Via Satellite GEOINT archive illustration.

Amid popular, and some military, concern about the timely accuracy and applicability of artificial intelligence (AI), the National Geospatial-Intelligence Agency (NGA) is launching a pilot program for accrediting geospatial intelligence (GEOINT) models for the National System for Geospatial (NSG) intelligence.

The agency’s Accreditation of GEOINT AI Models (AGAIM) “is the evaluation of the methodology and robustness of a program’s model development and test procedures,” NGA Director Vice Adm. Frank Whitworth told a Defense Writers Group breakfast on Friday, Aug. 30.

“The accreditation pilot will expand responsible use of GEOINT AI models and posture NGA and the GEOINT enterprise to better support the warfighter and create new intelligence insights,” he said. “Accreditation will provide a standardized evaluation framework, implements risk management, promotes a responsible AI culture, enhances AI trustworthiness, accelerates AI adoption and interoperability, and recognizes high quality AI, while identifying areas for improvement.”

According to the NGA’s Sept. 3 announcement, NGA established and launched a GEOINT Responsible AI Training for all coders and users of GEOINT AI. NGA held pilot classes in April and May. The plan is for training to eventually be broadly available to anyone in the NSG coding or using GEOINT AI capabilities.

A marquee DoD AI effort over the last nine years has been Maven, which is now an NGA program of record and which serves as an operating system for the Palantir Technologies’ Maven Smart System (MSS) — essentially a graphical user interface for military applications in areas such as targeting and situational awareness.

Another NGA AI program of record, and one that Whitworth has said is a “larger mission” than targeting, is the Analytic Services Production Environment for the National System for GEOINT (ASPEN), which globally monitors behavior changes to update and, if needed, warn U.S. officials. ASPEN is a five-year program that began in May last year, NGA said.

“Underscoring all these tenets of targeting, warning, and safety is this issue of distinction and being certain, as best we can, in making assessments and keeping an archival record of our observations,” Whitworth said on Aug. 30.

While NGA has used computer vision “for decades” for detection and for some automated conclusions, the agency can now train models through machine learning algorithms, he said.

A new report by Georgetown University’s Center for Security and Emerging Technology on the use of MSS by the Army’s 18th Airborne Corps noted AI successes — for example, a 20-member time critical targeting cell in the corps that matched the performance of a 2,000 member such cell during Operation Iraqi Freedom, and a U.S. Army goal now to leverage MSS so that firing units are able “to make one thousand high-quality decisions — choosing and dismissing targets — on the battlefield in one hour.”

Yet, the report also laid out AI growing pains. “Besides the challenges internal to the DOD, MSS contractors identified frictions in some of their relationships with government personnel,” the study said. “These frictions were summarized by company leaders as being viewed as ‘just contractors’ or ‘dirty contractors,’ a theme that has been echoed by companies that voluntarily contributed to the response in Ukraine.”

“This friction may have also coincided with initial skepticism about the utility of AI for artillery fires, which one soldier bluntly summarized as: ‘This s— don’t work,’” the report said. “Those opinions were initial obstacles that were overcome with experience, and later matured into what could be considered as justified skepticism when operators purposefully disabled algorithms when they noticed a degradation in accuracy.”

AI models, including large language models, “are starting to emerge at a quite a pace,” Whitworth said on Aug. 30. “People who are innovative want to apply models in all walks [of life].”

“In GEOINT, getting back to that issue of distinction, it is so important that we make sure these are good models because the positive identification effectively underlies whether you’re gonna be correct or whether we might have an apology on the part of our nation or an alliance,” he said.

This story was first published by Defense Daily