fair use and AI training
In plain wordsThe unsettled US legal question of whether AI companies may copy books, articles, pictures and music to train their models without permission, under the rule that allows some unlicensed uses of copyrighted work.
The application of the US copyright doctrine of fair use (17 U.S.C. § 107) to copying works into training datatraining data, n.: The examples an AI system learns from. Whatever patterns, skills, gaps and biases are in the training data tend to show up in the finished system.. Courts weigh the statute's four factors: the 'purpose and character of the use' (above all whether it is transformative, and whether it is commercial), what kind of work was copied, how much of it was taken, and how far the use harms the work's actual or potential market. Developers argue that training extracts statistical patterns rather than expression, citing Authors Guild v. Google (2015), which allowed book scanning for search; rights holders stress wholesale copying, pirated sources, regurgitated output and lost licensing markets, citing the Supreme Court's narrowing of transformative use in Warhol v. Goldsmith (2023). In June 2025 two federal judges in California found training on books to be fair use — in Bartz v. AnthropicBartz v. Anthropic, prop. n.: A US lawsuit by book authors against the AI company Anthropic that ended in a record $1.5 billion settlement, after a judge ruled that training on purchased books was legal but keeping pirated copies was not. (though not the retention of pirated copies) and Kadrey v. Meta (on a thin record, while warning that market dilution could defeat the defense) — whereas in Thomson Reuters v. Ross, a non-generative case, fair use was rejected, and in September 2026 the Third Circuit, in an opinion issued under seal, affirmed that result in the first US appellate ruling on fair use for AI training. The US Copyright Office's 2025 report concluded that some training uses are fair and others are not.
Their whole defense rests on fair use and AI training being transformative, so the market-harm evidence will decide it.
Origin
Fair use, rooted in Folsom v. Marsh (1841), was codified in the Copyright Act of 1976; its application to machine learning was debated from the late 2010s and became urgent with lawsuits against generative AI developers from late 2022.
Usage note
Fair use is a US doctrine; most other jurisdictions rely on narrower statutory exceptions such as the EU's text and data mining exceptiontext and data mining exception, n.: A rule in many countries' copyright laws that lets people use computers to copy and analyze large amounts of text, images or data without asking permission — sometimes only for research, sometimes unless owners opt out.. Courts increasingly treat the lawfulness of acquiring training copies separately from the lawfulness of training itself.
Related
- AI copyright litigationAI copyright litigation, n.: The many lawsuits in which authors, artists, news organizations, record labels and others accuse AI companies of copying their work without permission to build or run AI models.
- text and data mining exceptiontext and data mining exception, n.: A rule in many countries' copyright laws that lets people use computers to copy and analyze large amounts of text, images or data without asking permission — sometimes only for research, sometimes unless owners opt out.
- Bartz v. AnthropicBartz v. Anthropic, prop. n.: A US lawsuit by book authors against the AI company Anthropic that ended in a record $1.5 billion settlement, after a judge ruled that training on purchased books was legal but keeping pirated copies was not.
- New York Times v. OpenAINew York Times v. OpenAI, prop. n.: A major lawsuit, filed in 2023, in which The New York Times accuses OpenAI and Microsoft of copying millions of its articles to train ChatGPT and of letting the chatbot reproduce its journalism.
- training datatraining data, n.: The examples an AI system learns from. Whatever patterns, skills, gaps and biases are in the training data tend to show up in the finished system.
- memorizationmemorization, n.: When an AI model stores specific pieces of its training data — a paragraph, a phone number, an image — closely enough to reproduce them, instead of only learning general patterns.
- data licensingdata licensing, n.: The legal permissions that say who may use a dataset or a body of content, for what purposes and on what terms — including whether it may be used to train AI.
- text and data mining exceptiontext and data mining exception, n.: A rule in many countries' copyright laws that lets people use computers to copy and analyze large amounts of text, images or data without asking permission — sometimes only for research, sometimes unless owners opt out.
Sources
- US Copyright Office (2025). Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version, May 2025).
- IPWatchdog (2026). Third Circuit Affirms Revised Fair Use Ruling Against ROSS' AI Legal Research Platform in Sealed Opinion. September 30, 2026. ipwatchdog.com/2026/09/30/third-circuit-affirms-revised-fair-use-ruling-against-ross-ai-legal-research-platform-in-sealed-opinion
Mentioned in
AI artAI art, n.: Images, music or other artworks made with the help of AI — from early computer drawing programs to today's prompt-based image generators — and the debate over whether and how such work counts as art., AI copyright litigationAI copyright litigation, n.: The many lawsuits in which authors, artists, news organizations, record labels and others accuse AI companies of copying their work without permission to build or run AI models., training datatraining data, n.: The examples an AI system learns from. Whatever patterns, skills, gaps and biases are in the training data tend to show up in the finished system., AI authorshipAI authorship, n.: The question of whether a work or invention produced by AI can be protected by copyright or patent, and if so who counts as its author or inventor: the AI, the person using it, or no one., data licensingdata licensing, n.: The legal permissions that say who may use a dataset or a body of content, for what purposes and on what terms — including whether it may be used to train AI., opt-outopt-out, n.: A way for creators, websites or individuals to say 'do not use my work or data to train AI' — instead of being asked for permission first, which would be opt-in., text and data mining exceptiontext and data mining exception, n.: A rule in many countries' copyright laws that lets people use computers to copy and analyze large amounts of text, images or data without asking permission — sometimes only for research, sometimes unless owners opt out., web scrapingweb scraping, n.: Using software to automatically copy information from websites — text, prices, images — in bulk, which is how much of the data used to train AI was gathered., Bartz v. AnthropicBartz v. Anthropic, prop. n.: A US lawsuit by book authors against the AI company Anthropic that ended in a record $1.5 billion settlement, after a judge ruled that training on purchased books was legal but keeping pirated copies was not., Books3Books3, prop. n.: A collection of nearly 200,000 pirated e-books, assembled in 2020 and used to train several AI language models, which became central to lawsuits by authors over AI training., New York Times v. OpenAINew York Times v. OpenAI, prop. n.: A major lawsuit, filed in 2023, in which The New York Times accuses OpenAI and Microsoft of copying millions of its articles to train ChatGPT and of letting the chatbot reproduce its journalism., vibe portingvibe porting, n.: Using AI coding agents to translate a whole software project into another programming language, usually checking the result against the original project's tests.