Illustration by Sagar Vikmani
In late 2017, knowing and understanding the term “deepfake” would have meant one thing: You were a regular browser of niche Reddit forums and a connoisseur of a very particular type of AI-generated celebrity pornography. Now, just a few years later, it’s a word synonymous with many of the regular obstacles of being online and using digital products: fake news, transparency, digital literacy, data protection, media manipulation, and disinformation—all things that threaten the fabric of what society believes to be true.
Now, “deepfake” refers to any audiovisual content created by artificial intelligence that impersonates a public figure or private individual. While pornographic content still accounts for the majority of deepfakes uploaded online, sexually explicit media isn’t the biggest challenge for designers, developers, and legislators.
Beyond confected sleaze, deepfakes have become most newsworthy in recent years for their threat to the democratic process. In April 2018, BuzzFeed created its own deepfake video to warn viewers of the potential of this new technology to be used for nefarious ends. Taking an existing video of former President Barack Obama at a formal public address, it mapped the mouth of comedian Jordan Peele onto his face and, using Peele’s finest Obama impersonation, issued a stark warning:
“We’re entering an era in which our enemies can make it look like anyone is saying anything at any point in time, even if they would never say those things… Moving forward, we need to be more vigilant about what we trust from the internet.” Preach, Peele.
Designing for a new era of trust
In their 2019 report for Data & Society, Deepfakes and Cheap Fakes, Britt Paris and Joan Donovan write that “currently, technologists, policymakers, and journalists are responding to deepfakes with calls for what scholars call technical and legal closures—that is, regulations, design features, and cultural norms that will determine the role of this technology.”
The UX design community is rallying to support these aims. “As designers of digital products for the next decade, we need to focus our efforts on designing for transparency and encouraging critical thinking from our users,” wrote Fabricio Teixeira in a January 2020 op-ed for UX Collective.
But what does transparency actually mean in the fight against deepfakes, and can improving society’s digital literacy suppress the sensationalized false narratives and compelling confirmation biases on which deepfake content thrives?
The answers are not immediately clear as technological development continues to outpace users’ grasp of what’s possible.
While the BuzzFeed video of Obama relied on Peele’s skills of vocal impersonation, software already existed that could render his role redundant. In 2016, Adobe gave its first public demonstration of Project Voco, a program capable of creating AI-generated speech from neural networks trained on just 20 minutes of a real voice. Voco was billed as the Photoshop for audio, but despite its huge creative potential, more column inches were made of its ability to cause harm.
“Inadvertently, in its quest to create software to manipulate digital media, Adobe has [already] drastically changed the way we engage with evidential material such as photographs,” said Dr. Eddy Borges Rey, a lecturer in media and technology at the University of Stirling to the BBC. “This makes it hard for lawyers, journalists, and other professionals who use digital media as evidence.” Perhaps in light of this, Voco has yet to be commercially released.
To lay the blame for deepfakes at the feet of any single piece of software is misguided and ignores the complex history of society’s relationship with “truth.” Numerous programs and platforms now exist that allow video, audio, and images to be doctored, but before their invention, photography was always a medium prone to falsification.
In Deepfakes and Cheap Fakes, Paris and Donovan argue that the new breed of audiovisual deepfakes are simply the next step in a continuum of evidentiary abuse that has undermined the veracity of audiovisual material since it first became admissible as evidence by law. Society has lived with doctored photos in courtrooms and in the media for decades, the only thing that’s different about today’s problem is the speed and scale at which misinformation can be distributed across a hyper-connected community.
Paris refers to this new phenomenon as “the libidinal use of information technology,” in which the sharing of information, real or fake, scratches some kind of innate itch within us. When we’re confronted with a piece of sensational media that marries with our own preconceived ideas about the world, we’re immediately compelled to share it far and wide.
Studies have shown that disinformation travels faster and further among online communities simply because it is fake.
Why share content that conforms to a banal worldview when a video of Facebook’s CEO admitting to his membership of a sinister global cabal could travel so much further?
“A lot of the time, people just don’t need things to be true before they share them,” says Paris. “People don’t necessarily see a video or photograph as evidence, just that it matches with a basic and deeply-held personal truth. Once a piece of content achieves a certain level of ubiquity, it seems to take on this character of truth.”
To date, human censors have struggled to tackle the problem of deepfakes and misinformation, despite attempts by some platforms to take measures against them. YouTube now insists on disclaimer copy specifying the source of each video to give viewers a clearer picture of the motives of content creators. The Guardian now dates all of its stories shared on social channels to prevent them from being recontextualized and used to fuel false narratives. Neither goes far enough.
Pending U.S. legislation known as the DEEPFAKES Accountability Act would mandate that any fake content uploaded to the web be labelled with a watermark that shows its fraudulence. “But I don’t know if that’s a very realistic solution,” says Paris. “The idea is that by requiring humans to comment and give notes on every upload that it would cut back on the volume of what’s being uploaded.” Such measures require a lot of good faith on the part of the creators of fake content, who have yet to show much concern for the chaos they’ve caused to date.
Ironically, a more sophisticated solution to the problem involves using neural networks, the same type of artificial intelligence behind the creation of deepfakes. At the Computer Vision and Machine Learning Lab at the University at Albany, Siwei Lyu has been training AIs to spot deepfakes since they first arrived on Reddit. His work is a constant game of cat and mouse—every time he trains his own systems to detect deepfake videos, fakers develop new and inventive ways to outsmart them.
Lyu’s initial methodology trained his neural networks to observe patterns in eye movement. Normal humans blink between every 2-10 seconds. AI-generated humans tend to blink far less. For a short time Lyu’s methods proved highly effective, resulting in a 95% detection rate, but when he published his research, deepfake creators changed their approach. Now his AIs look for artefacts left over from the first generation algorithms themselves—tiny details embedded in files that are undetectable to the human eye.
Despite advances in his research, Lyu believes the solution to deepfakes is neither technological nor legislative.
“User education and media literacy is the biggest issue here,” he says. “A lot of people can be fooled by these fake videos because they simply don’t know that videos can be manufactured and manipulated in this way. It’s like a virus. When a virus attacks a society for the first time, people get sick, not because they don’t know how to protect themselves, but because they don’t know that there is a virus. But once they know about the virus and take proper measures to protect themselves, then the virus can be controlled.”
How quickly the understanding of this particular virus improves depends not on individuals, but on platforms and their designers, all of whom have a responsibility to take definitive action. “The slow action on behalf of platforms to address any sort of misinformation happens primarily because it’s good for business,” says Paris. Lyu agrees. “It was only congressional pressure last year that made Facebook and Twitter take action.” It’s not hard to see why, in October 2019, senator Elizabeth Warren labelled Facebook a “disinformation-for-profit machine.”
Real action likely means addressing the fundamental design of systems that allow misinformation to spread, potentially impacting functionality and damaging revenue—moves unlikely to happen very soon. For digital designers that means “designing tools to filter out fake content, making users more aware of the treachery of deepfakes, and stopping the spread of misinformation,” writes Teixeira in his op-ed, “and raising awareness inside of our organizations, establishing principles around truth, and reporting how our platforms might be misused by agents with hidden agendas.”
While we wait for platforms to design responsible solutions to a problem they helped create, Paris offers a more mindful solution:
We all [think] a little bit more carefully about who we’re communicating with and why it might be helpful for society across the board.
Better still, if, as Lyu suggests, deepfakes do conform to the same patterns as organic viruses, perhaps it’s simply easier to self-isolate and switch off.