At Google’s campus in Mountain View, California, executives are trying to assuage thousands of employees protesting a contract with the Pentagon’s flagship artificial-intelligence initiative, Project Maven. Thousands of miles away, algorithms trained under Project Maven—which includes companies other than Google—are helping war fighters identify potential ISIS targets in video from drones.
The controversy around Silicon Valley’s cooperation with the military may intensify in coming months as Project Maven expands into new areas, including developing tools to more efficiently search captured hard drives. Funding for the project roughly doubled this year, to $131 million. Now the Pentagon is planning a new Joint Artificial Intelligence Center to serve all US military and intelligence agencies that may be modeled on Project Maven. “It’s exceeding my expectations,” says Bob Work, who established Project Maven in April 2017 while serving as deputy secretary of defense, before retiring later in the year.
Google’s precise role in Project Maven is unclear—neither the search company nor the Department of Defense will say. Two people familiar with the project said another company built the systems deployed on drone missions overseas.
Project Maven is formally known as the Algorithmic Warfare Cross-Functional Team. A seal for the group in a recent presentation, from project chief Lt. Gen. Jack Shanahan, depicts a trio of cheery cartoon robots under a Latin motto that Google Translate renders as “Our job is to help.”
The effort was created to demonstrate how the Pentagon could transform military operations by tapping AI technology already established in the private sector. On a trip to Silicon Valley last summer, Defense Secretary James Mattis lamented how his department lags the capabilities of tech companies he visited, such as Amazon and Google.
Processing drone video was selected as Project Maven’s first mission, Work says, because the Pentagon’s analysis tools can’t keep pace with the tidal wave of high-resolution aerial imagery swamping US bases. The plan was to deploy machine-learning techniques that internet companies use to distinguish cats and cars to spot and track objects of military interest, such as people, vehicles, and buildings. The initial goal was to have a system helping analysts in the field by December 2017.
That target was met handily. The Defense Department said in December that algorithms bought from unidentified contractors were helping on bases fighting ISIS. At a conference in Washington this month, Lt. Col. Garry Floyd said technology developed for Maven was being used by the US military’s Middle East and Africa commands, and had been expanded to a half-dozen combat locations. William Carter, deputy director of the technology policy program at the Center for Strategic and International Studies, says that progress is remarkable for a department famed for glacial acquisitions processes. “By DoD standards, this is literally a work of magic,” says Carter, who has been briefed by Shanahan and others on Project Maven.
The technology fielded under Maven can automatically annotate objects such as boats, trucks, and buildings on digital maps. Work notes that this helps analysts with tasks like identifying targets or understanding a group’s pattern of activity, by reducing time spent scouring screens just to find objects of interest. The software deployed to bases also has features that let analysts help retrain the algorithms, by quickly tagging new objects of interest or flagging errors.
Google’s exact function in all that is unclear. The company says it is helping the Pentagon use its open source TensorFlow machine-learning software to train algorithms on unclassified drone imagery, and that the technology is limited to “non-offensive” uses. Google’s director of AI told WIRED the work is “mundane” when asked about the internal protests last month. A Pentagon spokesperson said Project Maven “includes many leading technology and artificial intelligence companies,” but declined to identify any. Carter of the Center for Strategic and International Studies, along with another person familiar with Project Maven, say that a company other than Google developed the technology deployed in operations against ISIS.
In his recent talk, Shanahan said the project is beginning to grow. That includes deploying Project Maven’s drone surveillance algorithms more widely. The initial system was developed for smaller drones that fly at relatively low altitudes, such as the 1.4-meter, 20-kilogram ScanEagle. Shanahan said his team is now “refining” algorithms for drones that fly higher, and will soon work on high-altitude surveillance aircraft. His slides depicted the 15-meter-long Global Hawk, which flies at up to 18,000 meters (60,000 feet) and carries sophisticated conventional and infrared cameras. Ultimately, the goal is to integrate Maven’s algorithms onto drones themselves, he added.
Shanahan also said Project Maven will soon start applying AI to new areas of military operations. One is speeding the process of sorting through material captured in raids—machine-learning algorithms could be used to help analysts look for the most important material on captured hard drives. He said Project Maven will look at how AI could help military or intelligence analysts assess the relative importance of different enemy targets.
Project Maven’s future may be more expansive than the handful of projects Shanahan described. Pentagon R&D chief Mike Griffin is due to submit a proposal to Congress this summer outlining the joint artificial intelligence center to accelerate military and intelligence use of AI. “My understanding is that more money is being pushed in the Maven direction and it will be a big part of the Joint AI Center,” says Work, who co-chairs a task force on AI at the Center for a New American Security. Maven or a unit like it could become a kind of universal AI shop inside the new center, helping all US intelligence and military organizations build AI projects with commercial contractors.
If a vocal minority of Google’s more than 80,000 employees have their way, the search company won’t be one of those contractors. More than 4,000 of them signed a letter saying Google should forswear all defense projects. Work worries that might encourage other companies to make similar pledges. He also says that the Pentagon would still find companies competent in AI that are willing to help.
In part, thanks to companies like Google being open with their research and software, artificial intelligence expertise is more dispersed. “They clearly have other people to go to,” says Amir Husain, CEO of startup SparkCognition, which works on government AI projects, including with the Air Force. “The scale of artificial intelligence talent in the US is significant.”