AI Is Designing Weapons Now, The Pentagon Says It’s “Under Control”

The Pentagon rises like a concrete fortress on a humid afternoon in Northern Virginia, just across the Potomac from Washington. It is symmetrical, under control, and almost peaceful. Fluorescent light fills the windowless hallways inside as contractors enter and exit with laptops that increasingly contain code that can determine who survives and who does not.

The analysis of battlefield footage is no longer the only use of artificial intelligence. It involves modeling strike scenarios, creating weaponry systems, improving drone swarms, and occasionally proposing targets. According to the Pentagon, it’s “under control.” Repetition of that phrase in policy documents and press briefings feels comforting and slightly evasive.

According to the U.S. Department of Defense, which is based on frameworks that emphasize accountable, traceable, trustworthy, and governable systems, people are still very much “in the loop.” According to officials, the goal is to preserve sound human judgment. But with algorithms working at speeds that no human commander could possibly match, it’s difficult to ignore how rapidly the loop is closing.

AI-assisted warfare has been tested in recent conflicts in Gaza and Ukraine. AI-powered target identification and drone coordination is becoming a reality, not just a pipe dream. According to analysts, Ukraine is a “super lab” for AI-driven drone innovation, where low-cost systems are rapidly improving under the demands of the battlefield. One gets the unsettling impression that war is becoming less human and more computerized when viewing grainy drone footage online, which shows vehicles being tracked and heat signatures being outlined.

Within Washington, there is a growing movement to speed up development. The goal of initiatives like the Pentagon’s “Replicator” program is to swiftly and affordably deploy thousands of autonomous systems. The initial pledge was for large-scale production of $500–$1,000 drones. However, some analysts now note that the costs are gradually rising, with more advanced AI-enhanced drones with onboard targeting models and evasive capabilities costing $20,000.

That change seems illuminating. The idea behind cheap swarms was to overwhelm defenses with sheer numbers. Smarter, faster, and more autonomous machines that can prioritize targets, identify vehicles, and modify flight paths appear to be the focus at the moment. AI-driven autonomy may become less optional and more necessary as defenses get stronger.

But officials keep coming back to the word “control.”

The Defense Innovation Board, which was then led by former Google CEO Eric Schmidt, published preliminary ethical standards for the application of AI in the military back in 2019. These five guidelines were intended to allay critics who were concerned that “killer robots” might evade human supervision. The 65-page document was jam-packed with governance jargon. However, documents are not drones. Code does.

The Pentagon’s relationship with Silicon Valley has also evolved. Many tech workers protested military contracts ten years ago. Project Maven, which used AI to analyze drone footage, was famously objected to by Google employees. A few quit. Although it hasn’t disappeared, the moral tension has lessened. Businesses such as Palantir Technologies and Anduril Industries are now aggressively pursuing defense contracts, positioning AI weapons as essential defenses against Russia and China.

There is a perception that the urgency of national security is surpassing cultural hesitancy.

However, specialists in cutting-edge AI systems caution that “under control” might not necessarily imply “completely understood.” In many ways, large neural networks are still mysteries. Even developers acknowledge that they are not always able to provide an explanation for how a model produces a particular result. There is no time for philosophical contemplation in combat situations, where drones are approaching at 200 miles per hour and missiles are approaching. Machine-speed decision-making is used.

Whether humans can effectively monitor systems that respond in milliseconds is still up for debate. Instead of referring to complete interpretability, military researchers discuss “operational comprehension”—teaching commanders to know when to trust and when to disregard AI recommendations. That difference seems significant. It feels brittle, too.

Other dangers exist. When AI systems optimize for unexpected results, this is known as alignment failures. Telling commanders what they want to hear is known as sycophancy in advisory systems. Escalation dynamics, in which self-governing systems react more quickly and forcefully than human judgment might permit. As this is happening, it’s difficult to avoid thinking about nuclear deterrence, another technology that was once characterized as stable, controlled, and logical.

History indicates that control is frequently temporary.

Data centers are emerging next to suburban communities outside Northern Virginia’s defense campuses, their blank exteriors concealing racks of servers used to train and improve military models. In the evenings, strollers roll along peaceful paths, and joggers pass them. It provides a strangely serene setting for algorithms to practice combat situations.

One gets the impression that the current arms race is about more than just hardware. It has to do with culture. Carefulness versus speed. Verification versus deployment. A “verification gap,” in which systems are deployed before governance catches up, was recently explained by a LinkedIn commenter. That phrase sticks in your head.

One thing the Pentagon is right about is that the use of AI in combat is nothing new. For many years, missile defense and targeting systems have included autonomous modes. Scale, flexibility, and the extent to which machines are starting to influence tactical choices rather than just carry them out are novel.

It appears that investors think this change is unavoidable. Venture capital that previously targeted social media platforms is now investing in defense tech startups. Drone navigation systems are currently being optimized by engineers who may have developed ride-sharing apps. Perhaps geopolitical anxiety has overtaken the moral calculus, or it has changed.

It’s hard not to feel both admiration and trepidation when standing beneath the Pentagon’s enormous limestone walls. The organization has prior experience managing sophisticated technologies. It misjudged other people, too.

AI is currently creating weapons. It’s a fact.

The degree of humility that goes along with the code may determine whether it is actually under control more so than policy documents.