AP Photo/Alex Brandon, File “Military AI is going to be a race … where the risks to U.S. national security of moving too slowly outweigh the impacts of imperfect alignment,” Secretary of War Pete Hegseth said in announcing the Department of War’s artificial intelligence strategy.
The strategy’s theme is “Speed Wins,” and it directs the Department to “measure and manage cycle time and adoption rates as decisive variables in the AI era.” For many areas of work at the Pentagon, we applaud rapid adoption of agentic AI. However, when the consequences are lethal, speed must be balanced with rigorous testing and a better understanding of where AI could speed us to a dangerous outcome.
Speed is appropriate when automating non-mission-critical back-office functions. Freeing personnel from routine, repetitive tasks enables focus on higher-impact activities leveraging human judgment and creativity. Many of these back office tasks, like coding, logistics planning, and paperwork management, are indistinguishable from what commercial companies do. For instance, rather than combing through thousands of pages of budget materials manually, personnel can use natural language AI platforms like Obviant to search and analyze data.
Similarly, the department can use AI coding agents like those developed by OpenAI and Code Metal to automate coding tasks like software development and modernization, while tools like Watchtower can automate logistics planning and simulation.
Accordingly, the Pentagon launched GenAI.mil, an enterprise-wide AI platform for cutting-edge AI model access. Its 1.5 million users have already built more than 100,000 agents since the platform was launched in December 2025, automating database creation, personnel and account management, and drafting statements of work.
In early May, the Pentagon announced agreements with leading AI companies to deploy models on classified networks, enabling a new set of workflows that rely on classified data access. The Pentagon will now have to match its desire for speed by investing in more compute hardware for AI to be deployed more widely — especially at the edge within combatant commands.
Although many back-office functions are well-suited for rapid AI adoption given their low operational risk and clear efficiency gains already proven in commercial enterprises, a different standard must apply to mission-critical, lethal applications. In defense against an incoming drone swarm, AI-based fires should proceed with all due speed. But in an offensive kill chain, the tolerance for error should narrow significantly. In these contexts, speed cannot come at the expense of reliability, accountability, and human judgment. A higher standard of more rigorous testing, validation and oversight is required.
During Operation Epic Fury, the U.S. military struck 11,000 targets in 38 days, with the first 1,000 strikes happening in 24 hours. The AI system used for offensive targeting — Maven Smart System — compressed the kill chain from hours to minutes. Speed is precisely what makes the system so powerful.
However, on the opening morning of the war, the U.S. struck an elementary school, hitting the building at least twice during morning classes, killing about 175 people, most of them children. The target package identified the site as an Islamic Revolutionary Guard Corps facility, making it appear similar to other military targets. But the military had since moved, and the location was clearly listed as a school in recent Iranian business directories and on Google Maps.
This tragedy underscores how outdated targeting data can go unchallenged with devastating consequences. Although the AI system itself was not at fault for using the obsolete data it had been fed, AI should have been applied to cross-reference potential targets against open-source databases and maps. Tools like Danti AI, which enables analysts to query geospatial and open-source datasets using natural language, illustrate how AI could address this gap in the future.
Safeguards must be applied to AI agents when they are used for offensive targeting at a speed that exceeds the capacity for human verification. For example, data used for targeting should be the most recent and should be cross-checked against local open-source databases.
Since Maven Smart System uses computer vision to detect, tag, and track targets, the system should also demonstrate acceptable error rates across diverse environments, lighting conditions and target types — especially distinguishing combatants from civilians. Confidence levels on targets should be presented and assessed. If the system is only 50 percent confident about a target, then human judgment must be applied on whether to strike.
Responsible governance also requires a named officer accountable for system performance. A single official, separate from the operations commander, should certify the system as a whole for accuracy and reliability. Testing should also be conducted by adversarial teams specifically trying to fool the AI — presenting decoys, camouflage, spoofed signals, or civilian environments that mimic military ones.
Finally, every AI-assisted strike should generate a post-action review comparing what the AI predicted versus what was actually hit, with results fed back into model improvement. We must be diligent in monitoring model drift and misalignment.
The Secure and Military Accountable AI Act, introduced by Sen. Kirsten Gillibrand (D-N.Y.) is a step in the right direction toward requiring human accountability for high-stakes military decisions, rather than letting AI substitute for judgment in lethal targeting, nuclear command and control, or other sensitive consequential actions.
As we confront the most important question in military ethics today, the answer is not fast or slow, but both. We should absolutely prioritize speed in many cases. But for offensive targeting, we need to apply safeguards and judgment that live up to our historical standards of minimizing collateral damage in warfare.
Mike Brown and Maggie Gray are both with Shield Capital. Brown was previously the director of the Defense Innovation Unit, a White House Presidential Innovation Fellow, CEO of Symantec and Chairman and CEO of Quantum Corp.
Add as preferred source on Google Tags Pete HegsethCopyright 2026 Nexstar Media Inc. All rights reserved. This material may not be published, broadcast, rewritten, or redistributed.
Comments: Link copiedMore Opinions - National Security News
See All
Opinions - National Security The Korean War offers lessons for a fragile US-Iran ceasefire by Harlan Ullman, opinion contributor 2 days ago Opinions - National Security / 2 days ago 
