4; Directive (EU) 2019/790, Arts. 3–4.
Limitation
Licensing markets are developing quickly, which undercuts the infeasibility argument and strengthens the market-harm factor for plaintiffs.
Plaintiffs struggle to show that training itself harms the market for their works. Courts have asked for concrete evidence of lost sales or licensing, not speculation.
In Kadrey v. Meta (June 2025), Judge Chhabria granted summary judgment to Meta because the authors had not shown market harm, while stressing the ruling turned on their evidence, not a general holding that training is fair use.
Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417 (N.D. Cal. June 25, 2025).
Limitation
Chhabria's opinion said market dilution from floods of AI-generated substitutes could weigh heavily against fair use in a better-argued case.
When the output competes directly with the inputs, the use is not transformative in the relevant sense. Generative models trained on news, fiction or art produce substitutes for those works.
In Thomson Reuters v. Ross (Feb. 2025), a court held that copying Westlaw headnotes to train a competing legal-research AI was not fair use because it served the same market. The Supreme Court in Warhol v. Goldsmith (2023) narrowed transformativeness when uses share the original's purpose.
Thomson Reuters v. Ross Intelligence, No. 1:20-cv-613 (D. Del. Feb. 11, 2025); Andy Warhol Foundation v. Goldsmith, 598 U.S. 508 (2023).
Limitation
Ross was a non-generative system directly competing with the plaintiff. General-purpose models are further from any one plaintiff's market.
Models can memorise and reproduce training material, which shows the works themselves are embedded in them, not just abstract patterns.
A Munich court ruled in November 2025 (GEMA v. OpenAI) that memorised song lyrics in a model infringed copyright. The New York Times' complaint against OpenAI exhibits outputs reproducing long passages of its articles.
LG München I, GEMA v. OpenAI (11 Nov. 2025); The New York Times Co. v. Microsoft Corp. and OpenAI, No. 1:23-cv-11195 (S.D.N.Y., filed Dec. 2023).
Limitation
Memorisation affects a small share of training data and can be reduced. It bears more on outputs than on whether training itself is fair use.
Fair use was not designed for commercial copying of nearly all human creative output to build products worth hundreds of billions of dollars. Creators should share in that value or be able to refuse.
The US Copyright Office's 2025 report on generative AI training concluded that some training uses would likely not be fair use, especially where outputs compete with the originals and licensing markets exist.
U.S. Copyright Office (May 2025). Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version).
Limitation
This is an argument about what the law should be; courts apply the four factors as they stand.
Acquiring works from pirate sites is treated differently from training on lawfully obtained copies.
Outputs that reproduce protected works can infringe regardless of whether training was fair use.
There is no appellate ruling on generative AI training yet; the law is unsettled.