According to Bloomberg reporter and author Katrina Manson, more than 179 data feeds now flow into the Maven Smart System, the AI platform at the heart of US military operations. Before large language models were integrated, the system could process around 1,000 targets a day. With LLMs, including Anthropic's Claude, that number jumped to 5,000. AI is also helping to speed up the "find and fix" part of the targeting cycle, the steps before a strike is authorized.

That speed comes with a cost. The US military maintains that commanders are always making the final call. But as Manson explains, the process underlying those calls is moving so fast that the ability to verify or cross-check is quietly eroding.

Manson describes an army artillery targeter who was initially skeptical, then came around, and found himself clicking "accept, accept, accept"as potential targets surfaced on screen, with little visibility into why the machine flagged them. "We know that when you rely on AI, you're also relying on it to make mistakes," Manson says. "That is intrinsic to the very nature of AI." The US military is aware of AI's susceptibility to sycophancy, escalation, bias, and hallucination. It does not yet have sufficient fixes for any of them."