Project Panama ...

The risk with generative AI, as this single instance with Anthropic indicates, is that we cede the means of production of our large language lives: that we turn from creators to consumers.’
Photograph: Dado Ruvić/Reuters

Should we destroy all the books in the world?

An answer to this question can be found in the court documents of Bartz v Anthropic PBC. The northern California district court case, decided in late July this year, highlighted the improbably named “Project Panama”, one of the AI company Anthropic’s efforts to improve its large language model Claude. “What is Project Panama?” court exhibit 21 asks, in an internal memo. The answer: “Project Panama is our effort to destructively scan all the books in the world.” The memo advises discretion: “Why use a codename? … Because we don’t want it to be known that we are working on this.”

Destructive scanning was Anthropic’s solution to a problem: in order to “train” Claude, Anthropic had to procure a large, high-quality language dataset, preferably one created before 2022 and the corrupting influence of generative AI on contemporary text. Claude needed as many language combinations as possible, to improve its ability to predict language outcomes. Anthropic needed data, lots of it, of very high quality. Books, as it happens, remain one of the best sources for complex, high-quality, long-form text. As the court’s decision relates, Anthropic hoped that books’ “well-curated facts, well-organized analyses, and captivating fictional narratives” would help “Claude write as accurately and as compellingly as Authors”.

Anthropic had a choice: it could have secured copyright permission to use existing e-books. This would have required the “legal/practice/business slog”, as Anthropic’s co-founder and CEO phrased it, of managing copyright. Rather than engage with the texts’ owners, Anthropic first chose to use pirated sources instead, a decision informing the company’s $1.5bn out-of-court settlement with authors. When that approach seemed too complicated (or, as the court decision phrases: “Anthropic became ‘not so gung ho about’ training on pirated books ‘for legal reasons’”), Anthropic turned to destructive scanning. As it happened, the judge ruled that using proprietary material to “train” an LLM did not, in and of itself, constitute an infringement of copyright. To the court, it would seem, “training” a corporate product is equivalent to training any human, teaching how to read in order to learn how to write.

Burning books by the Nazis in 1933 was the attempt to erase knowledge from society. Anthropic's destruction of books by different means does the same thing. Like the Nazis, they got away with it.

The Nazi book burnings were a campaign conducted by the German Student Union (German: Deutsche Studentenschaft, DSt) to ceremonially burn books in Nazi Germany and Austria in the 1930s. The books targeted for burning were those viewed as being subversive or as representing ideologies opposed to Nazism. 

Epilogue

But perhaps the most telling aspect of this case is that it was so important to Anthropic to have access to a dataset of complex, long-form, uncorrupted text. Let’s ask ourselves why that text should seem so profitable. One response to Bartz v Anthropic PBC might be to refuse to devalue our lives as readers and writers, to claim ownership of the cultural spaces in which our thought and words are created, shared and preserved. The risk with generative AI, as this single instance with Anthropic indicates, is that we cede the means of production of our large language lives: that we turn from creators to consumers, and hand the generative promise of our work to large language models and their proprietors