A 5,000-book order is an arresting clue. It is not an identified buyer.

That distinction matters in reports that independent booksellers are receiving unusually large orders spanning obscure, unrelated titles. Sellers suspect the books may be headed for AI training. The pattern makes that hypothesis plausible. The available reporting does not establish who is buying the books or what happens after delivery.

What the booksellers actually saw

The Guardian spoke with secondhand booksellers in the UK and Ireland who said the orders had increased over recent months. The requests did not resemble a normal collection built around a subject, author, period, or customer interest.

Barter Books co-owner Stuart Manley described orders that jumped among an Estonian translation of a John le Carré novel, a particular edition of Anne Brontë’s Agnes Grey, and a 1983 issue of a naval magazine. He said three buyers had purchased hundreds of seemingly random books for about £4,000.

Other sellers described the same lack of a theme. MW Books owner Jim Shaughnessy cited titles ranging from agricultural tools in 18th-century Africa to biographies of 1950s racing drivers. One anonymous UK seller said similar buyers had ordered 6,000 books since January and paid the asking price without seeking a bulk discount. Tom’s Hardware separately reported an order for about 5,000 books received by a bookseller in Galway.

The orders also shared logistical oddities. Different buyer names could point to the same delivery location, and one postcode examined by the Guardian led to freight warehouses near Heathrow. Some buyers operated in the recycling or resale trade. The identities behind the orders remained opaque even to experienced sellers.

These are the confirmed observations: exceptional volume, unrelated titles, repeated buyer patterns, unusual shipping links, and buyers willing to pay full price.

Why AI training is a reasonable hypothesis

Large language models need large text collections. Books can offer edited, long-form material that is absent from ordinary web archives, especially when they are older, obscure, out of print, translated, or never digitized.

There is also a documented precedent. The Guardian connects the new orders to earlier reporting that Anthropic spent millions of dollars acquiring books, removing their spines, scanning them, and recycling the physical copies. Anthropic said book sourcing is widely used in language-model training and that its programs do not buy and destroy rare or antiquarian books.

That history explains why sellers are considering AI data acquisition rather than only resale. It does not connect any company to the current mystery orders.

What remains unknown

No buyer identified in the current reports has acknowledged purchasing these books for AI training. The booksellers do not describe seeing a scanning facility or a model-training contract. A freight warehouse is a shipping clue, not proof of the final destination.

There are alternatives. Libraries, universities, archival projects, recyclers, and international resellers can place large or varied orders. Some requested books are already available free in the public domain online, which complicates a simple theory that every title is being bought to obtain otherwise inaccessible text.

The most accurate conclusion is therefore narrower than the viral version of the story:

  • large and unusual secondhand-book orders are real;
  • several booksellers suspect AI training because the selections look broad and machine-driven;
  • prior AI-company book acquisition makes that suspicion credible;
  • the buyer, purpose, scanning process, and fate of these particular books remain unverified.

If a company, buyer, or logistics provider identifies the project, the evidence boundary will move. Until then, “booksellers suspect” belongs in the headline.

Three more developments worth knowing

Colorado families need the district rule, not a statewide assumption

Colorado school districts are beginning the year with different approaches to student AI access. Axios reports that Douglas County provides Gemini access in upper grades, Littleton blocks district AI access for kindergarten through eighth grade while permitting approved Google tools in high school, and Cherry Creek frames its approach around using AI safely, intelligently, and socially.

Those are district decisions, not one Colorado-wide student rule. The variation sits beside a statewide preparedness gap. In the Colorado Department of Education’s 2026 survey of more than 42,000 teachers and school-based staff, 54% of respondents said they were comfortable using AI. Only 40% of school leaders reported clear guidelines for student AI use, and 32% of all respondents said they felt prepared for future AI integration.

For a family, the practical sequence is simple: check the district’s current student-use policy, then the school’s guidance, then the teacher’s instructions for the specific assignment. Access to a tool does not automatically make its use permitted on every task.

1666: Amsterdam says its demo’s AI paintings are gone

During a hands-on preview of 1666: Amsterdam, GamesRadar noticed paintings that appeared to be generated with AI. Creative director Patrice Désilets said the team had used AI to explore variations early in development and had become so accustomed to seeing some of the images that insufficiently retouched versions remained in the demo.

Désilets said those paintings have now been replaced with work made for the game and that the final release will contain no AI-generated paintings. The account is valuable because it describes a concrete prototype-to-production decision, not a general studio position.

It is still a developer statement made before release. The final build has not yet been independently inspected, so the correct status is “the studio says the assets were removed,” not “the shipped game has been verified AI-free.” Early access is scheduled for August 25.

WeatherNext can inform experts, but it is not an official warning

Google DeepMind says WeatherNext 2 can forecast wind, precipitation, and pressure, generate six-hour forecasts four times a day, and analyze more possible scenarios than its earlier models. Its experimental cyclone system can produce 50 possible paths and intensity scenarios up to 15 days ahead.

During Hurricane Melissa in 2025, Google says WeatherNext predicted a Category 5-strength landfall in Jamaica five days ahead with 80% confidence. National Hurricane Center forecasters used it alongside physics-based models, satellite observations, and hurricane-hunter data. That combination is important: the model contributed evidence to expert judgment; it did not issue the public warning by itself.

Google’s own Weather Lab page carries the essential limit. Its predictions come from models still under development and are not official weather reports or warnings. Families should use their local meteorological agency or national weather service for decisions about preparation and evacuation.

This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.