Two recent U.S. court decisions – Bartz v Anthropic (decided June 23, 2025) and Kadrey v. Meta Platforms, Inc. (decided June 25, 2025) have added important clarity – and complexity – to the global conversation around copyright and artificial intelligence. Judges in the Northern District of California ruled that using copyrighted books to train large language models (LLMs) can qualify as fair use. Fair use is a legal doctrine permitting limited use of copyrighted material without permission under certain conditions.
These rulings are among the first to directly address how copyright law applies to AI training, and they carry implications not just for U.S. companies, but for AI developers, publishers, and policymakers worldwide.
Critical Legal Distinctions Emerge from Court Analysis
Training Data Acquisition vs Model Development
Both courts separated the act of training an LLM from the earlier act of acquiring the training data. In Anthropic, Judge Alsup ruled that while training was fair use, building a library of pirated books was not. He emphasized that these pirated copies displaced legitimate demand and were retained for multiple uses.
In contrast, Judge Chhabria in Meta viewed the initial downloading of books as part of the transformative process of training LLaMA. For him, the end goal – training an AI – justified the means.
Takeaway: Courts may treat pre-training copying as fair use only if it’s tightly linked to the training itself.
Unauthorized Sources and Fair Use Defense
The judges also differed on the significance of using pirated copies. Judge Alsup was skeptical that downloading from pirate sites could ever be justified under fair use. Judge Chhabria, however, rejected the idea that using unauthorized copies automatically invalidates a fair use defense, focusing instead on market harm.
Takeaway: The legality of using pirated data may hinge less on the source and more on the impact on the market.
Market Impact Analysis Reveals Judicial Split
Economic Harm Assessment Diverges
The biggest split came over market dilution. Judge Alsup dismissed concerns that AI-generated content would flood the market, likening LLM training to teaching students to write. Judge Chhabria disagreed, warning that AI can produce vast amounts of competing content quickly and cheaply—posing a real threat to authors’ markets.
Takeaway: Future cases may turn on whether plaintiffs can prove that AI-generated content meaningfully harms the market for original works.
Strategic Implications for AI Development
These divergent interpretations create both opportunities and challenges for AI companies operating globally. The rulings suggest that fair use protection for AI training may depend heavily on specific factual circumstances, including data acquisition methods, training purposes, and demonstrable market effects.
Companies developing AI systems must now navigate a complex landscape where the same training practices might receive different legal treatment depending on jurisdiction and specific implementation details.
Preparing for Evolving Legal Standards
The legal landscape is still evolving. More cases are pending, and appeals could reshape the current understanding. Meanwhile, companies may wish to: (i) audit their training data sources; (ii) stay informed on jurisdictional differences; and (iii) engage with policymakers and industry groups.
The international implications of these U.S. decisions remain significant, as many global AI developers look to American precedent when assessing their own legal risks. However, different jurisdictions may develop distinct approaches to balancing innovation incentives with creator rights.
Need help with your IP, Technology and Commercial Strategy? Contact our team at CrossBorder IP for a hassle-free, complimentary consultation.