AI Training and Destroyed Books: First-Sale Doctrine and the Panama Scanning Fight

HN resurrected a Forbes story asking whether civilization is trading irreplaceable books for marginal benchmark gains. The thread points to Forbes reporting from August 17, 2026 and follow-on 404 Media warehouse tracking: AI labs allegedly buy used and antique volumes in bulk, slice spines, scan at industrial throughput, then shred print copies to feed large language model training—all while staying inside U.S. first-sale doctrine today.

This is not a model release; it is training-data politics every builder touches when choosing corpora, signing enterprise deals, or explaining Etsy’s anti-scrape terms to customers. Below: what investigations claim, why paper came back, legal lines courts drew, and practical guidance for teams shipping models.

TL;DR

What investigations describe

Forbes ties together:

1. Authors’ lawsuit filings against Anthropic referencing Project Panama—contractors debinding books, scanning, shredding, with internal language about scanning “all the books in the world.”
2. Market signals — used booksellers reporting spikes in pre-2022 inventory and suspicious bulk orders.
3. 404 Media’s tracker story — a rare-book seller cooperated with reporters to AirTag a shipment, following skids to an Amazon-coded warehouse (VGT3) where workers described all-day scan jobs and a logo of a dinosaur eating a book.

None of that replaces a courtroom judgment, but it explains why archivists call the workflow “evil incarnate” while counsel can still argue lawful ownership.

Why labs buy paper in 2026

Forbes lists twin drivers also visible in our Dougan leak and publisher friction coverage:

Liability. Scraping LibGen-style shadow libraries invites nine-figure lawsuits from authors, news outlets, and photographers.

Data quality. The open web now contains AI slop; long-form human-edited books carry structure and style harder to fake. Physical pre-2022 imprints guarantee no synthetic contamination in the source ink.

Teams without destruction budgets still face the same quality curve—see Apple alignment limits research on how data shapes behavior.

Legal landscape (United States, not legal advice)

Forbes summarizes:

• First-sale doctrine — buying a copy confers disposal rights, including shredding.
• Transformative fair use — some rulings treated scanning owned books into internal training copies as transformative, though appeals continue.

Critics say first-sale was never meant to enable million-title digitization factories without paying authors. Expect licensing middlemen and opt-in marketplaces to market themselves as the compliant path—similar to how Amazon’s product graph memo wrestles with agent scraping.

What this means for builders

Enterprise fine-tunes: Customers increasingly ask data provenance questionnaires. “We scraped indefinitely” fails procurement; contractual corpora win even if more expensive.

Open model teams: Destructive scanning headlines fuel EU and UK transparency rules. Document retention, deletion, and objection channels before regulators ask.

RAG products: You may not train, but ingesting client PDFs triggers parallel ethics when those PDFs are scanned rare books. Prefer publisher APIs or explicit licenses.

Safety research: Irreplaceable texts lost to shredding cannot be reconstructed from model weights—another reason weights are not archives.

Hacker News thread: what developers argued

The November 13 HN discussion mixed archival ethics with practical ML takes. Common points worth separating:

• Preservation vs throughput — archivists argue destructive scanning removes the last copy of niche imprints; labs argue they legally bought and therefore may digitize.
• Web corpus decay — engineers note that synthetic web text pollutes crawls, making pre-2022 print attractive for base-model quality.
• Scale skepticism — commenters question whether shredding millions of volumes is economically real or lawsuit-flavored hyperbole; court filings still anchor the order-of-magnitude claims better than rumor.

We did not treat anonymous comment scores as verification; they explain why the Forbes piece resurfaced, not new facts.

International angles beyond U.S. first-sale

Forbes focuses on U.S. copyright doctrine. EU and UK teams face database rights, preservation exceptions, and contract law in vendor agreements that may forbid scanning even when you own a physical copy. Multinationals should not assume American first-sale analysis transfers to Frankfurt or London procurement.

Publishers are also tightening platform terms, as in Etsy’s AI training clause and ongoing publisher leak battles.

Dataset hygiene checklist for ML leads

1. Maintain a provenance registry: title, seller, license, destruction yes/no.
2. Separate pretraining corpora from enterprise fine-tune data reviewers can audit.
3. Ban shadow library URLs in corporate networks if legal risk is unacceptable.
4. Prefer synthetic or licensed data for customer-facing assistants regardless of base model sourcing.
5. Document DPIA impacts when personal data could appear in scanned ephemera.

Why this story persists in 2026

The Forbes article is from August, but the HN bump coincides with broader copyright tension in AI: California freeze lawsuits, publisher investigations, and consumer backlash against “AI slop.” Destroyed books are a visceral metaphor for irreversible training choices—even when law currently permits them.

Copyright licensing scroll illustrating legal questions around AI book scanning datasets

Anthropic Project Panama in context

Forbes ties Project Panama to litigation discovery: bulk purchase, debinding, scanning, shredding. That label matters for vendor diligence—if your enterprise buys Claude, procurement may ask whether training data practices affect indemnities. We do not have the full unredacted filings in this run; treat dollar figures inside Panama discussions as allegations until courts rule.

Alternatives labs actually use

Not every frontier lab destroys paper. Common alternatives include licensed corpora, publisher deals, synthetic textbooks, and web crawls with robots.txt compliance. Each carries cost and quality trade-offs the Forbes piece outlines: shadow libraries risk suits; web text risks slop; paper risks cultural backlash.

Museums and libraries responding

Archivists argue digitize without destroy through non-destructive scanning. Industrial debinding is faster for skewed bindings, which is why contractors prefer it. Policy fights may shift toward mandatory deposit copies if legislatures react to viral stories—watch EU Text and Data Mining debates separately from U.S. first-sale.

Questions compliance officers should ask vendors

1. Do you destroy physical sources after scanning?
2. What share of pretraining data is purchased print versus licensed files?
3. Can you provide title-level provenance on request?
4. How do you exclude pirated scans from contractors?
5. What is your retraining deletion policy if a title is later challenged?

Media literacy for technical readers

Photos of shredded spines trigger moral outrage faster than spreadsheeets of token counts. Engineers should separate emotional weight from legal analysis: something can be lawful and still damage brand trust, especially for education and publishing customers.

Connecting to shadow-library lawsuits

Forbes contrasts legal print purchases with pirated shadow libraries. Courts have not uniformally blessed either path. Builder teams should map their corpus to the least risky tier their counsel accepts—often licensed news plus public domain, not gray-market dumps.

Training pipeline documentation template

Maintain a one-page data sheet per corpus: source type, legal basis, destruction flag, languages, date range, dedupe method, PII scrub pass, and renewal date. Auditors increasingly ask for this during enterprise AI rollouts regardless of base model vendor.

What changes for fine-tune-only shops

If you never pretrain, you still face questions when customers upload scanned PDFs into RAG. Add terms of use banning unauthorized book scans and scan-for-train workflows in your product.

Historical note on destructively scanned collections

Google Books and HathiTrust famously negotiated preservation scanning with libraries under different rules than commercial labs buying used inventory on eBay. The current controversy is scale and destruction, not digitization itself—another reason headlines conflate unlike practices.

Bottom line for practitioners

You do not need to pick a side in the culture war to make a engineering decision: document provenance, refuse gray-market book dumps in corporate fine-tunes, and tell leadership when vendor training stories rely on destructive scanning. Legal permissibility and reputational risk diverge—that gap is where explainx.ai readers actually work.

Readers searching after the HN thread should also read primary court filings when they become public, not only Forbes summaries—docket text defines what was alleged versus admitted.

Schools and libraries watching this story may tighten deaccession policies for volumes labs approach—another second-order effect for local AI meetups that scan donated books without counsel review.

For educators using AI in classrooms, the story is also a prompt to teach citation and provenance alongside prompting skills—students will hear about shredded books whether or not your district buys frontier APIs.

None of that requires your team to take a press-side moral stance; it requires clear internal policy before the next procurement cycle.

Related reading on explainx.ai

• OpenAI grader misalignment incident
• Etsy AI training terms for listings
• Dougan publisher leak claims
• California AI freeze lawsuit
• Public Citizen regulatory fights
• AI security social engineering

Sources: Forbes August 17, 2026 article; secondary coverage of the 404 Media tracker investigation from The Decoder; Euronews on Project Panama and the unsealed filings; follow reporting cited there from 404 Media and Anthropic litigation filings (check current docket for admissibility).

Legal summaries describe journalism’s characterization, not legal advice.