The U.S. Military’s AI Policy Is Up For Debate

The use of artificial intelligence in military operations has prompted a fierce discourse in Washington: Does the U.S. government have the appropriate guardrails in place to drop bombs using AI?

A relatively obscure Pentagon policy document is now at the center of that debate.

President Donald Trump has ordered the Department of Defense to update its “Directive 3000.09” — the policy governing the deployment and testing of autonomous weapons — by early September.

The last time that policy, first established in 2012, was updated was in 2023; it includes detailed protocols for how U.S. military officers should engage with guided missiles and drones, and includes oversight processes for the use of these weapons. While the use of AI is constrained by traditional rules of engagement, Directive 3000.09 does not cover the use and specific testing protocols for the emerging technology.

The timing of Trump’s order to update the policy, and the Pentagon’s recent push to deploy AI-powered systems in large-scale military operations like those in Venezuela and Iran, has raised concerns among congressional Democrats that the Trump administration wants to undermine safety protocols and human input.

“What I don’t want is for the DoD to move so fast and just start implementing policy without actually thinking about the consequences of what that means,” Sen. Ruben Gallego (D-Arizona) told NOTUS. “Once you allow AI to take humans out of the kill chain decision making process, it becomes easier to kill people, which means it’s easier for you to go to war.”

Gallego sent a letter to Defense Secretary Pete Hegseth last week asking for more visibility into what considerations are going into the update of these directives. He’s yet to hear back from his office.

“We should be thinking about it in a very constructive and deliberative manner before we just run head on into it, because bad things could happen,” Gallego said.

A Defense Department official told NOTUS in a statement that humans will remain integral to deploying automated weapons.

“Our military operates in full compliance with all U.S. laws and established Department policies, such as ensuring a human is always in the loop for critical operational decisions,” the official told NOTUS. “The Department maintains in Directive 3000.09 that a human operator has always been in the loop when using autonomous capabilities. The responsibility for the lawful use of any AI tool rests with the human operator and the chain of command, not within the software itself.”

The Trump administration has pushed the Pentagon to rapidly adopt AI across the agency. In January, Hegseth released a memorandum to accelerate adoption, encouraging “becoming an ‘AI-first’ warfighting force across all components.”

Traditionally, humans have been in charge of planning military operations, but militaries across the world have developed advanced AI capabilities such as missiles and automated drone swarms that can reach their targets with minimal human input.

Advanced AI systems like Anthropic’s Claude and OpenAI’s ChatGPT can analyze massive amounts of data and help troops identify likely targets in conflicts like the war in Iran. That usage came under scrutiny when the U.S. mistakenly struck an Iranian all-girls school labeled as a military facility due to reportedly outdated data in the early days of the war in Iran. Claude reportedly helpedorchestrate1,000 attacks in the first 24 hours of “Operation Epic Fury.”

Keeping up to date with these systems remains a challenge, and Democratic and Republican lawmakers have expressed concerns with how the U.S. military is deploying these tools.

“This is one technology where, you’re going to create a policy, and then you’re going to find six months later it’s outdated because of new innovations — it’s moving that fast,” Sen. Mike Rounds (R-South Dakota) told NOTUS.

“You just have to try to keep up as best you can,” he said. “And you have to include in that discussion what the adversaries are doing because we’re going to have to be able to respond to what the new capabilities that our adversaries are going to be introducing to battle.”

Both Democratic and Republican senators have been pushing to establish additional guardrails to the use of AI in military operations.

The Senate’s version of this year’s National Defense Authorization Act, which has yet to make it to the floor, would forbid the military from using AI to decide if a nuclear weapon should be deployed, to track constitutionally protected groups or individuals in the U.S., or to plan military attacks without the “incorporation of appropriate levels of human judgment,” the text of the bill reads. However, the bill’s text falls short of detailing what that human judgment would look like.

The bill would also enshrine in law some of the testing protocols in 3000.09 for fully autonomous weapons, like requiring the DOD to report incidents related to autonomous weapons and share details of their AI safety testing with Congress.

“I do not believe that current law is keeping up with advancing AI technology. Lethal decisions require a conscience, not just an algorithm,” Sen. Kirsten Gillibrand (D-New York), who is behind some of the provisions included in the NDAA, said in a statement.

Michael Horowitz, a former senior Pentagon official under President Joe Biden who was tasked with updating 3000.09 in 2022, told NOTUS that the current version of the document was drafted to be adaptable to new technologies.

He said that he’s not as concerned with possible updates to 3000.09 since he trusts the military will prioritize reliability in the weapons they deploy. However, he said that this moment presents Congress with an opportunity to ensure human judgment in military attacks.

“Legislation that enshrines the centrality of the human in making the decision about the use of force is a great example of a place where Congress could make a difference,” Horowitz said.