AI Training Copyright Cases Show Legal Uncertainty, Not Theft

Courts Begin Weighing AI Training Against Copyright Law

More than 100 lawsuits testing whether AI companies can train models on copyrighted works without permission have produced their first two decisions—and the results reveal more ambiguity than clarity. In Kadrey v. Meta and Bartz v. Anthropic, both from a California federal district court, judges suggested that AI training itself likely qualifies as fair use under copyright law, but they diverged on critical details that will shape future litigation.

In Bartz, the court ruled that Anthropic's training of its Claude model was "spectacularly" transformative and therefore fair use, since generated outputs weren't substantially similar to ingested works. However, the judge condemned training on books pirated from shadow libraries like Library Genesis, leading to a $1.5 billion settlement—the largest in copyright history. The infringement, in this view, was the piracy, not the training.

The Kadrey decision took a different stance. Judge Chhabria found Meta's training fair use because plaintiffs couldn't retrieve more than short quotations from the AI, failing to demonstrate substantial similarity under established doctrine. Unlike the Bartz court, this judge didn't consider downloading from shadow libraries inherently unfair. More significantly, Judge Chhabria expressed regret that plaintiffs hadn't argued a "dilution theory"—the idea that AI-generated content could flood markets and devalue human creators' works even without producing substantially similar outputs.

Why it matters

The legal uncertainty benefits large AI companies and major content owners who can afford prolonged litigation, while individual creators lack clear copyright protection against AI systems trained to compete with them. Without definitive rulings or legislation, most creators cannot count on existing law to ensure compensation when their work trains models designed to replace them. The split between these two judges signals that courts will evaluate AI copyright questions on narrow, fact-intensive grounds rather than establishing bright-line rules—meaning clarity may not emerge for years.

What Copyright Law Actually Protects

Copyright law protects original expression, not ideas or information. The "substantial similarity" test determines whether copying goes too far, but U.S. courts apply this test inconsistently. Fair use doctrine allows some copying without permission when uses are sufficiently transformative—giving works new purpose or meaning—and don't harm existing or reasonably foreseeable markets.

AI training complicates these principles. Companies acquire massive datasets through web scraping, licensing, or scanning purchased books. The data trains models that generate new outputs from user prompts. Several legal scholars argue this process learns patterns rather than memorizing expression, making it non-infringing—except when models reproduce distinctive copyrighted elements like recognizable characters.

Open Questions Remain

Key uncertainties persist: whether AI companies or users bear liability for infringing outputs, whether "dilution" without substantial similarity can constitute infringement, and whether ignoring technical opt-out measures like robots.txt affects fair use analysis. Future cases where plaintiffs demonstrate AI generating clearly infringing outputs—such as reproducing large text chunks or copyrighted characters—will prove decisive.

The framing of these lawsuits reflects strategic choices: litigators sue AI companies for both training and outputs, likely because training alone might be fair use while outputs alone would make users, not companies, directly liable. Proving companies contributed to user infringement remains difficult.

These details were first reported by The Bulletin of the Atomic Scientists in an analysis examining the intersection of copyright law, AI technology, and the political dimensions of creative work compensation.