2022 As They Digitize Them for Training Data, a Practice That Has Sparked a Sharp Debate Over Cultural Preservation

AI firms are destroying millions of physically printed books published before 2022 as they digitize them for training data, a practice that has sparked a sharp debate over cultural preservation. The books are prized because they are entirely human-written and free of AI-generated content, making them ideal for training large language models. Data broker ISB Ndb supplies developers with bulk collections of such works, and its digitization methods have drawn scrutiny: the company cuts off book spines and feeds pages through high-speed scanners, a process that destroys the original physical copies.

The approach is far cheaper than slower techniques that preserve the books' integrity, but critics say it amounts to cultural loss. Some of the scanned volumes are rare copies that survived wars and fires, and unlike web content, they cannot be replaced once destroyed. ISB Ndb calls the work digital preservation, but it guards client identities with strict confidentiality agreements, suggesting an attempt to manage backlash.

In the United States, a recent court ruling involving Anthropic found that scanning legally purchased books for AI training can qualify as fair use under certain conditions, with the destruction of the physical copy seen as replacing an existing legal copy rather than creating a new one. Elon Musk, responding to the controversy, said he has instructed his SpaceXAI team to preserve rare books in libraries and scan them using slower, safer methods. The aggressive scanning is also designed to prevent model collapse, the quality degradation that occurs when systems are trained on their own outputs, and to avoid data-poisoning techniques used by authors to block AI training. Still, the quiet destruction of irreplaceable physical works continues to raise serious questions about the long-term fate of cultural heritage.