US Military AI Push Stumbles Over Anthropic Feud as China Advances

The Pentagon's scramble to integrate advanced artificial intelligence into its operations has produced a familiar pattern of abrupt policy shifts and internal conflicts, even as companies building the underlying models pursue training data through methods that raise separate preservation concerns. Recent directives from the Air Force requiring contractors to eliminate Anthropic products from weapons systems by early September were quickly walked back, underscoring the difficulty large institutions face when technologies evolve faster than procurement rules or oversight structures can adapt.

This turbulence stems in part from a public dispute between Anthropic chief executive Dario Amodei and Defense Secretary Pete Hegseth over limits the company sought to place on military use of its models. The episode illustrates how commercial decisions about AI safety and deployment now intersect directly with national security planning. The National Security Agency, for instance, has relied on capabilities like Anthropic's Mythos model to identify software vulnerabilities, yet the broader contracting environment remains unsettled amid concerns about reliability and control.

At the same time, reports have surfaced of AI developers acquiring large volumes of used books, scanning their contents, and discarding the originals. The practice has drawn criticism from booksellers and collectors, particularly when rare or historically significant volumes are involved. Proponents argue that training systems on canonical literature from authors such as Shakespeare or Austen offers a necessary counterweight to the low-quality material that dominates online sources. Critics counter that the destruction of physical artifacts reflects a broader institutional carelessness toward cultural transmission.

These parallel developments point to a common underlying problem: the absence of coherent mechanisms for deciding what information should shape powerful AI systems and under what safeguards they should operate. The military's push for dominance over Chinese capabilities has accelerated adoption timelines, yet the same agencies continue to manage equipment designed decades earlier. Rapid model updates outpace the slower processes of contracting, legal review, and interagency coordination.

China's own advances in military AI have intensified pressure on U.S. planners, creating incentives to prioritize speed over stable governance arrangements. Internal administration debates have produced contradictory signals, with some officials favoring tighter restrictions on certain providers and others emphasizing the need for broad access to frontier capabilities. The result is an environment in which private firms wield significant influence over how their technology reaches the world's largest defense apparatus.

Preservation questions around training data add another layer. If AI systems are to incorporate humanity's accumulated written record, the means of ingestion matter. Physical copies carry metadata, annotations, and contextual details that digital scans may not fully capture. Treating books as disposable inputs risks eroding the very historical depth that advocates claim is essential for more capable models.

Effective responses would require institutions capable of setting durable standards for data sourcing and model deployment. Current arrangements, shaped by competing commercial interests and shifting political priorities, have instead produced cycles of restriction and reversal. Whether those institutions can develop steadier approaches before both military applications and cultural archives are shaped by short-term decisions remains an open question.