Likewar: The Weaponization of Social Media

by P.W. Singer

be found in other machines. Recent breakthroughs in neural network training hint at what will drive machine evolution to the next level, but also save us from algorithms that seek to manipulate us: an AI survival of the fittest. Newer, more advanced forms of deep learning involve the use of “generative adversarial networks.” In this type of system, two neural networks are paired off against each other in a potentially endless sparring match. The first network strains to create something that seems real—an image, a video, a human conversation—while the second network struggles to determine if it’s fake. At the end of each match, the networks receive their results and adjust to get just a little bit better. Although this process teaches networks to produce increasingly accurate forgeries, it also leaves open the potential for networks to get better and better at detecting fakes. This all boils down to one important, extremely sci-fi question. If both networks are gifted with ever-improving calibration and processing power, which one—the “good” AI or the “bad” AI—will more often beat the other? In the answer quite possibly lies not just the fate of content moderation policy, but also of future wars and elections, as well as democracy, civilization, and objective reality. Within a decade, Facebook, Google, Twitter, and every other internet company of scale will use neural networks to police their platforms. Dirty pictures, state-sponsored botnets, terrorist propaganda, and sop...