Copyright Law in 2025

Taken together, these cases establish early guideposts for AI developers, publishers, media companies, and enterprises deploying generative AI systems. Below, we summarize the most important copyright decisions and pending cases shaping the law in 2025.

Bartz et al. v. Anthropic PBC (N.D. Cal.)

Fair Use for Lawfully Acquired Training Data – With a Hard Stop on Piracy

In the first summary judgment ruling squarely addressing AI training and fair use, the Northern District of California held that Anthropic's use of lawfully acquired books to train its large language models (LLMs) constituted fair use under Section 107 of the Copyright Act. The court emphasized the highly transformative nature of LLM training, reasoning that the models do not function as repositories of expressive content but instead extract statistical relationships to enable new capabilities.

The court rejected plaintiffs' arguments that AI training inherently creates competing works that substitute for the originals – one of the four key factors to the fair use analysis. Critically, however, the court drew a firm line regarding unlawfully acquired content. It allowed infringement claims to proceed based on Anthropic's retention of pirated books in a permanent internal library, even where those works were no longer being used for training.

Practical Takeaway: Fair use may protect AI training on lawfully obtained data, but companies face significant risk where datasets include pirated or improperly sourced material.

Kadrey et al. v. Meta Platforms Inc. (N.D. Cal.)

Divergence on Market Harm, Convergence on Piracy Risk

Shortly after the Anthropic decision, another Northern District of California court granted summary judgment to Meta on fair use, while sharply criticizing the authors' market-harm arguments. The court agreed that LLM training is transformative but emphasized that market harm remains the most important fair use factor and cannot be dismissed by analogy alone.

Unlike the Anthropic court, this judge rejected comparisons to human learning, reasoning that AI systems can generate large volumes of content at scale and therefore raise distinct competitive concerns. Even so, the court found that the plaintiffs failed to present concrete, non-speculative evidence of market displacement. As in Anthropic's, claims alleging Meta obtained training data through piracy were allowed to proceed.

Practical Takeaway: Courts agree on the importance of lawful data sourcing, but they may diverge on how rigorously to scrutinize alleged market harm from AI outputs.

Andersen v. Stability AI Ltd. (N.D. Cal.)

Training Is Not the Same as Output

Although earlier in procedural posture, Andersen remains one of the most influential generative-AI copyright cases. Visual artists allege that image-generation models were trained on copyrighted artwork and that the resulting outputs constitute infringing derivative works.

Courts addressing these claims have consistently emphasized a critical distinction between training processes and outputs. General allegations that models "store," "compress," or "memorize" copyrighted works have been met with skepticism. Instead, courts require plaintiffs to plausibly allege that specific AI-generated images are substantially similar to identifiable copyrighted works.

Practical Takeaway: Plaintiffs face a high bar when alleging that AI training alone constitutes infringement absent demonstrable copying in outputs.

Tremblay v. OpenAI, Inc.

Pleading Standards Matter in AI Copyright Cases

In Tremblay, book authors alleged that OpenAI infringed their copyrights by training LLMs on their works and enabling the generation of competing content. Courts considering these claims in 2025 reinforced traditional pleading standards, rejecting conclusory assertions of copying and market harm.

Practical Takeaway: AI copyright claims that rely on generalized fears of competition or unsupported assumptions about training data are unlikely to survive early scrutiny.

The New York Times Co. v. OpenAI/Microsoft

A Harder Case for Fair Use

Although unresolved on the merits in 2025, this case looms large. Unlike many book-author suits, it directly implicates news content, real-time markets, and established licensing regimes. The Times alleges that AI outputs substitute for journalism and undermine its ability to monetize content through subscriptions and licenses.

Practical Takeaway: Fair use arguments may be weaker where AI outputs directly compete with copyrighted content in active licensing markets.

Getty Images v. Stability AI (U.K. and U.S.)

International and Evidentiary Implications

Getty's litigation – particularly in the United Kingdom – raises additional issues that U.S. courts may increasingly confront, including database rights, cross-border training, and evidence of verbatim reproduction such as retained watermarks.

Practical Takeaway: Multinational AI developers must consider non-U.S. copyright and database regimes in dataset design and deployment.

Music Industry Cases (Sony, UMG, and Others)

Sound-Alikes, Voice Models, and New Theories of Harm

In 2024–2025, major music labels filed suits challenging AI models that generate sound-alike vocals and music. These cases push beyond traditional training-data disputes and raise questions about derivative works, right of publicity, and unfair competition. They signal that courts may confront AI copyright issues differently where models replicate distinctive human performance attributes rather than abstract expressive patterns.

Eddie Richardson v. Karim Kharbouch (7th Cir.)

Reinforcing Evidentiary Rigor

Although unrelated to AI, the Seventh Circuit's decision requiring proof of actual sampling – rather than mere similarity – to establish sound-recording infringement reinforces a theme common to AI cases: copyright liability requires evidence, not inference.

Conclusion and Recommendations

The 2025 decisions reflect cautious but meaningful progress in defining how copyright law applies to generative AI. Courts are increasingly receptive to fair use arguments for training on lawfully acquired data, deeply skeptical of speculative market-harm claims, and uniformly intolerant of piracy. At the same time, cases involving direct competition, news content, and human likeness may test the limits of these early rulings.

For questions about how these cases could reshape copyright law, please reach out to Edward D. Lanquist and Benjamin West Janke.