Baker Donelson lawyers summarize 2025 AI-training copyright decisions

What happened

On 10 February 2026, attorneys Edward D. Lanquist, Jr. and Benjamin West Janke of the law firm Baker, Donelson, Bearman, Caldwell & Berkowitz published a review of 2025 U.S. copyright decisions on AI training. On Bartz v. Anthropic, they report that the Northern District of California held on summary judgment that training large language models on lawfully acquired books was fair use, while allowing infringement claims to proceed over Anthropic's retention of pirated books in a permanent internal library. The review also discusses Tremblay v. OpenAI (pleading standards for author claims against OpenAI) and The New York Times Co. v. OpenAI/Microsoft, noting the Times case may be a harder case for fair use because it involves news content, real-time markets and established licensing regimes.

Why it may matter to you

As a statement from identified lawyers about specific cases, this is a useful map of where courts drew lines in 2025: training on lawfully obtained data has been treated as transformative fair use in early rulings, while pirated sourcing was uniformly treated as carrying infringement risk. The authors also flag that cases involving direct market competition (news) or human likeness (music/voice) may test the limits of those early rulings.

What is not yet known

These are district-court decisions, not appellate precedent, and the authors note the rulings do not settle all open questions — including emerging judicial disagreement over how to assess market harm from AI outputs. The New York Times case remained unresolved on the merits at the time of writing.