Physical locks and digital keys built into computer chips make killer robots impossible to build.
April 20, 2026
Original Paper
Constrained Intelligence: A Human-in-the-Loop Architecture for AI-Guided Defense Systems
SSRN · 6304978
The Takeaway
This human-in-the-loop architecture uses hardware constraints rather than software rules to prevent autonomous weapon firing. The AI is physically unable to bypass these locks because the security is baked into the silicon itself. This approach claims that constrained AI is actually more effective on the battlefield because it removes the risk of unpredictable behavior. Safety moves from fragile policies that can be hacked to immutable laws of physics and math. This architecture provides a realistic path for deploying AI in defense while ensuring humans always hold the final key.
From the abstract
<p>The Pentagon's February 2026 demand that AI companies remove restrictions on autonomous weapons applications forces a design choice with existential implications: whether AI weapons systems will be built with architectural constraints requiring human judgment, or without them. This paper argues that the unconstrained approach is not only strategically dangerous but tactically inferior, and proposes a specific alternative.</p> <p>The paper presents a five-layer human-in-the-loop hardware archi