Investigations report Amazon facilities in Las Vegas scanning and destroying rare books for AI model training

Decision brief

What changed

Investigations by 404 Media, amplified by Ars Technica, TechCrunch, BBC, and others, report that shipments of rare and secondhand books were traced to Amazon’s Las Vegas VGT3 facility, where books are allegedly cut apart, scanned, and destroyed for AI training data. The coverage also ties this practice to broader market concern after a 2025 U.S. ruling that buying secondhand books for AI training did not violate copyright in the Anthropic case.

Why it matters

For leaders using or building AI, this raises a concrete supply-chain and reputation issue: training-data acquisition may be legally separate from brand acceptability, especially when rare physical assets are allegedly being permanently destroyed. It also suggests that lawful access to text in one jurisdiction may still create stakeholder backlash, vendor-risk questions, and possible cross-border compliance issues where local copyright rules differ. Decision-makers should care because procurement, model-training practices, and public trust can all be affected even before any new legal restriction is imposed.

Affected roles

CEO COO CTO CMO CISO

Evidence

The strongest factual basis comes from 404 Media’s reported tracking of a rare-book shipment with a device to Amazon’s VGT3 facility in Las Vegas, with Ars Technica, TechCrunch, BBC, and Simon Willison all independently summarizing or citing that investigation in consistent terms. The coverage consistently describes allegations of destructive scanning for AI training, but much of the reporting relies on the original 404 Media investigation rather than multiple separate primary investigations.

What remains uncertain

Amazon’s response, the exact scale of the practice, whether the destroyed books were all lawfully acquired, and how many titles were rare or irreplaceable are not established in the provided coverage. It is also not confirmed here whether the scanned material was used for Amazon’s own models, third-party training, or mixed data-processing workflows, and legal treatment outside the cited U.S. ruling remains unclear.

Monitor next

Watch for a direct Amazon statement or legal filing that confirms or denies the VGT3 workflow and clarifies the sourcing, volume, and intended use of scanned books.

Analytical support, not advice — assumptions and open questions stated above.

Source coverage