Notes on the Trickery of Deepfakes

Kailyn Slater

Deepfakes remain an enigma to the news-literate public. Most confuse these deepfakes with cut-and-paste Photoshop video manipulation or Snapchat filters, rather than as algorithmically fabricated videos that utilize machine learning, found footage, and deep neural networks (Bates 2018). Communication scholars Chadwick and Vaccari argue that deepfakes are politically contextualized in ways that impact levels of uncertainty and trust in the news the public witnesses on social media. Understanding the very prescient presumption that a basic standard of truth can not be established due to the proclivity of false political rumors or “fake news” existing online, they conducted a survey utilizing the viral deepfake of actor Jordan Peele impersonating former President Barack Obama. In the video Peele is speaking as if he is Obama himself, easily recognizable as one of Peele’s classic impersonations and sitting at what is presumed to be the White House. The video begins with “Obama” in front of CGI, proclaiming calmly, yet pensively: “President Trump is a total and complete dipshit,” leaving viewers wondering if this was Obama himself talking. Obama is known for calling out Trump’s bullshit in his own neoliberal way, but to hear a possible representation of him say something this explicit and almost obvious about the current President is both visually and communicatively confusing. This eerily possible image of Obama begins to explain why Chadwick et al. chose to study this deepfake: this mechanism of trickery is easily translatable across contexts, and as such went viral after Buzzfeed News published an article about it in 2018.

While there are programs able to detect which aspects of these videos are artificially intelligent, they are not always reliable or accessible to the average user. Like most aspects of technocapitalism, the closest thing to accessible in this case requires the user to establish an understanding of artificial intelligence—most corporate projects are in testing phases, or are open-source amateur applications not bound by ethics. Programs like Google Duplex utilize machine learning within Alphabet’s data analytic framework to perform reverse engineering (Floridi 2018). Google Duplex claims it can detect the authenticity of a painting—a common example in AI research that is cringeworthy to those of us who know that paintings’ authenticity relative to the art world is determined by institutional capital — or fabricate an artificial voice for mobile assistants like Google Home (Floridi 2018). The latter aspect of Duplex, designed as a tool to literally assist by providing a voice from text, can conduct “natural” conversation to carry out “real world” tasks as well as distinguish between types of voices (Floridi 2018). Beyond commercially available technology, though, most viewers of deepfakes are not able to detect the difference unless it is visually obvious and in their face, or if they have seen enough CGI.

Deepfake applications are becoming more available and refined, as more videos go viral and gain circulation on the web. With creating content in mind, experienced programmers know that it’s easier to construct a machine of your own control with a Windows operating system. Windows machines have less barriers to entry for the average user to gain access to complex machine learning protocols, video rendering software, and back-end development compared to any Mac OS. Those who know how to make a deepfake application, theoretically, should then be conscientious of the knowledge acquired of internet protocols relative to the average user; unfortunately, these folks tend to opt-out of design and history courses in the process. DeepFaceLab, with almost five thousand forks on Github, states that they want this kind of technology to exist without the intention of “denigrat[ing] or ... demean[ing] anyone” (Github 2019). FakeApp is similar, but boasts a more user-“friendly” interface (Zucconi 2019). As these tools become more accessible, who will decide how deepfakes are received by the media, as silly memes made by amateurs to cast aside or criminally tampered evidence in a public trial? Will there be ethical committees formed? More importantly, do deepfake creators actually feel an ethical responsibility to do right by both the technology and the subjects they’re distorting?

Deepfakes had initially gained notoriety in the subreddit r/NSFWdeepfakes. Deepfake porn hasn’t categorically disappeared: they tend to dominate celebrity porn categories on PornHub, even when the site was instructed to ban deepfakes in 2018. Obviously, celebrities aren’t the only victims of non-consensual revenge porn: globally, deepfakes have been found depicting female politicians engaged in sex acts as a means to disparage their integrity (The Economist 2019). As deepfakes tend to utilize the likeness of celebrities (theoretically, anyone) using publicly accessible images and videos, they complicate the context of consent present in the video (Citron & Franks 2014). No one is able to consent to their likenesses in the video, as any of their likenesses were posted without permission from any management or studio. In any case, the specific demographic of deepfake creators has not been explicitly named; but based on the evidence provided we can assume they don’t think of women. Women do in fact favor the criminalization of revenge porn over men, as women and girls are by and large the primary targets of harassment online or off (Citron & Franks 2014).

We can ground the public’s frightened view on deepfakes in their unsettling, uncomfortable affect. This speaks to Marashiro Mori’s uncanny valley theory: a sense of eeriness, fear, or repulsion towards an object of near-human likeness, typically of prosthetics or humanoid robotics (Mori 2012). At first glance, deepfakes seem real, but to a fault. This fault, or the point at which difference can be detected, is obviously the product of an imperfect algorithm; the algorithm doesn’t define this as an error, but the human does. Mori’s theory performs the same way: the object of near-human likeness remains its own object; it is our human perception that creates fear or repulsion towards said object. The scholarship on the uncanny valley has varied, from understanding atypical feature response (i.e., responding to strange or ‘off’ features) to ideas about how viewers categorize and compartmentalize certain aspects of the human face (MacDorman et al. 2016). Videos that appear human, but are not quite, are uncanny because they appear to be attaining a realism that viewers attach to human features. These videos express a human affect that then confuses or frightens the viewer, such as the participants who could tell that it was Jordan Peele, not the former President, who called Trump a dipshit. While this theoretical concept is used to understand video games, films, or television, deepfake videos remain understudied and have the potential for insidious harm, considering the algorithms within the deep neural networks end up encouraging the uncanny themselves.

In their study, Chadwick and Vaccari rightfully point out that previous scholarship on disinformation has noted a creator’s “‘need for chaos’ [and] a desire to ‘watch the world burn’ without caring about consequences” (Petersen et al. 2018). As Chadwick et al. discuss, uncertainty concerning the validity of political information increases as trust in news on social media decreases, expounding deception. They argue that uncertainty in regards to intaking news on social media becomes systemic once an individual is exposed to multiple “contradictory, nonsensical, and disorienting messages that malicious actors introduce into digital discourse” (Chadwick et al. 2018; Phillips & Milner 2017). This is the root of the problem: a lack of understanding as to how deepfakes are made points to their public perception as inaccessible and deceptive, and, at an exponential scale, is uncanny. Deepfakes have been consistently typified as frightening or harmful, mainly due to their early notoriety as a means to fabricate non-consensual revenge porn (Bates 2018). Considering all of this, this typification ultimately stems from the nature of the content it is altering, and the censoring bodies hired to monitor them. As so much of our daily life functions through privatized metadata, it only seems logical to be weary of the slippery slope we face ahead regarding our trust in social systems, dealing with surveillance capitalism, and confronting the future grounded in facial recognition technology.

References