Leaked Emails Show Google Expected Lucrative Military AI Work to Grow Exponentially
Lee Fang | The Intercept | Source URL
FOLLOWING THE REVELATION in March that Google had secretly signed an agreement with the Pentagon to provide cutting-edge artificial intelligence technology for drone warfare, the company faced an internal revolt. About a dozen Google employees have resigned in protest and thousands have signed a petition calling for an end to the contract. The endeavor, code-named Project Maven by the military, is designed to help drone operators recognize images captured on the battlefield.
Google has sought to quash the internal dissent in conversations with employees. Diane Greene, the chief executive of Google’s cloud business unit, speaking at a company town hall meeting following the revelations, claimed that the contract was “only” for $9 million, according to the New York Times, a relatively minor project for such a large company.
Internal company emails obtained by The Intercept tell a different story. The September emails show that Google’s business development arm expected the military drone artificial intelligence revenue to ramp up from an initial $15 million to an eventual $250 million per year.
In fact, one month after news of the contract broke, the Pentagon allocated an additional $100 million to Project Maven.
The internal Google email chain also notes that several big tech players competed to win the Project Maven contract. Other tech firms such as Amazon were in the running, one Google executive involved in negotiations wrote. (Amazon did not respond to a request for comment.) Rather than serving solely as a minor experiment for the military, Google executives on the thread stated that Project Maven was “directly related” to a major cloud computing contract worth billions of dollars that other Silicon Valley firms are competing to win.
The emails further note that Amazon Web Services, the cloud computing arm of Amazon, “has some work loads” related to Project Maven.
Jane Hynes, a spokesperson for Google Cloud, emailed The Intercept to say that the company stands by the statement given to the New York Times this week that “the new artificial intelligence principles under development precluded the use of A.I. in weaponry.” Hynes declined to comment further on the emails obtained by The Intercept.
THE SEPTEMBER EMAIL chain discussing the recently inked deal included Scott Frohman and Aileen Black, two members of Google’s defense sales team, along with Dr. Fei-Fei Li, the head scientist at Google Cloud, as well as members of the communications team.
Black provided a summary of the Project Maven deal, which she described as a “5-month long race among AI heavyweights” in the tech industry. “Total deal $25-$30M, $15M to Google over the next 18 months,” she wrote. “As the program grows expect spend is budgeted at 250 M per year. This program is directly related to the Sept 13 memo about moving DOD aggressively to the cloud I sent last week.”
The September 13 memo sent by Black was not included in the emails obtained by The Intercept. It appears to be a reference to the Joint Enterprise Defense Infrastructure, a contract worth $10 billion over 10 years, a project that Google has expressed an interest in obtaining. The JEDI program was announced on September 12.
The project had finally come together and was moving along rapidly, Black wrote. The Pentagon was “really fast tracking” Google’s cloud security certification, a development she called “priceless.”
The Google executives discussed the potential for a public relations fiasco from the Project Maven contract. Whether or not to reveal the deal was a point of concern.
“This is red meat to the media to find all ways to damage Google. You probably heard Elon Musk and his comment about AI causing WW3,” wrote Fei-Fei.
“I don’t know what would happen if the media starts picking up a theme that Google is secretly building AI weapons or AI technologies to enable weapons for the Defense industry,” she continued. “Google Cloud has been building our theme on Democratizing AI in 2017, and Diane and I have been talking about Humanistic AI for enterprise. I’d be super careful to protect these very positive images.”
The Google team noted that it has no press plan for the rollout of the contract and agreed that the company should work to set the “narrative” as quickly as possible. The “buzz” generated by the contract could be a positive, Black suggested.
The government sales team noted that Project Maven had been concealed through a contract awarded to ECS Federal, an arrangement first reported by The Intercept.
“The contract is not direct with Google but through a partner (ECS) and we have terms that prevent press releases from happening without our mutual consent,” wrote Black. The Defense Department “will not say anything about Google without our approval.”
Despite the secrecy, Black cautioned that news will eventually leak and that information about the contracting process could be obtained by the public through the Freedom of Information Act. Google’s involvement with Project Maven “will eventually get out,” Black warned. “Wouldn’t it be best to have it released on our terms?”
The project, however, was never announced publicly until news broke in March 2018.
AN INTERNAL WORK timeline about Project Maven, also obtained by The Intercept, provides a window into the quick progression of the contract.
On October 27, 2017, a team from Google Cloud visited Beale Air Force Base — a major hub for drone pilots — to “meet operational users (Air Force data analysts) who will be the end users of our technology, and primary testers starting June 2018.”
The previous week, Lt. Gen. John N.T. “Jack” Shanahan, who helped spearhead Project Maven, visited Google’s Advanced Solutions Lab to meet with 50 members of the team working on the project. Shanahan declared that “nothing in DoD should ever be fielded going forward without a built-in AI capability,” according to the timeline.
The timeline describes how Google engineers were continually working with the military to improve the product, including the user interface. “While the initial core technology focus will remain detection, classification, and (limited) tracking of certain classes of objects, we are considering how to address customers’ concern regarding more challenging use-cases that solve user’s real problems,” the document notes.