We have to have confidence it will exercise violence only where we intend
Defense Secretary Hegseth pressures the Pentagon to deploy AI without ideological limits, while Admiral Bradley warns troops must ensure AI violence occurs only where intended. Trump rejected an executive order on AI safety to preserve U.S. technological advantage over China, canceling the measure hours before signing.
- Trump canceled an AI safety executive order hours before signing to preserve U.S. technological advantage
- Defense Secretary Hegseth rejected AI systems that limit military applications
- Pentagon blacklisted Anthropic and terminated a $200 million defense contract over AI safety disputes
- Army's 18th Airborne Corps used AI to guide artillery with 2,000 fewer soldiers
- Air Force used AI bots to instantly downgrade classified intelligence during Iran war
The Trump administration seeks to expand AI use in military operations without safety restrictions, despite warnings from senior military officials and disputes with tech companies like Anthropic over autonomous weapons and ethical safeguards.
The White House is moving to unleash artificial intelligence across the military without the safety guardrails that some of the technology's own makers are demanding. The push comes from the top: President Trump canceled an executive order on AI safety just hours before he was scheduled to sign it, worried that restrictions might erode America's technological edge over China. "We're ahead of China, we're ahead of everybody, and I don't want to do anything that gets in the way of that," he told reporters.
Defense Secretary Pete Hegseth is driving the acceleration. He has told Pentagon officials to use AI in any legally permissible way they see fit, and he has made clear he will reject any AI system that limits military applications. In January, speaking to SpaceX employees, Hegseth said he wanted technology that operates "without ideological restrictions that constrain legal military uses." But this ambition is colliding with resistance from inside the military itself, and from the technology companies being asked to build these systems.
Admiral Frank Bradley, who commands U.S. Special Operations, raised the concern most directly at a conference in Florida. He acknowledged that AI could one day decide which targets to strike. But he also said something that cuts to the heart of the tension: "We, as human beings, have to have confidence that it will exercise violence only where we intend it to." Bradley oversees the units responsible for the military's most difficult and dangerous operations. His caution reflects a real problem—the faster AI systems operate, the less time humans have to verify that the machine is doing what it was supposed to do.
The disagreement has become unusually public. Anthropic, a major AI company, refused to let the Pentagon use its Claude chatbot on classified networks without restrictions. The company's CEO, Dario Amodei, worried about autonomous armed drones and mass surveillance powered by AI. The Pentagon responded by formally blacklisting Anthropic, terminating a $200 million defense contract and barring other government contractors from working with the company. Anthropic sued, arguing the Pentagon was retaliating illegally. The military has since pivoted to Google, OpenAI, and SpaceX—companies more willing to cooperate with the Pentagon's vision.
Within the military, there are two competing narratives about what AI will actually do. Pentagon officials describe it as a tool to help troops identify targets faster and accelerate attacks. But officers in Special Operations talk about it differently. Sergeant Major Andrew Krogman said AI could handle administrative work, freeing operators to focus on their missions. Melissa Johnson, the command's top acquisitions officer, said AI should "reduce cognitive load on routine tasks." The difference matters: one vision treats AI as a weapon accelerant; the other treats it as a staff assistant.
Both things are happening. A case study published two years ago by Georgetown University's Center for Security and Emerging Technology showed how the Army's 18th Airborne Corps used AI to guide artillery strikes with the efficiency of the best-performing unit in recent American military history—while using 2,000 fewer soldiers. More recently, the Air Force's Special Operations Command revealed that it used AI bots to instantly downgrade classified intelligence so drone operators could access it during the war with Iran. The technology is making it possible to operate at new speeds and scales.
Helen Toner, who directs Georgetown's center, noted that both descriptions are accurate. There is enormous potential for AI in bureaucratic military work, and the Pentagon is actively exploring it. But there is no question that AI is also helping the military find and attack targets faster. The question now is whether humans will remain in control of those decisions, or whether speed and competitive pressure will gradually push that responsibility to the machines. The Trump administration has made its choice clear: it wants the advantage, and it does not want to wait.
Notable Quotes
We're ahead of China, we're ahead of everybody, and I don't want to do anything that gets in the way of that.— President Trump, on why he canceled the AI safety order
We, as human beings, have to have confidence that it will exercise violence only where we intend it to.— Admiral Frank Bradley, commander of U.S. Special Operations
The Hearth Conversation Another angle on the story
Why did Trump cancel the AI safety order? It seems like a straightforward governance move.
He was afraid it would slow down American AI development relative to China. He saw the order as a constraint on competitive advantage, not a safeguard. That's the core tension—safety measures and speed are being treated as opposites.
But the military officers seem worried. What exactly is Admiral Bradley concerned about?
He's worried about the gap between intention and execution. If an AI system decides to strike a target in milliseconds, there's almost no time for a human to verify it made the right call. The faster the system, the harder it is to maintain control.
So the Pentagon and Anthropic are fighting over the same thing—who gets to decide what the AI does?
Exactly. Anthropic wanted to build in restrictions. The Pentagon wanted a blank check. When Anthropic refused, the Pentagon blacklisted them and moved to companies more willing to cooperate.
Is the military actually replacing human judgment with AI, or just speeding up the process?
Both are happening simultaneously. In some cases, AI is genuinely just handling paperwork. In others, it's identifying targets and accelerating strikes. The military's own case studies show it can do both. The problem is we don't have clear rules about where the line is.
What happens next?
Anthropic is suing. The military is building relationships with Google and OpenAI. And the technology keeps getting faster. At some point, the speed of the system will outpace the human ability to oversee it, and we'll have to decide whether that's acceptable.