US Military AI Push Stumbles Over Anthropic Feud as China Advances

Reports indicate that several artificial intelligence firms have acquired large volumes of used books, scanned their contents for training data, and disposed of the originals. This practice has drawn criticism from booksellers and collectors who argue that rare or out-of-print volumes represent irreplaceable physical records of cultural history. The approach allows AI systems to encounter works by authors such as Shakespeare, Austen, and Dostoevsky rather than relying solely on internet content often described as low-quality material.

Proponents of the scanning method note that depriving advanced models of foundational literature could limit their capacity to process complex human ideas. An AI shaped only by contemporary online output risks developing patterns detached from centuries of accumulated thought. At the same time, the disposal of physical copies raises questions about whether digital alternatives fully substitute for original texts that may contain marginal notes, unique editions, or material details lost in scanning.

This development coincides with separate tensions in federal efforts to integrate AI into national security systems. The Pentagon has issued conflicting directives to military contractors regarding the use of products from Anthropic, a company whose models have been under consideration for defense applications. One order required removal of such tools by early September, only for a subsequent message to suspend that requirement. Defense Secretary Pete Hegseth has publicly clashed with Anthropic chief executive Dario Amodei over usage limits the company sought to impose on its technology.

These shifts reflect broader difficulties within government agencies accustomed to long development cycles for conventional equipment. Officials have struggled to adapt procurement and policy processes to AI capabilities that evolve rapidly. Restrictions and reversals have created uncertainty for contractors, potentially slowing integration of models that could address vulnerabilities in existing systems.

Observers point to patterns where centralized decision-making produces inconsistent rules that affect private firms developing the technology. In contrast, market-driven training of AI on established literature occurs without direct bureaucratic oversight, though it still encounters resistance when physical assets are involved. Historical warnings from early 20th-century futurists against excessive attachment to museums and archives underscore that cultural materials serve living purposes rather than static preservation alone.

The episodes illustrate trade-offs between advancing computational tools and maintaining tangible records of prior knowledge. Private acquisition of books for scanning proceeds from commercial incentives, while military applications encounter layers of administrative friction that can delay or distort outcomes. Data on AI performance suggests models benefit from exposure to structured literary works, yet the loss of originals may eliminate future verification options not captured in digital form.

Contractors and technology developers continue to navigate these overlapping pressures as both commercial training practices and defense policies evolve.