The Future of National Security Is AI, And AI Is Stats. The Facebook Testimony Shows How Unprepared Congress Is.
Invoking the term thirty-five times in two days of hearings, Mark Zuckerberg typified how the phrase “AI” has become a tool for the technical haves to bob and weave around questions from the technical have-nots. The Facebook CEO was treading on territory so well-worn it has a name: Tesler’s Theorem states that “AI is whatever hasn’t been done yet”, and each appeal Zuckerberg made to AI to solve a problem can be thought of as saying the solution is something that a machine likely but not yet definitely can do. The bounds of what machines can do has, and will continue, to grow largely from methods to make sense of torrents of data, or “big data”, which in turn relies heavily on machines deciding what parts of that data matter, or “machine learning”. Throughout Zuckerberg’s testimony, senators made clear that the future of our national security demands a grasp on the statistics that underpin the future of machines. Jihadi propaganda and Russian disinformation proliferate across Facebook today, enabling our adversaries to reach a global audience and achieve strategic objectives, and time and again senators could have laid that bare had they possessed a stronger hold on statistics.
We spent last summer immersed in the methods jihadis deploy on Facebook, trawling thousands of accounts by hand to find their profiles and propaganda. To meet our criteria, propaganda needed to originate with an official Telegram channel or dark web site of a jihadi group, and accounts needed both to share that propaganda and an image of the account owner with a weapon. The richness of Facebook as a source for those kinds of profiles cannot be overstated, and we quickly built a library of 1,200 profiles from thirty-five countries, helped along by Facebook’s recommendations, which tracked my browsing habits to recommend we browse more jihadist profiles, and stopping only when we had surpassed our goal by 20%. Today, 618 of the accounts remain active. Attributing the rest wholly to Facebook removals, the platform posts an abysmal takedown rate over nine full months of 48.5%.
“We’ve actually been very successful at deploying A.I. tools already,” Mark Zuckerberg told Congress, “Today, as we sit here, 99 percent of the ISIS and Al Qaida content that we take down on Facebook, our A.I. systems flag before any human sees it.” The questioning senator thanked the Facebook CEO, who would respond with the same metric twice more, twice more thanked for his work in fighting terrorism. But the ratio of manually flagged content to automatically flagged content tells us only that of the 582 removed accounts in our library, 576 were flagged by AI, a point wholly irrelevant to determining why 618 gun-toting jihadi-propaganda-spreading accounts have eluded content reviewers for three quarters of a year.
Facebook has empowered jihadis to achieve tactical goals like recruiting and spreading fear, and has even helped along strategic objectives like coercing civilian population to agitate for military confrontation. But their capacity to manipulate the platform for their own ends pales next to the resources a nation-state can bring to bare. Russian-backed trolls, semi-automated accounts, and bots used Facebook to spread fake news articles that would become four out of the five most viral Facebook posts of 2016.
An advertiser’s campaign achieving the same would have become Facebook’s call to prayer, recalled by sales reps five times a day and commencing and closing each and every shareholder meeting. Instead, as the hearings revealed, Facebook has shown no real movement to address the threat posed to democracy. Zuckerberg boasted of 740 removals of accounts traced back to the Internet Research Agency, but, like with ISIS and Al-Qaeda content removed, gave no estimate of the total. Radio Free Europe reported in 2015 there were no fewer than 800 IRA agents on Facebook, while Buzzfeed the year before obtained files showing each agent maintained at minimum six Facebook accounts, leaving a floor for IRA accounts on Facebook in 2015 of 2,400. By the time of the presidential election, the number, including bots, was at least 7,000, and was likely to be far greater.
“There were many others that our systems catch,” Zuckerberg continued, “which are more difficult to attribute specifically to Russian intelligence, but the number would be in the tens of thousands of fake accounts that we remove.” But Facebook has always invested heavily in fake account detection, and the bots to advertise online games, porno sites, and gambling pages that number in the tens of thousands and that Facebook has always removed are “more difficult to attribute specifically to Russian intelligence”. That Facebook would radically alter their current work to detect specifically IRA-built bots is unlikely. That they would do so, and then their CEO would carefully choose his words to leave open the possibility they hadn’t beggars belief.
Despite the constant appeals to AI, exposing the flaws in Zuckerberg’s testimony required no insight into hyperparameters in keras-net, vector multiplication in TensorFlow, or any of the impenetrable technical details of machine learning. Simply asking what share of the total a metric represents and why that’s the best metric for the problem at hand would have bypassed the headline numbers and led to a real debate. The ideal world is one where senior policymakers can wade into big data’s minutiae, but a prerequisite is the capacity to ask those questions. Had members of congress done so, the threat to national security posed by Facebook’s inaction would have been made clear.