Cut, copy: Why books are being destroyed to feed AI, Pg12

Key Highlights:

• AI companies, including Anthropic, are reportedly making bulk purchases of physical books in Australia and Europe.
• These books are being subjected to "destructive scanning," a process where they are destroyed after their content is digitized for training AI models.
• This practice, revealed in unsealed documents from a copyright lawsuit against Anthropic, has raised significant concerns among booksellers and copyright holders.
• The demand for physical books is driven by the need for high-quality, human-written data for Large Language Models (LLMs), particularly older content free of AI-generated text.

Detailed Insights:

• The destructive scanning process involves removing book bindings and feeding loose pages into automated scanners, with remaining pages often recycled or shredded.
• This method is favored for its speed and ability to produce flat, high-quality images for Optical Character Recognition (OCR) compared to non-destructive scanning.
• Anthropic's Project Panama was previously identified as a large-scale effort to acquire and digitize millions of physical books for building training datasets.
• AI developers view this practice as a "middle path" to obtain data, balancing the need for content with avoiding contentious web scraping and negotiating licenses.
• The legal implications under Indian copyright law are being debated, specifically regarding the applicability of the Fair Dealing Exception under Section 52 of the Copyright Act of 1957.
• While digitisation for AI training might be considered "non-expressive uses," concerns of authors and publishers extend beyond current copyright interpretations.
• Proposed solutions to address these concerns include mandatory labelling of AI-generated content and establishing funds to support authors and artists affected by AI adoption.
• This current trend echoes the earlier Google Books project, which also involved mass digitisation but primarily aimed at making books searchable rather than solely for AI training.
• Institutions like the Internet Archive advocate for and utilize non-destructive digitisation methods to preserve the integrity of original physical copies.

Key Concepts Involved:

• Large Language Models (LLMs): AI models that learn statistical patterns from vast text data to generate human-like text.
• Generative AI: A type of artificial intelligence capable of producing new content, such as text, images, or audio.
• Destructive Scanning: A digitisation method that disassembles physical books by cutting their spines to facilitate high-speed scanning.
• Fair Dealing Exception: A provision in copyright law, such as Section 52 of the Copyright Act of 1957, allowing limited use of copyrighted material without permission for specific purposes.