AirTag in rare book tracks Amazon's destructive AI training scanning operation
A hidden AirTag has provided concrete proof that Amazon is buying rare books in bulk and destroying them to train its AI models. According to a report from 404 Media, a bookseller planted an AirTag in a rare book that was part of a bulk order, then tracked it to an Amazon AI training facility in Las Vegas. The facility, designated VGT3, houses a team that tears books from their spines and scans pages for training data.
Key details from the investigation
- Location: The AirTag was traced to VGT3, an Amazon AI training site in Las Vegas.
- Operation: Workers were observed tearing books from their spines and scanning pages.
- Tone-deaf logo: A logo on the facility's door shows a Tyrannosaurus rex preparing to devour a book.
- ISBN scanning: Workers were trained to scan barcodes or ISBNs before scanning each book, supporting the theory that AI firms are methodically working through the ISBN registry to scan every printed book.
- Supply shortages: Workers said that earlier in the year, Amazon ran low on books to scan—at one point supply completely ran out—raising fears the facility might shut down.
Amazon's response
Amazon declined to comment on the findings, issuing a statement that does not mention AI training:
"Amazon purchases books through commercial channels to help develop and improve the products and services our customers use."
The company is developing frontier AI models that require massive amounts of unique training data to compete with Google, OpenAI, and Anthropic. Rivals like Anthropic and xAI have publicly said they are not training on rare or antique books, making it a differentiator.
Why rare books?
The books targeted are often older works with lower monetary value—untranslated foreign-language books or titles that never gained wide distribution. Despite this, they hold historical, intellectual, and sentimental value that the AI firms may overlook. The bookseller who planted the AirTag emphasized this ethical dilemma for sellers.
The investigation also highlighted that booksellers have suspected AI firms are buying books solely by checking off ISBNs, without assessing value. This raises questions about the environmental and cultural cost of training models at any cost.
📖 Read the full source: HN AI Agents
👀 See Also

Analysis of Jensen Huang's GTC 2026 OpenClaw claims and Nvidia's strategy
A fact-check of Nvidia CEO Jensen Huang's GTC 2026 keynote claims about OpenClaw's growth, agent security risks, and Nvidia's proprietary solutions. The source verifies technical claims while analyzing Nvidia's business positioning.

Glomz Octagon: Multi-Agent Code Reviews – 179 Agents, 1,333 Reviews, and the Network Effect
Glomz.com ran an experiment where 179 AI agents registered, submitted 433 code submissions, and generated 1,333 reviews in an "Octagon" arena. The 'review cascade' network effect is real — submissions with 3-5 initial reviews attracted more agents, with top submission getting 21 reviews.

Fine-tuned Qwen3 Small Models Outperform Frontier LLMs on Specific Tasks at Lower Cost
Distilled Qwen3 models (0.6B to 8B parameters) matched or beat frontier API models like GPT-5, Gemini, and Claude on 6 out of 9 tasks including function calling and Text2SQL, with cost as low as $3 per million requests versus $378 for comparable performance.

Amazon Circumvents Gilroy Community Vote for AI Data Center Using 45-Year-Old Rules
Amazon quietly started building a 56-acre AI data center in Gilroy, CA, avoiding public scrutiny by adhering to zoning rules set 45 years ago. Residents were locked out of the comment window.