The Pentagon Can Build Killer Drones. It Needs to Move Faster.

The debate over U.S. autonomous weapons policy is dominated by concerns over moving too fast. Detractors warn that we may risk losing control of the systems themselves or the escalation dynamics they will impact. But, as we argue in a recent report, the reality is that the Pentagon is moving too slow owing to funding, procurement, organizational, and oversight systems that are poorly suited to fielding and continuously updating the most decisive capability on the modern battlefield. While increasingly autonomous systems are not preferable in an ideal world, they are being developed in the real one. If the U.S. neglects to develop them proactively, it risks having to develop them reactively in a crisis, with far less time to conform with our ethical standards.

The services understand this. The Navy, Air Force, and Marine Corps all have distributed operations concepts that internalize the dilemma: disperse or die. But dispersal trades one problem for another. Dispersed forces strain already-taxed command and sustainment networks. The data links that enable them to preserve coherence become extremely high-value targets. Degrade those links and dispersed forces cannot fight.

Human operators alone cannot close the gaps dispersal opens, especially as populations shrink and tolerance for casualties falls. We need to leverage machine intelligence and metal across the spectrum: logistics, distributed sensing, resilient networking, decision support, and at the far end engagement. The public argument fixates on the last slice, but this is the smallest requirement.

Take the Marine Corps' Expeditionary Advanced Base Operations concept. It depends on resupplying small units dispersed across contested littorals. An uncrewed, unarmed surface connector could run those routes without necessarily ever having to fire a shot. What it must do is harder than shooting: receive a commander's intent, plan a route through surveilled water, keep going when its link drops, recognize spoofed navigation inputs, reroute around a threat it detects, and fail by loitering or scuttling rather than drifting into a sea lane. None of that involves pulling a trigger. But it is in such activities where most of the demand lies.

On the civilian side, we see a different mode of thought. Every day, millions step out of increasingly capable self-driving cars and spend their day prompting AI tools. Self-driving cars will never be perfect, but they don’t have to be, they only need to be better than the alternative at acceptable risk. The same logic should apply to warfare. Humans make mistakes when exhausted and scared. The real questions is not whether machines will be flawless. It is whether they can perform certain functions reliably and at lower cost.

The binding constraint is not the technology. The United States has the engineers and the tools. The problem is organizational. The Pentagon is used to buying platforms in orderly, linear stages across fixed categories. Algorithmic war machines do not sit still. They are fusions of hardware, software, data, and models that change constantly. They are not like tanks that can be parked. They are living capabilities that have to be maintained as such.

The key is to find a solution today that doesn’t become tomorrow’s problem. This requires the flexible, multi-year funding already requested in the Fiscal Year 2027 budget to be dispersed in a competitive process to groups of soldiers and sailors mixed with civilian engineers, materials scientists, and software engineers, stood up inside the services and command headquarters that face most sharply the operational problem. Follow on money should be tied to creating capabilities for real units and not promising panel findings. A thin board, chaired by the Vice Chairman of the Joint Chiefs, should be established responsible only for interoperability, certification baselines, and audit standards. And every charter should be written to sunset, ensuring that the incubators are given top cover mature capability and don’t morph into a self-protecting empires that smother the potential for future innovation. Autonomy needs both protected space to grow and rivalry so that losers can be culled.

It also needs to respond to moral objections. Responsible autonomous targeting should be bounded. Human judgement should still feature in setting an envelope for a system: where it may operate, for how long, and against what class of target. Inside those envelopes, systems still need to clear confidence thresholds. When links drop or pictures blur, they must fail gracefully, with a default of waiting, returning, or shutting down. And accountability must flow to the commanders who set the parameters. That is a far cry from a machine hunting on its own, and far more realistic than pretending a human can pull every trigger at machine speed. We are more likely to get tested and bounded systems if we develop them in peacetime than in the heat of a crisis. Responding to the changing world is not an option. But by acting soon we can do so on our own terms.

A second objection is that by building these systems, the United States legitimizes them and speeds their spread. But the spread is already underway, and it does not wait on Washington's permission. Ukrainian analysts who dissected Russia's V2U loitering munition found a drone that navigates and picks targets on its own, with the communication links stripped out so there is nothing left to jam — assembled largely from commercial, dual-use parts and running on freely available open-weight AI software. Autonomy has become a software problem more than a hardware one: firms are now retrofitting computer-vision autonomy onto mass-produced drones without altering the airframe. There is no fissile-material chokepoint to guard and no cadre of weapons physicists to monitor. A norm of American restraint would restrain only America. The question is not whether these weapons exist, but whether the United States shapes the terms on which they do.

No technological development will make war clean. Innocents will be hurt as the result of errors, as they are in every war. The wager is that moving imperfectly will be less risky than not moving. The best way to protect civilians’ lives and prosperity is to deter conflict. A force that cannot disperse while maintaining data flows or generating relevant mass will not deter.

The money is on the table now. Congress can let it flow into the same machinery that, since the middle of the Cold War, has produced platforms too late, too expensive, and too few, or it can use it to rebuild how the Pentagon buys and sustains the capabilities the next war will demand. The technology will arrive either way. The real open question is whether the United States or its adversaries decides what it does. Does the U.S. develop the systems of the future on its terms in advance where we can set appropriate guardrails and train our forces to use them or is it forced to react in a crisis?

MG Ferrari is a Senior Fellow at the American Enterprise Institute. He previously served as the Director of Program Analysis and Evaluation for the U.S. Army and was the Commanding General of White Sands Missile Range. Dillon Prochnicki is his research assistant.

image: U.S. Army Sgt. Jared Brust, left, and Spc. Andrew Hwang, right, both assigned to 4th Infantry Division set up a reconnaissance unmanned aerial system to be used during Project Convergence Capstone 6 at Fort Irwin, California, July 20, 2026. PC-C6 is the premier Army-hosted experiment that allows the joint force and multinational allies and partners to integrate people, equipment, and technologies to support continuous transformation efforts aimed at shaping how future warfighters will operate in fast-paced, high-tech environments. (U.S. Army photo by Sgt. Nathan Arellano Tlaczani)