Shake-Up at Pentagon Intelligence Agency Sparks Concern
The director of the agency responsible for analyzing satellite imagery says he wants to modernize the work. Some employees fear they’re being replaced by artificial intelligence.
When Kim Jong Un gears up to launch a ballistic missile, analysts at the National Geospatial-Intelligence Agency comb through satellite imagery, looking for distinct signs on the the ground in North Korea indicating test preparations are underway.
Now, the U.S. agency is in the midst of a concentrated push—what some have called a reorganization—emphasizing the use of advanced technology to do analysis typically done by humans, five sources with knowledge of the matter told Foreign Policy.
That shift in priorities is worrying some veteran imagery analysts who fear that their jobs might drastically change, and the technology being pushed isn’t mature enough to replace human skill and analytic capability. Those working inside and with the agency say it’s unclear exactly what the changes entail, but it’s scaring some employees, who worry the reorganization is part of a push to move work done by human analysts to artificial intelligence, and to outsource some of NGA’s work.
The agency’s director, Robert Cardillo, appears to be “doing away with imagery analysis, NGA’s bread and butter,” at least the way such analysis has historically been done, one former intelligence official with knowledge of the reorganization told FP.
NGA is an important, albeit low-profile, part of the intelligence community. While the National Reconnaissance Office is responsible for the satellites that collect earth imagery and data, NGA plots the information on maps for use by the military and the intelligence community. In places where it is near-impossible to send in human sources, the military and intelligence community can get a bird’s-eye view of the landscape.
In 2011, for example, analysts at the agency helped locate the Abbottabad compound in Pakistan where Osama bin Laden was living in his final years. The agency even helped build the replica of the compound used to train special operations forces for the 2011 raid that led to the killing of al Qaeda’s founder.
In an interview with FP, Cardillo confirmed that there were changes underway that involved advanced technology, but he denied it was pushing people out of jobs or moving too fast with technology.
“It might feel like a really big reorganization to some folks,” Cardillo told FP. “The fact of the matter is, from this office, I’ve done very little reorganization. We’ve closed down a shop or two to realign some efforts. Most recently, I did change my top tier of leadership.… I now call it an executive committee.”
About a year and a half ago, Cardillo named a new head of the directorate of analysis within NGA, what he calls the “heart” of the agency. The director, Sue Kalweit, is “trying to create an entrepreneurial spirit,” he said. “But we start and finish our day with tradecraft.”
But technology, particularly machine learning that can allow computers to scan the massive stockpile of imagery in NGA’s possession, is attractive to the agency.
While some inside the intelligence agency feel these changes are inevitable and will help move NGA into the 21st century, the restructuring is troubling some of its employees, particularly its veteran imagery analysts, who are worried their jobs are at risk and are seeking positions in other government agencies or considering early retirement.
Those inside the agency are also concerned that artificial intelligence is not yet advanced enough to truly replace most aspects of human analysis. Even with advanced technology, “imagery analysis is so vital for all these hard targets we follow,” the former intelligence official said.
Some of those hard targets might include North Korea and Iran, where imagery analysts are critical to identifying nuclear sites, for example. NGA is one of the biggest contributors of intelligence on North Korea and places a premium on providing detailed insight into the country.
NGA’s plan is already sending shock waves throughout the intelligence community, with concerns that traditional imagery analysis is at risk of disappearing, according to the former intelligence official. The CIA’s Directorate of Analysis is working to replace some of the imagery analysis capabilities it fears might be lost under the reorganization, and it is assembling teams to focus on Russia and Iran, the former official said.
The CIA declined to comment.
Cardillo has in the past publicly advocated for moving toward artificial intelligence, including plans to replace three quarters of analysts’ tasks with computers. Cardillo is “all in on [artificial intelligence],” a second source, who does business with NGA, told FP, while expressing doubt that the technology is really at the level it needs to be to stand in for trained human professionals.
Cardillo doesn’t deny the emphasis on advanced technology but tells FP he “doesn’t like the term ‘artificial intelligence’” and instead prefers “computer learning and computer vision.”
“The fundamentals of our job are to take images of the planet from all sources, some government and some commercial, and create an understanding of man-made activity around the globe,” he said. “I’m optimistic about the advances in machine learning on that part” to track, for example, “a ship in a port, a plane on a runway.”
Artificial intelligence is booming, attracting talented scientists and researchers around the globe. China has invested billions in developing infrastructure to take advantage of breakthroughs in the field.
However, the technology remains rudimentary in many respects, particularly in imagery analysis. “Imagery analysis is a skill set—you have to be taught,” another former intelligence analyst told FP. “AI is not able to replace analysts in this sense, or any other. The capability is not there.”
In the commercial sector, for example, Google was infamously called out for labeling African-Americans as gorillas via its image recognition software, and the stakes are much higher in the intelligence world for such errors.
“We’re well aware of the impact of us making a mistake: It’s putting troops in danger,” Cardillo said. “We invest heavily … on testing and evaluation.” If the United States needs to “employ kinetic force somewhere … no target is struck until a human analyst affirms that what was automated was correct. We are not turning over that kind of decision to computers.”
Artificial intelligence is improving over time in pattern recognition and statistical analysis, said Todd Hughes, the chief technology officer at Next Century Corporation. “This occurs by giving a system a set of examples or class of objects that makes intuitive sense to an end user … like faces, vehicles, weapons, trees, or what have you,” he told FP in an interview. “It’s actually not that sophisticated.”
Hughes, who previously worked as a program manager at the Pentagon’s Defense Advanced Research Projects Agency, where he focused on research into automating imagery analysis, said machines could be well suited to rote tasks such as scanning endless numbers of pictures for specific objects. “Humans are actually not that great at it. Humans have limitations of attention and fatigue much more quickly than you would think,” he said. “Machines can go on forever.”
Machines might provide “tools that will help [analysts] do their jobs more efficiently,” Hughes said, but they wouldn’t be taking the place of humans working in the intelligence agencies anytime soon. “To think that there’s going to be some wholesale replacement … is roughly overstating it,” he said.
Regardless, for employees at the agency who believe their jobs are being taken by machines, the concerns about the reorganization are real. “Morale is very low,” said the source who does business with NGA.
Cardillo, however, said attrition rates from NGA are within historical rates. But he acknowledges he could lose analysts, particularly to private industry, where salaries are higher.
Looking into the future, Cardillo acknowledged the limits of using computing to replace humans, but he said it’s a matter of knowing where machine learning can be applied.
“Computers are very good at identifying what is in an image, but not good at identifying what’s not in an image,” he said. “To those who say this isn’t working, or we aren’t making progress, I think we are.”