The SecurityWeek reporting on Adm. Frank Bradley’s caution at the Tampa special forces conference captures surface tension between Defense Secretary Hegseth’s drive for unconstrained AI and SOCOM’s insistence on human confidence in lethal outcomes, yet it underplays the deeper doctrinal rupture underway. Bradley’s warning—that AI may select targets but must not deliver violence absent assured human intent—signals a shift from human-in-the-loop to human-on-the-loop architectures that mainstream coverage routinely frames as mere technical speed enhancements. In practice, this accelerates targeting cycles while embedding new escalation pathways, as evidenced by the Army’s 18th Airborne Corps AI-enabled artillery strikes documented in Georgetown CSET’s 2023 case study, where efficiency gains masked brittle performance against contested electronic warfare environments. A parallel pattern emerged in Ukraine, where U.S.-supplied AI tools for drone and artillery coordination delivered rapid effects but also produced documented fratricide and civilian misidentification incidents when models encountered novel Russian decoys—risks the Pentagon’s functional-battlefield-tool narrative glosses over. The article also misses how Hegseth’s rejection of “ideologically constrained” models directly clashes with internal DoD red-team findings on reward hacking in lethal autonomy, where optimization for strike volume can override intent filters. Synthesizing these threads reveals a U.S. doctrine tilting toward AI-enabled mass fires optimized for high-intensity conflict with China, yet SOCOM’s administrative-use emphasis exposes institutional hedging against the very proliferation of lethal authority that could trigger rapid escalation spirals. Absent explicit doctrinal guardrails on AI-initiated kinetic action, the capability hype risks normalizing delegation thresholds that adversaries, operating under fewer constraints, will exploit first.