Fair Use or Infringement? Bartz v. Anthropic and the future of AI Copyright
Ad
Judge William Alsup’s summary judgment in Bartz v. Anthropic stands out as one of the first serious judicial forays into how fair use applies to AI training. His ruling was intentionally nuanced: using books to train LLMs constituted “spectacularly transformative” fair use; scanning legitimately purchased books and discarding the physical versions was acceptable; but using pirated books, even for AI training, was clear copyright infringement.
How Anthropic Built Its “Central Library”
Anthropic compiled a vast library of copyrighted texts for training its Claude model. Some books were legally purchased; others were downloaded from pirate platforms like LibGen or PiLiMi. They then digitized prints and largely discarded the paper originals (original). This illuminated both the scale of their dataset and the variety of sourcing tactics they used.
Fair Use Analysis: What Passed and What Failed
1 - Transformative Use (Allowed): The court ruled that training AI constitutes a new use, creating novel outputs rather than just reprinting the original text. Thus, using legal sources this way was fair use.
2 - Format Shifting (Allowed): Digitizing purchased books was also treated as fair use, viewed as a space-saving format conversion, not exploitation.
3 - Pirated Copy Use (Rejected): Cropping corners by using pirated copies undermines market value and crosses the line. Even later purchasing the same titles didn’t retroactively legitimize prior infringement.
Unresolved Issues for Trial
Remaining questions include whether other internal uses of the central library (beyond training) might be infringing, and how damages will be calculated. The court also noted that if AI outputs themselves were infringing, that would open an entirely different realm of legal risk.
Broader Legal Context
Days after this, another AI-related fair use case (Kadrey v. Meta) reached a summary judgment, but under a different standard. Together, these rulings underscore how fact-dependent and complex fair use determinations are in the AI context.
Should Copyright Laws Be Strengthened for the AI Era?
Absolutely, and here’s why.
Current Law Lacks Clarity for AI
The Copyright Act wasn’t written with AI training in mind. Courts are forced to stretch its centuries-old constructs to fit these new realities and ambiguity breeds unpredictability, chilling both innovation and rights protection.
Transformative Use Alone Isn’t Enough
Alsup’s ruling rightly rewards transformative use, but it also highlights that courts must vigilantly guard against misuse of that doctrine. Without clearer guidelines, companies may toe the line or cross it, under the guise of “transformation.”
Pirated Content Must Be Severely Discouraged
The ruling underscores the importance of paying for content. But copyright mechanics need modern reinforcement; stronger tracking, stricter penalties, and clearer licensing protocols to deter misuse in training datasets.
Emerging Licensing Markets Need Legal Recognition
Authors argued that AI training licenses represent nascent markets but the court said the Copyright Act doesn't guarantee such opportunities. Laws should evolve to explicitly support licensing markets for AI use, ensuring creators benefit.
Transparency and Consent Should Be Mandatory
AI developers should be required to disclose data sources and obtain permission. This reinforces accountability and respects creators' rights.
So?
Alsup’s ruling is a thoughtful first move, but an imperfect fit in a rapidly evolving AI landscape. It's time to modernize copyright law: clarify fair use boundaries for AI training, outlaw use of pirated works emphatically, build new frameworks for AI licensing, and embed transparency and consent into training data practices.
Without this, we risk building powerful AI on shaky legal ground and shortening the creative ecosystem that makes AI possible in the first place.