Pentagon launches pilot program that will use AI to manage sensitive information

The Pentagon is launching a pilot program that will use artificial intelligence to manage its classification processes and handling of sensitive military information, according to a memorandum reviewed by DefenseScoop and sources familiar. 

Within the next six months, the department will start an “initial small-scale deployment of an automated security classification capability” intended to “overhaul” the way the department classifies info, according to a memorandum issued Monday by Deputy Defense Secretary Steve Feinberg.

The DOD will use an AI-backed system the Air Force recently developed, known as the Automated Classification Management Environment (ACME), for the trial program while tapping the service’s top civilian official as the executive agent for the pilot.

If the pilot is successful, according to the memo, the automated system will eventually serve as “the single, digital authoritative reference” for all of the DOD’s original classification decisions, a responsibility historically reserved for specially designated human officials who determine how sensitive information is categorized. 

Feinberg’s directive comes amid deep public concerns over the use of AI, which has evaded human oversight in several high-profile incidents over the last few months, setting off a recent firestorm of calls to regulate the disruptive technology. At the same time, defense officials have argued that AI is too powerful to sideline against adversaries who are also wielding it, and as the military continues to inject the technology into virtually all of its operations, often in the name of efficiency.

The memo framed the Pentagon’s use of AI as a way to shed “dysfunction” caused by outdated classification and declassification procedures, which it describes as “endangering” to the department’s mission. It also claims those outdated procedures hinder the Pentagon’s “timely release of critical information,” eroding the public’s trust in the department.

“The dysfunction caused by these procedures leads to incorrect classifications that burden our national security systems, impede interoperability, and slows innovation by trapping vital data in stovepipes,” according to the memo. It described ACME as “a suite of artificial intelligence-aided tools” created by the Air Force to modernize the department’s “security classification processes and its management of sensitive information.”

A source familiar with the matter confirmed the directive to DefenseScoop and a Department of the Air Force spokesperson said the service “will comply with the direction in the memo.”

The Pentagon did not immediately answer questions about which underlying AI models would be part of the pilot and whether the department had concerns about its use of the technology to manage sensitive information.

ACME, the Air Force system, was noted in a public request for information in August. Officials asked industry to come up with a classified management ecosystem to replace the “manual, human-intensive” process the military has used for decades, calling it “incompatible” with the volume and speed of information in the digital age. 

The notice said there are hundreds of officials who have the authority to initially classify information and that they are using relatively archaic programs to do so, resulting in outdated documentation, high error rates, frequent over-classification, and degraded data integrity. The notification said the DOD has a roughly 140-million-page hardcopy backlog.

Josh Wallin, who studies AI as a fellow at the Center for a New American Security’s defense program, told DefenseScoop Wednesday that automation could help with misclassification issues, such as when information gets over-protected, under-protected or mislabeled by humans. He also said it could help reduce information stovepiping, especially as modern conflict has shown that quick information flow can prove an advantage over adversaries.

Under-classified information can expose a capability or plan to an adversary, Wallin said, while over-classification can be withheld or stalled by a delayed review process, meaning it may not be shared with organizations that need it, including “in a coalition environment where information sharing is critical to coordination.”

“They are right when they say in the memo that there are concerns about over-classification, stovepipes,” Wallin said. “I think that you would actually find that in a lot of cases, the vast majority of information isn’t necessarily so sensitive that you’re worried about a misclassification.”

But, there is still a risk for misclassification, possibly at a faster scale under AI-backed systems, and concerns posed by potentially dodgy cybersecurity protocols from allies and partners who may receive information faster under the management system. 

“Is there a potential for an adversary to get information via them, even if they can’t get it through [U.S. systems]?” he said. “I think there is a misclassification concern. There is just an inherent concern that some information will be sensitive. Will it be misclassified and shared, sort of unintentionally, and will someone not catch it?” 

There is also an “aggregate” effect to declassifying information that may accelerate under such a system, where an adversary can surmise sensitive information from piecemealed declassified data. Wallin also questioned how short the pilot program seemed and the human oversight around using AI as a definitive reference for original classification authority.    

“I would want to know procedurally what checks are going to be in place for managing misclassifications because those are going to be inevitable” in both human and AI-drive processes, he said. “How [are you] going to actually have human oversight in place? That’s going to have to persist forever. This situation is going to evolve over time. The system itself is going to evolve over time, so you’re never going to have complete 100% trust.” 

Jon Harper contributed to this report.