Skip to playerSkip to main content
  • 7 weeks ago
Internal office memos seen during the copyright lawsuit against Anthropic AI read “What is Project Panama? Project Panama is our effort to destructively scan all the books in the world. Why use a codename? … [B]ecause we don’t want it to be known that we are working on this.”

Category

📺
TV
Transcript
00:00AI companies have been accused of allegedly destroying bulk loads of old
00:04and rare books in order to train their large language models. A copyright
00:09lawsuit brought against USA-based AI company Anthropic revealed that the
00:14company was engaging in a practice known as destructive scanning in order to
00:19train its large language model Claude. To scan the books most efficiently the
00:24company has allegedly been slicing off the spines of physical books and pulping
00:29its pages afterwards. In some cases it's been reported that the destroyed books
00:34are among the last surviving copies of that original work. The lawsuit alleged
00:40that the project, known as Project Panama, was intended to be kept a secret from
00:45the public. According to internal memos seen during the court case, Anthropic
00:50employees and bosses describe the project like this. What is Project Panama?
00:55Project Panama is our effort to destructively scan all the books in the
01:00world. Why use a code name? Because we don't want it to be known that we're
01:05working on this. The judge ruled the AI companies had acted according to US
01:11copyright law because the scanning process amounted to a transformative use of the
01:16books. If the AI companies sought to teach their language models off existing
01:21digital works, for example, via already scanned ebooks, the companies would have
01:26been required to secure copyright permissions from the tax owners. Rather
01:32than seek this permission, Anthropic originally chose to use pirated sources, a
01:37decision that costed the company $1.5 billion in out-of-settlement costs to the
01:43authors. To avoid incurring the further legal costs, the company turned to the
01:48destructive method of book transformation. And in an ironic twist, the AI companies
01:54were forced to reveal that they were buying and hoarding large quantities of
01:58books, preferably created before 2022, to ensure the models were being trained on
02:04human thought, and to avoid cannibalizing the corrupting influence of any AI
02:10generated content in new works. Anthropic said it hoped that the book's well curated
02:15facts, well-organized analysis, and captivating fictional narratives would help
02:21Claude write as accurately and as compellingly as human authors. As always, the
02:27genius of human minds is still required.

Recommended