Is It Legal to Train AI Models on Copyrighted Books and Articles?

An Open Legal Question, Not a Settled Answer

There isn’t currently a single, clean “yes” or “no” answer to whether training AI models on copyrighted books and articles is legal. Instead, this is one of the most actively contested legal questions in the entire AI industry, playing out through numerous lawsuits filed by authors, publishers, news organizations, and other rights holders against AI developers. The legal system is still working through how existing copyright doctrine — much of it developed long before large-scale machine learning existed — applies to the practice of using vast quantities of text to train a model.

Anyone looking for a definitive, universal answer right now is likely to be disappointed, because the honest state of affairs is ongoing litigation and genuine disagreement among courts, rights holders, and AI companies about where the law should land.

The Core Argument on Each Side

AI companies training on copyrighted material generally lean on the doctrine of fair use, a long-standing part of copyright law that permits certain uses of copyrighted work without permission under specific circumstances — commonly considering factors like whether the use is transformative, the nature of the copyrighted work, how much of it was used, and the effect on the market for the original. AI developers typically argue that training is transformative: the model isn’t republishing the book or article for readers to consume in place of the original, but rather analyzing statistical patterns across huge datasets to build a general-purpose system, which they argue is a fundamentally different purpose from the original work’s.

Authors, publishers, and news organizations pushing back generally argue the opposite on several of those same factors: that copyrighted works were copied in their entirety as part of the training process, that this was done without permission or compensation, and that AI systems trained this way can go on to generate output that competes with or substitutes for the very content used to train them — potentially harming the market for original works, which is one of the specific factors fair use analysis considers. Some rights holders have also raised concerns about how training data was obtained in the first place, including whether copyrighted works were used from sources without proper licensing.

Why the Outcome Is Genuinely Uncertain

This dispute is difficult to resolve cleanly because it doesn’t map neatly onto earlier fair use precedents, which were developed around more familiar situations like search engine indexing, parody, or academic quotation. Training a large language model is a distinct technical process, and courts are having to apply decades-old legal tests to a use case that didn’t exist when those tests were formulated. Different courts examining different sets of facts — different AI companies, different datasets, different alleged uses of the resulting models — have reached varying conclusions and are still working through appeals and related cases, meaning there is no single ruling that has settled the question industry-wide.

Adding another layer, some AI companies have separately chosen to pursue licensing agreements directly with publishers and content owners, paying for the right to use their material for training. This suggests that even amid legal uncertainty, parts of the industry see commercial and legal value in securing permission rather than relying solely on a fair use defense.

Bottom Line

Whether training AI models on copyrighted books and articles is legal remains a genuinely open, actively litigated question rather than a settled matter, with fair use arguments cutting both ways depending on the specific facts of each case. Anyone directly affected — as a rights holder or an AI developer — should follow current case developments closely and consult legal counsel rather than assume a fixed answer.

Go deeper

• What Major Lawsuits Have Been Filed Over AI Art and Copyright?
• Can You Use AI-Generated Images Commercially Without Legal Risk?
• Should AI-Assisted Books Be Labeled for Readers?