AI training on copyrighted books: what courts say now

Training an AI model on copyrighted books is not automatically legal or illegal in the United States. Recent cases make the result depend on how the works were acquired, what the model is built to do, whether the use is transformative, and what evidence exists about market harm, according to TechCrunch's review of the dispute.

Definition: AI-training copyright disputes ask whether copying and processing protected works for model development qualifies as fair use under U.S. copyright law.

Example: A court treated Anthropic's training use of lawfully acquired books differently from the company's retention of books obtained from shadow libraries.

Key takeaway: A favorable ruling for one training setup does not create a blanket safe harbor for every dataset or model.

Business impact: AI teams should treat data provenance, model purpose and market-substitution risk as design inputs before training, not as cleanup work after launch.

What does U.S. fair use ask about AI training?

U.S. fair use is a fact-specific defense built around four statutory factors: the purpose and character of the use, the nature of the copyrighted work, the amount used, and the effect on the work's market. The U.S. Copyright Office's AI initiative is still analyzing how those existing rules apply to training, so a company should not treat the phrase "fair use" as a categorical permission slip.

For AI training, the first and fourth factors are doing much of the visible work in current cases. A use looks more defensible when the model serves a meaningfully different purpose from the source and does not act as a substitute for the source. A use looks riskier when the model is built to compete with the source's market or when the record contains credible evidence of lost or threatened demand.

Why did Anthropic's book-training ruling help AI developers?

In Bartz v. Anthropic, Judge William Alsup ruled that using copyrighted books to train Anthropic's language models was lawful under the fair-use analysis described in the case, while separating that conclusion from Anthropic's acquisition of pirated copies. TechCrunch reported that the court later became associated with a $1.5 billion settlement concerning the authors' claims, but the core training ruling was more favorable to AI development than the headline-sized settlement suggests.

The distinction matters because the court treated model training as different from distributing a book to a reader. The court's reasoning, summarized by White & Case's case analysis, treated training as a transformative use while leaving potential liability for copying and retaining massive pirated libraries. For an AI operator, lawful acquisition and training purpose are separate questions that must both be documented.

Why did Ross Intelligence reach the opposite result?

In Thomson Reuters v. Ross Intelligence, Ross used headnote-derived material to train a competing legal-research tool, and the court rejected Ross's fair-use defense. The district court opinion treated the commercial competition and the potential market for AI-training data as important reasons the use was not sufficiently transformative.

Ross is a useful boundary case because the dispute was not simply about whether a model learned from text. The court viewed the training material as helping create a product that competed with the copyright owner's legal-research business. That fact changes the operator's risk analysis: a model that produces a substitute for the source's commercial function is harder to defend than a model whose use has a clearly different purpose.

What does the Meta case add to the picture?

In Kadrey v. Meta, the court also found Meta's use of books to train Llama fair use on the record before it, but the reasoning left a warning for future plaintiffs and defendants. The Ohio State University 2026 update explains that the court considered whether AI could create an indirect market substitute by enabling a large volume of competing works, even though the plaintiffs had not supplied enough evidence to win on that theory.

That split shows why the current cases are not a simple scorecard for either authors or AI companies. One court may find the training purpose transformative, while another part of the same analysis may still turn on market dilution, output similarity or the quality of economic evidence. Companies should preserve evidence about those questions instead of relying on a single favorable sentence from a ruling.

What remains unresolved for AI companies?

The U.S. Copyright Office has released a pre-publication Part 3 of its AI report on generative-AI training, but that document is an analysis of policy and law, not a universal court-issued license. The courts are still applying an older fair-use framework to new systems, and different records can produce different outcomes. That is why the AI training landscape remains favorable in some cases but carries material caveats.

The unresolved questions include how much weight courts should give to a potential licensing market, how to measure indirect substitution by AI-generated works, how output memorization affects the analysis, and how acquisition records should influence the result. Those questions matter to businesses even when they do not train foundation models themselves: a vendor's dataset decisions can become part of a customer's legal and procurement risk.

What should business operators do now?

Business operators evaluating a model or vendor should ask for a clear account of dataset provenance, licensing and filtering rather than accepting a generic statement that the model was trained on publicly available material. "Publicly accessible" does not answer whether a work was copyrighted, lawfully acquired, copied from a shadow library or used in a product that competes with the original market.

The operational checklist is concrete: record the source and permission status of training data, quarantine uncertain material, define the intended output and comparable markets, test for memorization or close reproduction, and route unresolved questions to qualified counsel. The same source-control discipline used for AI agents and their tool permissions applies here: the model is only one part of a system whose data paths and consequences need to be observable.

AI training on copyrighted books is therefore a moving legal-risk question, not a settled yes-or-no feature. Companies can learn from the Anthropic, Meta and Ross decisions, but they should not treat any one of them as a universal rule. As more courts receive better evidence about acquisition, outputs and market effects, the safest strategy is to make the dataset and the intended use defensible before training begins.

For background on how model inputs become part of broader AI systems, see what an AI agent does beyond a chat response.