Pentagon Opens Door to Broader AI Role in Military Targeting

Bloomberg reported in late June that the Pentagon has revised its classified targeting doctrine to envision artificial intelligence taking a far more active role in combat decisions. That finding sits awkwardly next to the public stance Hegseth staked out only weeks earlier: in April, Bloomberg reported him saying the US does not let AI make lethal targeting decisions. The classified revision, if the reporting holds, suggests the operational picture is more complicated than that framing.

Hegseth has been pushing for AI to be used for all legitimate military purposes, "free from usage policy constraints" set by AI companies, a stance that has put him in active dispute with developers who have insisted on restrictions around autonomous weapons. Existing US policy relies on Pentagon guidance calling for "appropriate levels of human judgment" in military systems, but that language, according to reporting, leaves critical questions unanswered about exactly when and how human review must occur.

Not everyone in uniform is moving at the same pace. Admiral Frank Bradley, head of US Special Operations Command, told a special forces conference in Tampa, Florida that troops "have to be very careful" about how AI is employed in delivering lethal force. According to Fortune's reporting on his remarks, Bradley envisioned AI helping identify military targets while stressing that humans must be sure a system is "going to deliver violence only where we intend it."

That gap, between public assurances of human control and a classified doctrine that reportedly broadens AI's role, is what practitioners and oversight bodies should be tracking. Congress has been working on legislation to draw statutory red lines around Pentagon AI use in targeting. Whether that effort produces enforceable limits before the pace of deployment overtakes the policy debate will determine whether this revision marks a durable shift or a temporary divergence between classified practice and public statements.