Two US decisions find that reproducing works to train large language models is fair use – Part 3: Comparing the Anthropic and Meta decisions

In Part 1 and Part 2 of this series, we discussed two recent US decisions (Anthropic and Meta), which found that the unauthorized use of copyright-protected books to train large language models (LLMs) was fair use.¹ In this final instalment, we examine three aspects of the decisions that may frame the battleground on future fair use cases relating to LLMs. Finally, we consider the state of Canada’s copyright laws.

Separating training an LLM from antecedent uses

Both cases separately considered use of works for training LLMs from antecedent uses. 

In Anthropic, Judge Alsup found training LLMs was fair use but that the prior building of a central library from pirated works was not. Central to Judge Alsup’s finding was that every pirated copy was retained for “any number of further uses,” only one of which was to train an LLM, and each copy displaced demand for the works.²

In Meta, Judge Chhabria held that Meta’s antecedent downloading must be “considered in light of its ultimate, highly transformative purpose: training Llama.”³ For Judge Chhabria, training the LLM (which was transformative) legitimized the earlier downloading.

This suggests copying prior to LLM training may be fair use when closely tied to the training, but otherwise, the court will consider the fairness of that use independently.

Significance of pirated copies

Judge Alsup and Judge Chhabria seemed to disagree on the significance of pirated copies. 

In Anthropic, Judge Alsup doubted that “any accused infringer could ever meet its burden of explaining why downloading source copies from pirate sites that it could have purchased or otherwise accessed lawfully was itself reasonably necessary to any subsequent fair use.” ⁴

Judge Chhabria disagreed with this approach and rejected the argument that “the fact that Meta downloaded the books from shadow libraries and did not start with an ‘authorized copy’ of each book gives them an automatic win.”⁵

Overall, this may be a superficial difference in approach. Ultimately, what seemed to concern Judge Alsup most about the piracy was the displacement of the demand for the plaintiffs’ works. While Judge Chhabria in Meta was less concerned about piracy for its own sake, he was explicit that harm to the market should be the most important factor in the fair use analysis.

Disagreement on market dilution

The greatest tension between Anthropic and Meta related to market dilution by an AI’s competing works. Judge Alsup in Anthropic analogized training LLMs to humans learning how to read and write and publishing competing works. He rejected the plaintiffs’ argument that “training LLMs will result in an explosion of works competing with their works,” which would dilute the market.⁶

Judge Chhabria in Meta explicitly rejected Judge Alsup’s analogy focusing on the speed and ease with which competing works could be created with generative AI:

Speaking of which, in a recent ruling on this topic, Judge Alsup focused heavily on the transformative nature of generative AI while brushing aside concerns about the harm it can inflict on the market for the works it gets trained on. Such harm would be no different, he reasoned, than the harm caused by using the works for “training schoolchildren to write well,” which could “result in an explosion of competing works.” According to Judge Alsup, this “is not the kind of competitive or creative displacement that concerns the Copyright Act.” But when it comes to market effects, using books to teach children to write is not remotely like using books to create a product that a single individual could employ to generate countless competing works with a miniscule fraction of the time and creativity it would otherwise take. This inapt analogy is not a basis for blowing off the most important factor in the fair use analysis.⁷

Nevertheless, Judge Chhabria found in favour of Meta based on the evidentiary record, but suggested plaintiffs will likely win with better evidence on market dilution:

In cases involving uses like Meta’s, it seems like the plaintiffs will often win, at least where those cases have better-developed records on the market effects of the defendant’s use. No matter how transformative LLM training may be, it’s hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars while enabling the creation of a potentially endless stream of competing works that could significantly harm the market for those books.⁸

Moving forward

These are the first cases to consider fair use in training LLMs. The decisions reveal commonalities in how the courts may approach these cases but also important divergences in judicial reasoning. They suggest the battleground for future litigation may focus on proving harm (e.g., through proliferation of near substitutes for the original works).

Connections to Canada

Rather than a general fair use exception to copyright infringement, Canada has a “fair dealing” exception requiring the reproduction to be both “fair” and for an enumerated purpose. Accordingly, had Anthropic and Meta been decided in Canada, the court would first consider whether the copying in issue was for: research, private study, education, parody, satire, criticism or review, or news reporting.⁹

If that were established, then the court would assess whether the copying was “fair” considering: the purpose of the dealing, the character of the dealing, the amount of the dealing, the nature of the work, available alternatives to the dealing, and the effect of the dealing on the work.¹⁰ These are factors similar to the fair use factors set out in section 107 of the US Copyright Act.

Cases relating to training AI are winding their way through courts in Canada. How the differences between fair dealing in Canada and fair use in the US impact the development of those doctrines in the context of LLM training remains to be seen.

In parallel, Canada is modernizing its Copyright Act to address generative AI.¹¹ From October 2023 to January 2024, the federal government consulted with stakeholders including cultural industries, technology industries, public interest groups, legal practitioners and scholars, and Indigenous communities. The consultation considered whether the Copyright Act should clarify when using copyright-protected works for AI training requires authorization from rights holders and when it should fall under an exception to copyright infringement, if at all.

In June 2025, Canada’s AI minister, Evan Solomon, announced the government is continuing to work on an updated “regulatory framework” for AI that will include provisions on copyright. Mr. Solomon acknowledged the government supports creators being compensated for their works, but the courts will determine where to draw the line.¹²