No rule governs this yet.
That’s the central fact behind this week’s most consequential AI governance story, and it didn’t come out of Brussels or Sacramento. It came out of a Special Operations conference in Tampa, where the gap between the Pentagon’s civilian leadership and its uniformed commanders became impossible to ignore.
Defense Secretary Pete Hegseth has made his position clear, on the record, more than once. Speaking to SpaceX employees in January 2026, he stated directly: “We will not employ AI models that won’t allow you to fight wars.” He has insisted, per AP News reporting, that the Pentagon should be allowed to use AI “any legal way it sees fit.” The position is operational: reject models that place constraints on lawful military applications.
According to AP News reporting, Adm. Frank Bradley, head of U.S. Special Operations Command, pushed back publicly at SOF Week 2026. Bradley urged that the military must be “very careful” about AI’s role in what he called “the delivery of lethality,” and insisted on human-in-the-loop confidence before any deployment in lethal contexts. These remarks come from a single AP News report with a broken source URL; Bradley’s specific language hasn’t been independently confirmed through a second source. The position itself, that human control must precede lethality decisions, is consistent with longstanding SOCOM doctrine and with the human oversight principles in the NIST AI Risk Management Framework. But the specific quotes belong to AP’s reporting, and readers should treat them accordingly.
The policy gap underneath this dispute is the real story. The exchange between Hegseth and Bradley isn’t a personnel conflict. It’s a governance conflict with no referee. The Pentagon hasn’t published a classified directive on battlefield AI deployment standards. No executive order defines what “human-in-the-loop confidence” means before a lethal operation. Defense contractors building AI-enabled systems for the U.S. military right now are operating without a definitive federal answer to the question that both Hegseth and Bradley are publicly arguing about.
Context matters here. The January 2026 SpaceX remarks weren’t a throwaway line, they were a policy declaration, confirmed by both the Department of Defense’s own official record at war.gov and by The Washington Post. Hegseth has been consistent. What’s new at SOF Week is the public pushback from a named commander at the operational level. That kind of visible disagreement between the Secretary of Defense and the head of SOCOM doesn’t appear often. When it does, it tends to get resolved through the legislative process, specifically, the National Defense Authorization Act.
The NDAA is the next real decision point. Congress uses the annual defense authorization to set policy guardrails on how the military acquires and deploys technology. If neither a classified directive nor an executive order resolves the Hegseth-Bradley dispute before the NDAA cycle closes, that’s where it lands. Compliance teams advising defense AI vendors, and engineers designing autonomous systems for military applications, should watch the NDAA 2027 markup process for specific human-in-the-loop language.
Unanswered Questions
- What contractual standard governs human-in-the-loop requirements for defense AI systems awarded before any NDAA language is enacted?
- Which DoD AI ethics principles, published 2020, not rescinded, are currently operative for acquisition teams in the absence of a new directive?
- If the NDAA 2027 mandates specific human-control thresholds, what is the remediation timeline for systems already under contract?
Warning
Defense AI vendors currently in the contracting pipeline are building to a standard that doesn't officially exist. The Hegseth-Bradley split isn't resolved by either a classified directive or an executive order as of May 31, 2026. Contracts awarded in this window may require renegotiation if NDAA 2027 codifies human-in-the-loop requirements more specifically than either party's public position suggests.
The catch is that the NDAA timeline doesn’t match the procurement timeline. Defense AI contracts are being awarded now, under contract terms that reflect neither Hegseth’s operational permissiveness nor Bradley’s caution requirement. Vendors building to one standard may find themselves out of compliance with whatever the NDAA ultimately mandates.
Don’t expect this dispute to resolve quietly. The Hegseth position has institutional momentum – it aligns with the administration’s broad approach to AI deployment across federal agencies. The Bradley position has operational credibility and the implicit backing of commanders who will be held responsible if AI-enabled decisions go wrong in the field. Both are defensible. Neither is codified. The organizations with the most exposure are defense AI vendors currently in the contracting pipeline – they’re the ones building to a standard that doesn’t yet officially exist.