Is It Legal to Train AI Models on Copyrighted Books? It's Complicated

Background and Context

A legal dispute that once seemed distant from daily life has quietly reshaped the foundational relationship between the publishing industry and the artificial intelligence sector. Most published authors have, without their knowledge or consent, indirectly become co-creators of the very AI tools that threaten to undermine their livelihoods. The words they wrote are scraped at scale, encoded, and fed into models, ultimately potentially transforming into generative products that could replace or compress their market value.

Modern large language models are typically built through massive text mining, scraping millions of books, articles, and web pages and converting them into vectors and statistical patterns the model can learn. Crucially, the text is not "copied" into a finished product readers can directly consume; instead it is broken down and recoded into internal parameter representations. One core function of copyright law is controlling the "reproduction" and "derivation" of works, yet the law offers no clear verdict on which category training actually falls into.

Deep Analysis

The true source of the dispute lies in this structural tension. Technology companies argue this constitutes transformative use, since it does not directly replace the original work but extracts linguistic patterns. Authors and publishers, meanwhile, fear the final "transformed" product precisely encroaches on their living space.

To judge fair use, jurisdictions such as the United States weigh four factors: the purpose and nature of the use, the nature of the original work, the quantity and proportion used, and the effect on the potential market of the original. The most decisive is often the last — whether the new tool forms a substantial substitute for the original's market. If AI-generated content simply stops readers from buying the original book, the fair-use defense grows fragile; if it merely offers a brand-new complementary tool, courts may lean differently.

From a commercial standpoint, the deeper contradiction is a severe mismatch between value creation and value distribution. The core competitiveness of AI companies rests on vast quantities of high-quality text, whose source is precisely the accumulated investment of authors and publishing houses over past decades. The creators of the trained text receive no direct return from the commercial value generated. When the marginal return on writing a book or long article is sharply compressed, the incentive for professional content production declines.

Industry Impact

For independent authors, the risk is especially concrete, since their income depends heavily on royalties and fees, which cheap AI-generated substitutes may directly erode. For the publishing industry, this is a battle for pricing power and scarcity, because the value of content markets is largely built on the scarcity conferred by the "human-original" label.

For leading AI companies, legal risk is no longer a theoretical question but a real source of cost and uncertainty. If courts rule in a key case that large-scale training constitutes infringement, the entire industry's model of acquiring training data could be rewritten, forcing companies toward costly paths such as paid licensing, self-built corpora, or data cleaning. This could raise industry barriers, placing resource-constrained startups at a disadvantage and further consolidating the dominance of leading players.

Authors currently sit at a clear disadvantage, since individual litigation is extremely costly while the gains from victory are hard to internalize — which is why class-action lawsuits and legislative advocacy have become focal points. For readers, the outcome directly affects the quality and diversity of content they can access, since thinning professional production raises the risk of markets being filled with low-quality generated content.

Outlook

Several signals warrant continued observation. First, precedent-setting key rulings, especially those addressing the "market effect" factor, are likely to become the watershed that defines the entire red line. Second, whether legislation introduces specific exceptions for text mining, as many countries already do for research, archiving, or specific purposes, will directly determine the industry's trajectory.

Third, whether market mechanisms form spontaneously — such as licensing platforms, revenue-sharing, or mature data-tracing technology — may reshape value distribution faster than litigation. Fourth, differences in international approaches, since different jurisdictions vary in their tolerance for fair use, may prompt AI companies to adopt different data strategies across markets.

Ultimately, "is it legal" lacks a simple answer because it is fundamentally not a black-and-white technical question but a value judgment seeking balance between encouraging technological innovation and protecting creators' rights. Until courts rule, legislators set rules, and markets provide solutions, this contest over text mining and copyright boundaries will continue evolving under uncertainty, and its final direction will profoundly shape the future form of the content ecosystem.