A Fair Use Fault Line: What the Anthropic Ruling Means for IP Valuation in the Age of AI
A recent decision from the Northern District of California in Bartz et al. v. Anthropic PBC offers new insights into how U.S. courts may handle copyright and fair use disputes involving AI training data. For companies and professionals involved in intellectual property valuation, this ruling clarifies the line between lawful use and infringement, especially as generative AI systems increasingly rely on copyrighted inputs.
Case Overview
The plaintiffs, a group of published authors, sued Anthropic, the AI firm behind Claude, for copyright infringement. Anthropic admitted to downloading millions of books from pirate sites, scanning purchased physical books, and retaining these texts to build a central digital library. These books, including the plaintiffs’ works, were used to train large language models (“LLMs”).
The court was asked to rule on whether these uses were protected under the fair use doctrine. The answer: It depends on the use.
Key Findings
- Training LLMs Was Fair Use
The court ruled that using copyrighted books to train LLMs like Claude was “exceedingly transformative” and qualified as fair use under Section 107. The court likened the process to a human learning to write by reading and internalizing literature. Importantly, the court noted that Claude did not output infringing material, no excerpts or copied texts were shown to users. - Digitizing Purchased Print Books Was Also Fair Use
Anthropic’s practice of scanning legally purchased books and storing digital copies for internal research purposes was deemed permissible. The court viewed this format change from print to searchable digital files, as a non-exploitative, transformative use, citing benefits like storage efficiency and search functionality. - Pirated Copies Were Not Fair Use
In contrast, Anthropic’s use of books downloaded from pirate websites was explicitly rejected as fair use. The court found no justification for retaining unauthorized copies, especially for books that could have been lawfully purchased. This was not transformative; it was a cost-avoidance strategy with no legal shield.
Implications for IP Valuation
From a valuation standpoint, this case reinforces a few key takeaways:
- Transformative Use May Reduce Liability Exposure—but Not Always
If a company can clearly demonstrate that its use of copyrighted materials meaningfully alters the original work’s purpose (e.g., training rather than distribution), courts may view that favorably. - Provenance and Licensing Matter
Valuation analysis should factor in not only the technical capabilities of AI firms but also the origin and licensing status of their training data. Firms that rely heavily on unauthorized content, even for seemingly internal purposes, may face legal and reputational risks, which in turn may impair value. - Ongoing Legal Uncertainty
While this ruling clarifies certain aspects of AI-related fair use, it leaves open questions, especially around derivative works, training on other types of IP, and secondary uses. This underscores the need for ongoing legal monitoring in AI-heavy IP valuations.
Conclusion
The Anthropic decision is not the last word on copyright and AI, but it sets a precedent with clear implications for IP strategy, risk assessment, and valuation. For companies developing or investing in AI technologies, ensuring proper data sourcing and documenting fair use positions will be critical, not just for legal compliance, but also for protecting enterprise value in M&A, licensing, and funding transactions.