The Consequences of Algorithmically Falsifiable Voice and Video for a Democratic Society

In 2017, the Justice Department charged thirteen individuals with interfering in American electoral and political processes. They had worked for the "Internet Research Agency," a well-funded Russian troll "factory" based in Saint Petersburg, which employs over four hundred people. Between 2015 and 2017, the agency is said to have created millions of tweets purporting to come from real people in America; their actions had spread and amplified both left and right wing conspiracy theories with the aim of sowing division. Other social media accounts created with the same aim had focused on fear mongering; some had even masqueraded as local news sources.

In 2018, NBC and FiveThirtyEight published large datasets of tweets from such accounts, many of which had been evicted from Twitter following the 2016 US presidential election. This archive gave an interesting insight into the tactics and content of an operation coordinated by a diffuse network of manipulators whose goal was to influence the narrative, the attention, and most likely the outcome of a national election.

Similar emerging tactics with various aims have popped up all over the world. In December 2018, the New York Times reported that for a period of time, half of YouTube traffic was "bots masquerading as people." Similarly, Facebook officially reports that up to 40 percent of its video engagement is bot-driven, though a recent suit filed by small advertisers asserts the figure is closer to 80 percent. The practice has become so common that bootstrapping marketing campaigns often involves manipulating engagement statistics. There are now companies that sell engagement such as likes, comments, and shares, which are fulfilled by "click farms" or bots. Today, one can buy five thousand YouTube views for as low as fifteen dollars.

Media platforms have of course pushed back against fake engagement, building algorithms to detect non-human traffic. They use machine learning to detect subtle patterns in behavior, such as mouse movements and time between clicks. If the patterns don't match those of typical humans, the clicks are rejected in the aggregate data. Naturally, fraudsters fight back, often with the same techniques used to detect them and simulate natural behavior to trick the filters. The two sides are now locked in an arms race. Just a few years ago, captchas, quick challenges only humans could solve, were used to help identify real humans from fake ones. Unfortunately, recent improvements in machine learning make it now increasingly harder to erect such barriers, because the gap in ability between man and machine is closing rapidly.

Another recent phenomenon capacitated through machine learning is porn with artificial faces, which often targets celebrities, created through the so-called deep-fake technique. The algorithm learns to perfectly map all expressions and view-angles of a given face, and "transplants" them to another allowing for a virtual face graft; the technology is very similar to the one I used to create the artwork in this project. This technique has rapidly proliferated on Reddit and porn sites; so much so, that in September 2018, Google added what it terms involuntary synthetic pornographic imagery to its ban list.

The use of this technique is not limited to celebrity pornography or revenge porn. In early 2018, for example, a video circulated on social media sites purporting to show Emma González, a pro-gun legislation activist, ripping up the US Constitution. The vide, fabricated using the deep-fake technique to create an impression of her being a traitor to the American ideals, did not aim to convince anyone across the political aisle, as much as to rile up those who already believe their Second Amendment rights are under attack. This shows that the use of this technique for political manipulation helps to polarize the population. What more, our general distrust in the media is eroding our trust in any objective truth, we become accustomed to question everything; if something, be it true or false, doesn't agree with our existing views and opinions, we discard it as fictitious. This effect is called the liar's dividend, in which a skilled manipulator can dismiss truth as false. For example, at a rally in July 2018, the president of the United States accused the mainstream media of propagating fake news, and exclaimed, "Just remember, what you're seeing and what you're reading is not what's happening."

The trend of faking individuals on a large scale in the media is having a profound influence on our collective beliefs, whether we lean towards the left or the right of the political spectrum. While the particular operation aimed at the 2016 US presidential election is often perceived to have tried to energize the Right in particular, a detailed analysis of the content of these tweets, content published by NBC and FiveThirtyEight, reveals that a significant fraction of the trolls aimed to provoke the Left into a frenzy by spreading fear and uncertainty. Overall, though, the goal of this coordinated activity was to create a more polarized society, or rather a significant Us vs Them mentality we are now all familiar with in public discourse. Note that the manipulation of the 2016 US electoral process was largely driven manually. By 2020, however, the technology for synthetic writing will have likely reached a level of believability such that similar tactics will be automatable, and will enable a very small number of people to orchestrate a very large and diffuse electoral process. In fact, in early 2018, OpenAI published a new algorithm for the generation of text, which the authors believed to be so realistic that OpenAI decided not to openly publish the model at all, for fear of misuse in the ways described earlier.

What struck me when observing all these present-day trends and novel techniques was the deception of identity at their very core. Historically, humans have always been able to easily verify someone's existence by interacting with them face to face. Nowadays, because of the disjoined way that we experience others' accounts on social media, especially public social media like Twitter, these fake accounts don't have to be particularly sophisticated to appear real. A simple avatar and an apparent location in the United States suffice. Unlike in a top-down approach of classic propaganda, here people are conned into believing that these fake identities are in fact true individuals who are "like them" and who espouse certain views. Prior to global social media, it was impossible for a foreign state to effectively access Americans in the United States; top-down information distribution channels such as newspapers and television are usually well under the influence of the government. It would have been moreover cost prohibitive for a foreign nation to send tens of thousands of spies to pretend to be normal people, to make friends with the locals, and then to start subtly changing the conversation. Today, however, social media make it an easy proposition by opening up a new way for bottom-up manipulation, which can be orchestrated from abroad, and which has already proven itself to be remarkably effective at sowing division and distrust of one's own country. The term fake news deliberately obscures the fact that the disinformation is spread with very specific intents. What we have here is not simply fake news, but rather a new form of coordinated propaganda.

The idea of imaginary people, central to the bottom-up approach, inspired me to explore this space artistically. Any collective belief, be it religious, economic, or political, can effectively become "real" if that fictitious entity affects the actions of the collective. In that sense, these imaginary people who express their views on social media really exist in the same way that deities do, as long as a group believes in their existence and reality. I was therefore curious to speculate who these fictitious humans might be or what they might look like, which became the inspiration for algorithmically generating faces to create portraits of imaginary people. In recent years, computer graphics algorithms, especially generative neural networks, have become rather proficient at producing and manipulating media such as image, video, and sound, to a degree that has become hard to distinguish from their real counterparts. WaveNet, an algorithm that models raw audio waveforms, is capable of producing extremely realistic human speech. The DeepDream algorithm, released in 2015, showed that a computer system that is good at perceiving image data can also be inverted to generate images. The Generative Adversarial Network (GAN), a machine-learning algorithm published in 2016, has been refined to specifically generate new images based on a set of training images. While tweets from 2016 predate these techniques, the appearance of such technologies gives us a taste of what's to come. As discussed earlier, fake identities and their subsequent opinions, also fabricated , were created and spread in 2016 by a small group of dedicated human beings; nowadays, however, we can expect this sort of work to become rapidly automated. Likewise, those who try to separate wheat from chaff are also increasingly turning to sophisticated machine-learning algorithms to find and filter inaccurate or fake content. We thus have now an arms race which is already affecting the nature of the internet, how we build our beliefs of the world around us, how we delegate trust in society, and what we base our own actions on.

Feb 2019, Mike Tyka