District Court Rules AI Training Can Be Fair Use in Bartz v. Anthropic
Summary
On June 23, 2025, Senior US District Judge William Alsup for the Northern District of California issued an order in Bartz v. Anthropic, granting partial summary judgment in favor of Anthropic. The opinion addresses the legal doctrine of fair use, which permits unlicensed use of copyrighted material for certain purposes, for three scenarios:
- Training large language models (LLMs) with copyrighted material;
- Creating digital copies of purchased, printed copyrighted material, before destroying the printed material; and
- Downloading unauthorized copies of copyrighted material that would later be used for fair use.
Plaintiffs alleged that Anthropic infringed their copyrights by downloading pirated copies of their books and scanning purchased copies to build a centralized “research library” that would be used to train its LLM, Claude.
The court held that using copyrighted material to train LLM models was “exceedingly transformative” and thus was fair use, granting summary judgment to Anthropic on this issue. The court emphasized that the training process used works to generate new text and did not result in distributing copies to the public. Further, the court likened the training process to how people read copyrighted material to learn and improve their own writing—which has long been held as outside of the Copyright Act.
Notably, the court distinguished the facts in Thomson Reuters v. Ross Intelligence. In that case, there was no fair use because the AI tool was a competing system for finding existing legal opinions and “not generative AI.” Here, LLMs create “fresh” writing extending beyond “anything that any copyright owner rightly could expect to control.”
Next, the court held that Anthropic’s practice of destructively scanning legally purchased print books to create digital, searchable replacements was fair use and similarly granted summary judgment in Anthropic’s favor. The court noted that established caselaw views such processes as transformative—the format change eased storage and searchability without creating additional copies or distributing them externally.
However, the court found that Anthropic’s downloading of over seven million pirated books for its permanent internal library was not fair use. The court emphasized that there is no workaround in the Copyright Act that allows AI companies to avoid paying for material that they could lawfully obtain, even if the works were later used for a transformative use such as training an LLM. This issue remains live and may proceed to trial.
