AI firms are shredding rare books for slop-free training data

Court filings unsealed in January named it Project Panama — Anthropic's internal effort, begun in early 2024, to 'destructively scan all the books in the world.' Purchased books went through a hydraulic cutter that sheared off the spines, then industrial scanners; the paper was discarded. From one planning doc: 'We don't want it to be known that we are working on this.'

A federal judge ruled that specific practice fair use — destroying the original means only one copy exists at a time. The separate $1.5B settlement, the largest copyright payout in US history, covered pirated books, not these.

The new twist is who's buying now. ISBNdb is openly pitching pre-2022 print books to AI labs as 'structurally guaranteed' free of AI slop, handling orders from 1,000 to a million titles while keeping buyers anonymous. Dutch and other European dealers told 404 Media they went from under 20 sales a week to several hundred overnight, weighted toward foreign-language and low-circulation stock. No current buyer has been confirmed — Anthropic or otherwise.

The 'knowledge lost forever' framing is overstated: scans survive and most titles have other copies. But for genuinely scarce editions bought blind off a bulk list, the physical object is gone.

Why it matters: clean pre-AI text is now scarce enough that labs will pay to destroy the last copy of it.