Washington has already settled the policy question. The 2022 Nuclear Posture Review committed the United States to keeping a human in the loop for actions critical to a presidential nuclear decision. Congress reinforced that principle in the FY2025 defense law, saying AI should not compromise the requirement for positive human actions in executing a presidential nuclear-employment decision.

That is a strong rule. It still needs a mechanism.

“Human in the loop” describes who is supposed to decide. Software architecture determines whether a high-consequence action can actually proceed without that decision.

Nuclear command systems are built to survive attack and move information quickly under extreme pressure. AI will make that information flow faster still. The risk is that software compresses the path from warning to recommendation so aggressively that a formal human checkpoint remains on the diagram while losing practical control over the sequence.

As AI moves deeper into warning, communications, sensor fusion, option generation, and decision support, the danger is not that a chatbot suddenly launches a missile. The more realistic problem is quieter: one automated function hands a recommendation to another, timelines compress, confidence scores become operational facts, and a nominal human checkpoint arrives after the system has already shaped the available choices.

The answer is to make human authority an enforceable technical condition.

Any AI-enabled component touching a nuclear decision chain should be wrapped in an external authorization layer that the model itself cannot rewrite or persuade. Decision-support functions could summarize warning data and generate options at machine speed. A transition into a protected action class would require a separate, affirmative human authorization token tied to a named authority, a defined action, and a short expiration window.

No valid authorization should mean no transition.

The same layer should fail closed. If authorization expires, communications become ambiguous, or the system cannot verify the relevant authority, the protected action should halt rather than inherit permission from an earlier state. Every attempted transition should also generate a tamper-evident record showing what was requested, what rule applied, who authorized it, and what the software did next.

This is less exotic than it sounds. The Pentagon’s current autonomy directive already requires systems to allow appropriate levels of human judgment over the use of force. Trump’s new memorandum now orders that directive updated to keep pace with rapidly evolving AI and to preserve the chain of command.

That update is an opportunity to turn a policy principle into an acceptance test.

A program office should be able to ask a contractor to demonstrate, under realistic failure conditions, that protected actions cannot execute without valid human authorization. Auditors should be able to reconstruct every attempted crossing of that boundary. Operators should know exactly which functions AI may perform independently and where the architecture forces the machine to stop.

THE WHITE HOUSE GAVE BIG TECH AN AI CHEAT SHEET — AND LEFT EVERYONE ELSE IN THE DARK

America should move quickly on military AI. Trump is right to demand it. Speed becomes strategically useful when political authority remains unmistakable at the point where software turns information into action.

For nuclear command and control, “a human remains responsible” should be more than a sentence in policy. The software should be unable to behave otherwise.

Burak Oktenli is an independent researcher based in Washington, DC. He holds a bachelor’s degree in computer science and engineering from the University of South Florida and an MBA, and is completing a Master of Professional Studies in Applied Intelligence at Georgetown University. His research focuses on authority, autonomy, assurance, and dependency in AI-enabled systems.