Abraham Gibson

The video is unmistakable and unforgettable. The grainy footage shows President Richard Nixon taking a seat and steeling his nerves. It was to be the most important speech of his career, one that he had hoped he would never deliver. “My fellow Americans,” he begins in a somber tone, “fate has ordained that the men who went to the moon to explore in peace will stay on the moon to rest in peace.” The implications were clear enough. Something had gone terribly wrong with Apollo 11, the first NASA mission to attempt a lunar landing, and astronauts Neil Armstrong and Buzz Aldrin were dead. “For every human being who looks up at the moon in the nights to come will know that there is some corner of another world that is forever mankind.” With these words, Nixon sought to console a grieving nation during one of the most tragic moments in its history.

Except the tragedy never happened. To be sure, the text of the speech was real. It had been written by Nixon’s speechwriter, William Safire, in 1969, in case something went wrong. Thankfully, the mission succeeded, and the speech was not needed. It went unread until 2019, when researchers at the MIT Center for Advanced Virtuality resurrected the script (and the president’s visage) for a special screening at the 2019 International Documentary Film Festival. The short film, In Event of Moon Disaster, was produced on a computer using video dialogue replacement technology powered by artificial intelligence (AI). The filmmakers sought to emphasize both the power and the danger of synthetic audiovisual media, otherwise known as “deepfakes” (Panetta and Burgund).

In general usage, the word deepfake refers to any video that has been altered using machine-learning algorithms to produce hyperrealistic videos that show actual people saying and doing things they never said or did. Some videos seamlessly transplant one person’s face onto another person’s body. Others create an entire person from scratch. In both cases, the algorithmic alterations are difficult, if not impossible, to detect with the naked eye. There are now several mobile apps that help users with no programming skills produce synthetic videos with face-swapping and artificial lip-synching, but most researchers do not consider these videos deepfakes since they do not use machine learning (Paris and Donovan). In late 2019, cybersecurity officials identified more than 14,000 deepfakes on the internet, a 100 percent increase in just seven months (Ajder et al.). The number has no doubt risen even higher since then.

The first deepfake to gain widespread attention was released in 2018. The video appeared to show Barack Obama saying, “Donald Trump is a total and complete dipshit.” In fact, the video was an effective public service announcement from actor and director Jordan Peele about the dangers of deepfakes. Acting as an invisible marionette, Peele used the Obama avatar to caution viewers that deepfakes could lead to “some kind of fucked-up dystopia” (Mack). In the years since, several other deepfakes have also gone viral. Perhaps you have seen the one that shows Bill Hader morphing into Arnold Schwarzenegger, or the one that places Will Smith’s face on Cardi B’s body. It isn’t just videos, by the way. Last year, a British energy company was duped out of nearly a quarter-million dollars when a scammer used AI to imitate the CEO’s voice and request a wire transfer over the phone (Stupp). Some have expressed fear that deepfakes herald nothing less than the “information apocalypse” (Silverman), the very “collapse of reality” (Foer).

Efforts to counter the deepfake menace have so far fallen into one of two categories. First, many in the tech sector promote technological solutions, especially automated detection. There are practical problems with this approach, however. After all, the automated detectors train the automated generators to fix their mistakes, which means that subsequent deepfakes may evade detection altogether. Other potential technological solutions include provenance stamps, geolocation tags, blockchain verification, and mandatory registration of content creators (Knight; Paris and Donovan; Chesney and Citron). Second, there are legal options. For example, China recently banned deepfakes outright. Germany has also passed stiff laws penalizing media companies that fail to remove racist or threatening content from their platform in an expeditious manner, and some believe that a similar tactic might work with deepfakes. At the present, however, most places have no laws to handle deepfakes, which means that legal challenges to deepfakes will be handled in a costly and inefficient case-by-case basis (Citron and Chesney; Woollacott).

Technological solutions and legal challenges are limited because they are post hoc fixes to larger, more complex sociotechnical systems. By comparison, the digital humanities (DH) are uniquely interested in both the social and the technical. Practitioners are trained to describe complicated systems, including their impact on society and culture, and can thus help support structural change. Ten years ago, the digital humanities might not have been ready to meet the deepfake challenge. The field was still young, still coalescing, still largely preoccupied with what qualifies as the digital humanities. The field entered a second phase around five years ago, as scholars sought to establish connections between the digital humanities and other fields of inquiry across campuses and throughout the academy. We have now entered a third phase, and not a moment too soon. This latest iteration insists that the digital humanities must embrace social activism, that we must move beyond the campus and engage the public at all costs. Rather than framing the digital humanities around advances in digital scholarship, Safiya Umoja Noble writes in Debates in the Digital Humanities 2019, perhaps we should interrogate how digital humanities too often obscure important features of the social, political, and economic landscape. In that same volume, Matthew K. Gold and Lauren Klein are more explicit still: “Our work within the digital humanities is enabled by larger social, political, and technological systems. In the present moment, we need work that exposes the impact of our embeddedness in those larger systems and that brings our technical expertise to bear on the societal problems that those systems sustain.”

First and foremost, scholars in the digital humanities should broadcast our core belief that technologies are best understood as sociotechnical systems. This broader view acknowledges that technology cannot be divorced from its social context. It encourages the citizenry to expect and, if necessary, demand answers about the provenance of any given technology. Rather than trying to identify and stamp out every deepfake on the web, digital humanists can help identify the deeper structural issues that allow tens of thousands of deepfakes to proliferate in the first place. We can advocate for reform, rejecting false creeds like “techno-solutionism” that promote technology as humanity’s savior (Morozov). As Catherine D’Ignazio and Lauren Klein explain, “we must look to understand and design systems that address oppression at the structural level.”

We can also help demystify deepfakes by showing that previous technologies engendered similar concerns (Pyne 2019). Take the printing press, for example. Several writers have shown that the rise of pamphlet culture in seventeenth-century England facilitated widespread disinformation during the first English Civil War (White; Peacey). The invention of photography was no less revolutionary, and it too facilitated disinformation. Famous examples from the American Civil War include the transposition of Abraham Lincoln’s head on to John C. Calhoun’s body, the insertion of Ulysses S. Grant into a battle scene near City Point, and Matthew Brady’s decision to rearrange dead bodies of soldiers to set up visual tableaux (Trachtenberg). A century later, Joseph Stalin famously ordered the mass alteration of photographs in Soviet Russia in an effort to purge his enemies from the history books (King). Meanwhile, the ability to deceive using audiovisual (AV) manipulation was, until recently, the stuff of Hollywood. Jurassic Park and Forrest Gump were among the first, using AV manipulation to resurrect dinosaurs and presidents, respectively. The visage of the late Peter Cushing was used in Rogue One, and there are plans to posthumously cast James Dean in a drama about the Vietnam War (Ritman).

These famous examples notwithstanding, the overwhelming majority of extant deepfakes are nonconsensual pornography. In fact, the word “deepfake” was coined in November 2017, when an anonymous Reddit user named “deepfakes” shared a software toolkit that would allow anyone to make synthetic videos replacing one person’s face with another. To demonstrate, the user posted a manipulated video that appeared to show actress Gal Gadot in a pornographic film (Rothman). Since then, deepfake pornography has exploded in popularity. In fact, one recent census found that approximately 96 percent of deepfakes on the internet are pornographic, and that approximately 99 percent of these fake videos target women (Ajder et al.).

None of this will come as a surprise to scholars in the humanities. As is well known, the pornography industry has often served as a catalyst for technological change. It has also been one of the most reliable early adopters for any new medium. Pornography played a crucial role in development or at least the expansion of Polaroid cameras, handheld camcorders, video cassette recorders (VCRs), cable TV, and the internet (Barss; Coopersmith). Deepfakes might seem like just another connection between pornography and tech, but they are much different. When it comes to deepfake pornography, the “participants” don’t even know it is happening. Celebrities are not the only target, either. Deepfakes can now be weaponized against anyone, including ex-lovers, professional rivals, and even random strangers, none of whom consented to “star” in a pornographic video.

Countless studies in the digital humanities and beyond have shown that technology has a disproportionately negative impact on women and other minoritized groups. Scholars have demonstrated time and time again that digital technologies are more likely to reify existing power structures than undermine them, that digital infrastructures invariably have an outsized impact on underrepresented groups, and that the digital humanities can help us elucidate structural inequalities in the nondigital world (Benjamin; Noble, Algorithms of Oppression; Broussard; Eubanks; Losh and Wernimont; Hicks; O’Neill; Gallon). As digital humanists, we should advocate for equity, dignity, and structural change.

Deepfakes will also create challenges that are unique to specific domains within DH. Consider their impact on digital history, for example. Some outside the field have suggested that deepfakes will be good for historians. Legal experts Danielle Citron and Bobby Chesney both herald the potential pedagogical value that deepfakes offer historians, writing that deepfakes make it possible to “manufacture videos of historical figures speaking directly to students.” Sure enough, programmers are already using OpenAI’s Generative Pre-trained Transformer-3 (GPT-3) to simulate correspondence with historical personalities from any era. Once again, the practical applications are not limited to famous people. For example, Microsoft filed a patent that would allow the company to digitally “resurrect” the recently departed in the form of chatbots and eventually deepfake avatars (Smith).

Even so, the preponderance of evidence suggests there are legitimate reasons for concern. First, malicious actors already mischaracterize the historical record in service of their agenda. Studies have shown that white supremacists routinely co-opt and misrepresent research on genetic genealogy, classical antiquity, and Viking genomics to suit their racist worldview (Panofsky and Donovan; Nelson; Strand and Källén). One can only imagine how neo-Confederates and neo-Nazis might like to revise the historical record with help from deepfakes. Second, deepfakes will generate different reactions depending on the context in which they are viewed. Consider the Nixon deepfake. While it was first displayed within the context of a museum exhibit, the film is now available on YouTube. Given that millions of Americans already doubt the moon landings ever took place, it is easy to imagine someone thinking that the deepfake is real, especially when it is divorced from its original context. Third, many of the strategies to combat deepfakes, from litigation to “life-logging” (Citron and Chesney; Eggers), presume that the target is still alive, but who speaks for the dead? Peter Cushing never consented to star in Rogue One, and President Nixon never eulogized Apollo 11. Did each surrender rights to his visage when he died? Does everyone? These questions are ripe for humanistic analysis of agency, power, authorship, and intellectual property, among other topics.

Historians who study pre-twentieth-century history, when moving-image technology was largely nonexistent, might think that they will be spared the worst effects of deepfakes, but generative adversarial networks (GANs) leave all media vulnerable to digital manipulation.1 This includes everything from photographs to cave art. In fact, the first piece of AI-generated art (a blurry portrait of a nonexistent person, created by GANs that were trained on more than 15,000 portraits) sold at Christie’s auction house in 2018 for more than $400,000 (Cohn). GANs can also read and manipulate digitized texts, which are the foundation of modern historical research. Combining machine learning with image processing techniques, researchers have built optical character recognition (OCR) engines that can read typeface and, increasingly, handwritten text (Mermon et al.). In fact, AI allows machines to not only read handwriting but also produce handwriting in any language. How will historians discern genuine primary sources from fake ones? Ironically, in an increasingly digitized world, we will be called on to vouchsafe the physical archives and safeguard authenticity like never before.

This is more than an academic debate. To quote George Orwell’s prescient aphorism, “Who controls the past controls the future, who controls the present controls the past.” Authoritarian regimes will not hesitate to suppress and misrepresent the historical record to retain power. Glenn D. Tiffert explains how officials in Communist China recently asked Cambridge University Press and Springer Nature to hide more than a thousand potentially offensive articles on Chinese history from users in China. Keen to remain in business with the most populous nation on earth, the publishers quietly acquiesced. For subscribers in China, the items simply disappeared from search results, as if they never existed. Such large-scale censorship would have been impossible a generation ago. Now, it can be done with a few clicks.

Tiffert drew his evidence from Communist China, but the past few years have shown that democracies are also vulnerable to authoritarianism. In the United States, Donald Trump actively promoted widespread disinformation throughout his presidency. He routinely called fake news real and real news fake. He often retweeted “cheapfakes,” relatively crude AV manipulations produced with mobile apps rather than artificial intelligence, and given that his only criterion for truth was whether or not the information flattered him, he would have undoubtedly promoted any deepfake that cast him in a positive light. Trump was eventually kicked off Twitter after he incited a deadly riot at the U.S. Capitol, but democracies remain vulnerable to disinformation in general and deepfakes in particular. Malicious actors are known to use fake profiles on social media using AI-generated faces (Gleicher), and research has shown that microtargeted deepfakes can have real and demonstrable effect on political attitudes (Dobber et al.).

To properly confront the deepfake challenge, DH scholars may find it necessary to use both computational methods and traditional methods of scholarly research. There is a rapidly growing body of scholarship on the computational humanities in general (Johnson, Mimno, and Tilton; Afanador-Llach et al.; Graham, Milligan, and Weingart; Arnold and Tilton, Humanities Data in R) and the “visual turn” in DH in particular (Tilton). Recent advances in machine learning and computer vision have transformed how researchers engage with visual media. Algorithms for machine vision can detect, identify, and qualify features within an image and then build predictive models that it can apply to future datasets. A growing number of scholars use both computational and humanistic methods in their analyses of visual media, thereby signaling potential ways of engaging with deepfakes from a DH perspective (Arnold and Tilton, “Distant Viewing”; Wevers and Smits; Lee 2021; Mittell). Meanwhile, just as DH scholars must be willing to use computational tools, so too must we insist that fields like computer science integrate humanistic perspectives. Doing so would bring critical perspectives to bear on technological advances. We need, in other words, more multidisciplinary research in the technology sector. DH intervention does not necessarily mean that every DHer should start researching deepfakes. Instead, we should advocate for embedding humanities perspectives in the tech world to anticipate problematic aspects of tech research.

In closing, DH scholars should bear three lessons in mind as we brace for potentially ubiquitous deepfakes. First, we should promote digital literacy at every turn. We do not necessarily need to learn code, but we do need to understand how algorithms influence every part of our daily lives if we are going to successfully navigate the cacophony of disinformation (McPherson; Schmidt). Second, DH scholars must also embrace advocacy. Deepfakes threaten everything from interpersonal relationships to international relations, but they are especially injurious toward women and other minoritized groups. Rather than resigning ourselves to nihilism, we can advise our neighbors to seek answers when confronted with a provocative video on social media. What is the source of this video? Who might have created it and why? Finally, we should recommit to values that have long defined the digital humanities. DH scholars have suggested several core principles to help unify the field, including openness, collaboration, collegiality, connectedness, diversity, experimentation, hope, compassion, and empathy (Nowviskie; Spiro; Noble, “Toward a Critical Black Digital Humanities”; Gold and Klein). Keeping the faith will not be easy. Deepfakes breed distrust, and fascism weaponizes credulity (Tiffert), but that only underscores the importance of the DH perspective. If we are earnest about our values, then the digital humanities can help the larger public stay resilient, rebellious, and real.