U.S. judges are poised to test whether training data is fair use, as publishers, artists and tech giants brace for rulings that could reshape the generative-AI economy.

In the first full workweek of the new year, the U.S. court system is becoming the main arena for a question that has hovered over generative AI since it burst into public view: can a company copy vast amounts of copyrighted material to train an AI model without permission, and then sell the resulting product?

For technology companies, the answer is existential. For publishers, artists and entertainment firms, it is about leverage — and, increasingly, survival. And for judges, it is a test of whether decades-old copyright doctrine can stretch to fit a technology that learns by ingesting the cultural record.

At the center of the fight is “fair use,” a legal safety valve meant to protect beneficial, transformative uses of copyrighted works. The doctrine has long covered things like search, quotation, parody and certain forms of research. AI developers argue training is similar: the models, they say, do not store books or articles the way a pirate site does, but analyze patterns to generate new text or images. Plaintiffs say the analogy breaks down because the copying is industrial-scale, commercially motivated and foundational to products that directly compete with the original works.

Courts have already offered conflicting signals, leaving a patchwork of early rulings that both sides are eager to interpret as momentum. Some judges have suggested that training can be transformative in the way courts previously found search engines and digitization projects to be. Others have voiced skepticism, warning that generative tools may do more than index knowledge — they can compete in the same markets as the works they learned from.

That tension is now converging on a pivotal year, with major cases in federal courts from New York to California moving from procedural skirmishes into the phase where judges confront the heart of the issue: what, exactly, is the “use” in AI training — and who bears the economic consequences?

Among the highest-profile disputes are those brought by news organizations that claim their reporting was copied to build products that can produce news-like answers on demand. They argue that the technology threatens audience traffic and subscription revenue by delivering summaries and responses that users might accept instead of clicking through to an article.

In closely watched cases, media plaintiffs have accused AI firms of using their articles without permission and then generating outputs that, at times, resemble or reproduce the underlying work. Judges have allowed core copyright claims to proceed even as some peripheral allegations have been narrowed, keeping the spotlight on whether training and output behavior combine to form infringement.

The tech industry’s defense leans heavily on a familiar theme: innovation depends on learning from what came before. Companies point to the long history of reading as a non-infringing act, arguing that training is a form of analysis — the machine equivalent of a human studying books to write a new essay. Plaintiffs counter that the training process requires making complete copies, often at massive scale, for commercial systems built to generate substitutive outputs.

Beyond journalism, authors have pressed similar arguments, saying their novels and nonfiction were swept into training datasets without consent. Many writers do not object to AI-assisted tools in principle; their core complaint is that the models were built atop uncompensated copying and may flood the market with low-cost substitutes.

Visual artists have raised parallel claims, emphasizing that image generators can mimic recognizable styles and aesthetics. In these cases, the debate often turns on whether a style is protectable and whether the models’ training involves making and retaining copies in a way that triggers copyright liability.

The entertainment industry has also sharpened its posture. Studios and music companies are wary of systems that can imitate characters, voices and production aesthetics, arguing that even when a single output is not an obvious copy, the cumulative effect is to erode the scarcity that underpins creative markets.

Tech companies reply that most outputs are not copies and that the law should not grant creators a veto over how others learn and create. They warn that requiring licenses for training would be unworkable, citing the vast universe of copyrighted material, complex rights chains and prohibitive costs, particularly for smaller firms and open-source projects.

While lawsuits dominate headlines, a quieter shift has accelerated behind the scenes: licensing. Some AI companies have pursued deals with publishers, image libraries and music firms to secure access to high-quality content and reduce legal risk, offering rights-holders an alternative to years of uncertain litigation.

Licensing, however, exposes a new fault line. If only the largest AI developers can afford comprehensive content deals, smaller competitors may be squeezed out. Conversely, if courts broadly bless training as fair use, creators fear the market will shift from negotiated compensation to uncompensated extraction.

Judges weighing these disputes are expected to focus on practical realities as much as abstract doctrine: how datasets are assembled, whether materials were lawfully obtained, what safeguards exist to prevent verbatim regurgitation, and whether model outputs siphon value from the markets copyright law was designed to protect.

However the courts rule, the outcomes are unlikely to be simple. Judges may distinguish between lawful and unlawful sources, separate training from outputs, or treat different media differently based on how substitution operates in each market.

For now, both sides are bracing for a year in which judges — not engineers — may define the operating rules of generative AI. As the cases advance toward decisions on the merits, a central question looms: in an economy where machines learn from culture at scale, who gets to decide what that learning costs?