Are AI Companies Buying Secondhand Books for Training Data? What Is Actually Known

*This is a research-based news analysis of reporting published August 15, 2026, plus public court records and provider documentation. The identities and purposes of the buyers behind the reported UK and Irish orders have not been established. We do not claim that any named AI company placed those orders.*

The short answer

Secondhand booksellers in the UK and Ireland have reported unusually large, apparently patternless orders for books across unrelated subjects, languages, and editions. Some sellers suspect the purchases are destined for scanning into AI-training datasets, but the cited reporting did not identify the ultimate buyers or prove how the books will be used.

The suspicion is plausible because books provide edited, long-form text—especially titles published before generative AI filled the web with synthetic material—and because documented AI-data programs have previously acquired and scanned physical books. Plausible is not confirmed. The immediate story is an opaque supply chain, not proof that a particular lab ordered these titles.

What booksellers reported

The Guardian reported that sellers received bulk orders from buyers in several countries, sometimes for hundreds or thousands of unrelated books. Sellers described unusual combinations, a willingness to pay full prices, repeated delivery patterns, and disruption to ordinary inventory and fulfillment.

Examples ranged across languages, obscure magazines, literature, biographies, and specialized historical subjects. That lack of a normal collector or subject pattern contributed to speculation that a machine-generated acquisition list was seeking broad textual coverage.

Marketplaces, intermediaries, recyclers, research organizations, resellers, digitization businesses, or AI companies could all participate in such a chain. Without contracts, shipping records, buyer confirmation, or a documented destination, the final use remains unknown.

Why old books are attractive as AI data

Books offer sustained arguments, edited prose, domain knowledge, translation pairs, uncommon vocabulary, and material that may not exist in clean digital form. Older editions also predate the recent flood of AI-generated web text, which can make them attractive to teams trying to reduce synthetic-data contamination.

Physical ownership does not automatically settle every right involved in scanning, copying, training, redistributing a dataset, or generating outputs. Those questions depend on jurisdiction, the works, licenses, exceptions, contracts, and the use. Buying a copy and acquiring rights to reproduce its contents are not necessarily the same transaction.

What is known about prior book acquisition

Public litigation and reporting have described large-scale book acquisition and scanning connected with model development. In the Anthropic copyright litigation, court records distinguished questions involving lawfully purchased books from allegations involving unauthorized digital libraries. Those findings are specific to that case; they do not identify the buyers in the current reports.

Anthropic told The Guardian that books are one source among publicly available web data, commercially acquired datasets, and company-generated data, and that its acquisition programs do not buy and destroy rare or antiquarian books. That statement does not confirm involvement in the reported orders.

What transparency should look like

AI developers should document major training-data categories, acquisition methods, rights controls, vendor requirements, retention, and whether physical materials are preserved or destroyed. Dataset intermediaries should maintain provenance records that survive resale and aggregation.

Booksellers and marketplaces can offer a disclosure path for unusually large digitization orders, publish preservation standards, flag rare-material risk, and distinguish ordinary resale from bulk data acquisition. That need not expose legitimate buyers' private information; it can provide aggregate transparency and contractual accountability.

Authors and publishers should avoid assuming a suspicious order proves infringement. They can preserve sales and rights records, monitor licensing activity, review current rights-reservation mechanisms where available, and seek qualified advice when evidence connects a work to an unauthorized dataset.

Why this trend matters for AI quality

The race for pre-AI books signals a data-quality problem. More internet text is now generated, summarized, translated, or rewritten by models. Training future systems indiscriminately on those outputs can amplify errors, flatten style, and obscure provenance.

High-quality human work is becoming more economically valuable at the same time that its creators demand clearer consent and compensation. Sustainable AI development needs both: strong source material and a credible chain of rights, attribution, preservation, and payment.

The larger trend

Model competition is shifting from raw data volume toward trustworthy data supply. The winning dataset is not merely large; it is useful, documented, legally supportable, and maintainable when a source is corrected or withdrawn.

The unexplained book orders are a warning precisely because the evidence stops short of the final buyer. An accountable AI-data market should make provenance easier to establish, not require the public to reverse-engineer it from shipping patterns.