The Divide
January 2026. A room full of SpaceX employees. Pete Hegseth, Secretary of Defense, delivers what
he clearly intends as a policy commitment: “We will not employ
AI models that won’t allow you to fight wars. We will judge AI models on this standard alone;
factually accurate, mission…”
The sentence is worth reading twice. It doesn’t just express a preference. It announces a
procurement criterion. AI models that impose constraints on lawful military operations won’t make
it into Pentagon contracts. The standard by which models will be judged is mission capability –
not safety posture, not constraint architecture, not human oversight design.
The Washington Post and the Department of Defense’s
own official record at war.gov both confirm the quote. This wasn’t a misquote or a rumor. It’s
the Secretary’s stated position, on record, at a named event.
Four months later, at SOF Week 2026 in Tampa, the premier annual gathering of the global
Special Operations Forces community, that position met its most visible public challenge.
According to AP News reporting, Adm. Frank Bradley, head of U.S. Special
Operations Command, warned the military must be “very careful” about AI’s integration into “the
delivery of lethality.” He insisted on human-in-the-loop confidence before any AI deployment in
lethal contexts. Bradley’s specific language comes from a single AP News report with a broken
source URL and no independent cross-reference for the exact quotes. The position itself, that
human control must precede lethality decisions, is entirely consistent with longstanding SOCOM
doctrine and with NIST AI RMF human oversight principles. But readers should treat the specific
phrasing as AP-attributed, not independently confirmed.
Two positions. Both public. Neither codified. That’s the governance gap.
The Policy Vacuum
No classified directive has been publicly disclosed that governs battlefield AI deployment
standards as of May 31, 2026. No executive order defines what “human-in-the-loop confidence”
requires before a lethal operation proceeds. The DoD AI Ethics Principles, published in 2020 and
not rescinded, include a “Human Responsibility” principle stating that humans must “exercise
appropriate levels of judgment and remain responsible for the development, deployment, and use of
AI.” But “appropriate” is doing a lot of work in that sentence, and neither Hegseth’s nor
Bradley’s position is technically inconsistent with it.
This isn’t unusual for emerging technology at the intersection of classified operations and
civilian oversight. What’s unusual is that the gap is now publicly visible, named, and contested
by principals at the top of the civilian and uniformed chains. The exchange illustrates a tension
visible across the defense establishment, between speed-of-deployment priorities and
safety-first perspectives, though no formal policy document currently governs this dispute or
assigns authority to resolve it.
Defense contractors operating in this space are building to contract terms. Contract terms
reference existing DoD guidance. Existing DoD guidance doesn’t resolve the Hegseth-Bradley
question. The result is that vendors are making design decisions, specifically, decisions about
kill-switch architecture, human override protocols, and constraint design in autonomous systems –
without a definitive federal answer to the question their systems will eventually have to answer
in operation.
Stakeholder Map
Four categories of actors have material stakes in how this resolves.
Civilian DoD leadership (Hegseth position). The Secretary’s January 2026
remarks aren’t isolated. They reflect a broader administration approach: AI deployment should be
governed by mission capability and legal authorization, not by constraint architectures that AI
vendors design for their own commercial risk management. The implicit argument is that commercially
motivated safety constraints, “won’t allow you to fight wars”, shouldn’t be baked into systems
the military pays for. This position has institutional momentum inside the current administration.
Uniformed commanders (Bradley position, per AP reporting). SOCOM commanders
carry operational accountability in ways that cabinet secretaries don’t. When an AI-enabled
decision in a lethal context goes wrong, it’s the commander’s responsibility, not the Secretary’s.
The push for human-in-the-loop confidence before deployment reflects that accountability structure.
Commanders want doctrinal cover before they’re in the field with systems that can act faster than
human review cycles. Adm. Bradley’s remarks, if accurately reported by AP News, represent that
institutional pressure made public.
Human Oversight Spectrum: Battlefield AI Design Options
What to Watch
Congressional oversight (NDAA cycle). Congress hasn’t yet weighed in
definitively. The National Defense Authorization Act is the annual mechanism through which Congress
sets binding policy on military AI acquisition. Several members, Senator Warren’s office, among
others, have raised general concerns about military AI and foreign adversary data access, though
those concerns aren’t directly tied to the Hegseth-Bradley dispute. The NDAA 2027 markup cycle is
where this conflict could receive legislative resolution. Or not, Congress could leave the gap
open, in which case it defaults back to contract-by-contract negotiation.
Defense AI vendors. This category includes frontier AI labs with active
Pentagon contracts, defense primes integrating AI into weapons platforms, and specialized defense
tech companies building autonomous systems. Their public positions on the specific Hegseth-Bradley
dispute weren’t available in this reporting cycle. What’s observable is that they’re making
architecture decisions now, constraint design, human override protocols, kill-switch
implementations, under conditions of regulatory ambiguity. The vendors whose constraint
architectures most closely match what the NDAA ultimately mandates will have a competitive
advantage in the next procurement cycle. Those who built to the Hegseth standard and find the
NDAA mandates Bradley-style controls face redesign costs. The reverse is also true.
Human-in-the-Loop Architecture: What “Strict Control” Actually Requires
The phrase “human-in-the-loop” covers a range of technical implementations. For practitioners
designing agentic AI systems for high-stakes applications, defense or otherwise, the ambiguity
matters enormously.
The NIST AI Risk Management Framework addresses human oversight
across several functions. The GOVERN function (1.7) calls for policies and procedures that ensure
AI systems’ decisions can be overridden and that humans maintain meaningful control over
high-stakes outputs. The MANAGE function (4.2) addresses incident response and the ability to
shut down or redirect AI systems when unexpected behavior occurs. Neither provision specifies a
technical implementation, they specify a requirement that the implementing organization must
translate into system design.
In practical agentic system design, “human-in-the-loop” can mean any of the following, with
very different operational and compliance implications:
Human-in-the-loop (HITL): A human approves every consequential action before
execution. Highest control, slowest cycle time. Inconsistent with many autonomous systems’
operational value proposition.
Human-on-the-loop (HOTL): The system acts autonomously but a human monitors
in real time and can intervene. Control is present but reactive. Whether “monitoring” counts as
meaningful oversight depends on the latency of action and the latency of human review.
Human-out-of-the-loop (HOOTL): Full autonomy. The system acts, the human
reviews after the fact. No pre-action control. This is the end of the spectrum that the Bradley
position, as reported by AP, explicitly rejects for lethal operations.
Hegseth’s January remarks don’t specify where on this spectrum he wants defense AI systems to
operate. They specify what he won’t accept: models that place categorical constraints on lawful
applications. That’s a constraint on the constraint, not a specification of the control level.
Bradley’s position, per AP reporting, gestures toward HITL or HOTL for lethal decisions, but
doesn’t define the technical standard either.
The real question is whether the NDAA will define it. If it doesn’t, the question defaults to
contract language, and contract language is currently being written by acquisition officers who
are reading the same gap that this brief is describing.
Who This Affects
Analysis
The organizations most exposed to NDAA-driven redesign costs aren't the frontier labs, they have resources and contract flexibility. They're mid-tier defense tech companies in multi-year contracts, making architecture decisions now that may not survive the next NDAA cycle. The gap between the Hegseth procurement standard and a potential NDAA human-control mandate is where the compliance risk lives.
Forward Signals
Three developments are worth monitoring.
NDAA 2027 markup. The House and Senate Armed Services Committees will begin
markup in mid-2026. Watch for any amendment language specifying human oversight requirements for
autonomous weapons systems or AI-enabled targeting. If such language appears, it will resolve the
Hegseth-Bradley dispute in the direction of the Bradley position, and it will do so with binding
authority that neither the Secretary’s remarks nor a SOCOM commander’s conference speech carries.
DoD AI acquisition guidance updates. The Defense Innovation Unit and the
Office of the Under Secretary of Defense for Research and Engineering periodically update AI
acquisition guidance. Any update to the responsible AI guidelines that specifically addresses
human oversight for lethal autonomous systems would represent a material development, and would
give acquisition officers something to write into contract terms.
Frontier lab contract disclosures. Several frontier AI labs have active or
pending Pentagon contracts. How those contracts describe human oversight requirements, if they’re
disclosed at all, would indicate which position is actually winning in the procurement process,
as opposed to the conference circuit. The existing TJS brief on the Anthropic-Pentagon contract
is directly relevant context for this monitoring question.
TJS Synthesis
The Hegseth-Bradley dispute isn’t a personality conflict. It’s a structural accountability
problem with a foreseeable resolution path. Commanders bear operational accountability for
AI-enabled decisions in the field. Secretaries set policy. When those two accountability
structures point at different technical requirements, the governing authority, Congress, in this
case, eventually has to choose. The NDAA is when that choice gets made.
The practical implication for compliance teams advising defense AI vendors: the constraint
architecture decisions being made in contracts awarded today are the decisions most at risk of
legislative override. A vendor whose system design reflects the Hegseth procurement standard –
no categorical operational constraints, may be well-positioned for contracts awarded in the next
12 months. A vendor whose system design reflects NIST AI RMF human oversight requirements may be
better positioned for the post-NDAA environment, if Congress writes the language Bradley’s
position implies.
The organizations that should be most concerned aren’t the frontier labs. They have the
resources to redesign. The organizations with the sharpest exposure are mid-tier defense tech
companies currently in multi-year contracts, building to a standard that may shift before delivery.
Consider whether the contracts you’re advising on include a provision for regulatory change, and
whether “human-in-the-loop” is defined specifically enough to survive an NDAA that defines it
differently.