Abstract
Deepfakes are an effective method of media manipulation because of their realism and also because truth is not a priority when people are consuming and sharing content online. Consumers are more focused on creating their own reality that aligns with their desires, opinions, and values. We explain how deepfakes differ from other sources of information. Their realism and vividness makes them unusually effective at depicting alternative facts, including fake news. Deepfakes are difficult to detect and will be even harder to detect in the future. However, people share deepfakes not necessarily because they believe them but because they want to reinforce their own identity and social position. The threat posed by deepfakes is that they can radicalize people by sowing chaos and confusion. They rarely change minds. We review the consequences of deepfakes in both the social sphere and private lives. We suggest potential solutions to reduce their negative consequences.
Keywords
Artificial intelligence is no longer mere fantasy. Among the unexpected and sometimes spine-chilling achievements in computing are deepfakes. Deepfakes are hyperrealistic digital impersonations or falsifications of images, video, and audio created through neural networks using a machine-learning technique called a generative adversarial network (GAN). GANs involve teaching a neural network to generate outputs that fool its adversary, another neural network whose job is to discriminate the real from the fake. The adversary looks for flaws and discrepancies produced by the deepfake over multiple rounds, finding editing errors in the generation of an exact likeness of a person (or anything else) that may not be noticeable to a human, and the GAN learns how to minimize those errors. By receiving feedback from the adversary many times, the generative network learns to fool the discriminator and, it turns out, to also fool people, allowing the fabrication of an event that never actually occurred.
Deepfakes appear in a wide range of contexts, from arts and entertainment to advertising and education. By far the most frequent application of deepfakes is in pornography; as of October 2019, 96% of deepfakes on the Internet were pornographic (Ajder et al., 2019). Nevertheless, it is their actual and potential contribution to the epidemic of fake news that has engaged academia and the media. Text transmits meaning symbolically and requires interpretation, but images and videos carry meaning by appearing to represent real life directly (Vaccari & Chadwick, 2020). Numerous deepfakes have appeared of politicians making claims that contradict their actual position, such as one uploaded to a hacked Ukrainian news website of Volodymyr Zelenskyy, president of the Ukraine, telling his soldiers to lay down their arms (Allyn, 2022) or Barack Obama using profanity toward Donald Trump (Silverman, 2018). If such deceptions quickly went viral, they could have an irreversible effect on world affairs.
Humans process visual data naturally and thus fluently. Visual learning is remarkably efficient (Prior, 2013). More critically, people believe what they see (Messaris & Abraham, 2001), suggesting that visual images have stronger persuasive power than text. Moreover, the detailed imagery of deepfakes has the potential to prime psychological proximity. Concrete misinformation (including disinformation) primes participants to think of events as being nearer and more probable, increasing their perceived threat (Rim et al., 2015) and likelihood of being shared (Kirkpatrick, 2020).
How Effective Are Deepfakes?
Studies have shown divergent findings about whether people are able to discriminate deepfakes from real images. Compared with traditional sources of fake news, such as text or audio, deepfakes have been found to be no more credible (Barari et al., 2021; Vaccari & Chadwick, 2020) or effective at implanting false memories (Murphy & Flynn, 2021) than their long-established counterparts. However, although these studies are recent, the technology is changing so fast that the studies did not deploy the most recent artificial intelligence (AI)-based image-creation technologies available online. Some studies of deepfake credibility use only a single video (Vaccari & Chadwick, 2020), bringing into question their applicability to the diverse environment of AI-generated content.
As demonstrated by Lago et al. (2022), although images generated by earlier forms of GANs can be distinguished from real images by the human eye, newer AI-synthesized images are perceived as real. Indeed, synthetic faces generated by the most state-of-the-art GANs (such as those produced by StyleGAN2) are judged as more real than real images (Lago et al., 2022), pointing to the potential of deepfakes to simulate reality and circumvent the eerie, unsettling feeling that arises when humanoid robots or computer-generated images are too close to the real thing (the uncanny-valley effect associated with synthetic faces; Mori et al., 2012). Even more convincing, Köbis et al. (2021) show that people cannot reliably detect deepfakes and that neither raising awareness nor introducing financial incentives improves their detection accuracy. They also found that people are more likely to mistake deepfakes as authentic videos than to make the opposite error and that they overestimate their detection abilities. An even more recent study (Lee & Shin, 2022) found that deepfakes are both more believable than fabricated images and text and that people are more likely to engage with them. In summary, although people may be able to distinguish some existing deepfakes from real images, rapidly evolving GAN technology will soon render deepfakes indistinguishable from genuine content if it has not already.
Do People Care About Accuracy?
Deepfakes remain a threat to society because of their potential to be circulated regardless of their believability. As it turns out, information veracity is not a deciding factor when users choose to share content online (Pennycook et al., 2021). Vosoughi et al. (2018) found that fake news is diffused faster and further online than factual information, reaching more users on Twitter. In fact, social media users are less likely to distinguish true from false information when indicating what they would share on social media compared with when directly asked about accuracy (Pennycook et al., 2020). This implies that even when users can identify deepfakes as untruthful, they still might share them within their social circle. If veracity is not the predominant driver of the consumption and shareability of deepfakes, what is?
A hint about why people share and view deepfakes can be found where they are most common. The term deepfake was coined by a Reddit forum created for sharing pornographic videos of women whose faces were synthetically swapped for those of others, mostly celebrities (Ajder et al., 2019). The appeal of pornographic deepfakes may be sexual pleasure, fantasy, to exert power and control over women, or some combination of these factors. Consumers of pornographic deepfakes are unlikely to be fooled by the image they are watching because the website or video title is typically marked as fake; there is no pretense of truth. Thus, viewers of pornographic deepfakes obtain whatever benefits they get despite or because of their knowledge that what they are watching is fake. This could also be true of political deepfakes. When people share political deepfakes, they need not care about accuracy; they may understand that they are viewing fiction.
Other uses of deepfakes similarly do not depend on their veracity. Art need not depict unadorned reality, and educational purposes can be obtained with (say) reenactments and simulations that do not mimic reality in all respects (LaMarre & Landreville, 2009). Deepfakes also allow creators to produce media that more closely resemble a narrative than the actual event, making them more appealing. A deepfake, like all counterfeits, is not constrained by the facts and presents an opportunity to offer whatever information is most engaging to its audience. This means they could also serve a positive, pedagogical purpose.
Most fake news is deployed for the purpose of getting ad revenue. It is reasonable to assume that most deepfakes on social media will be there to attract as many clicks as possible. To achieve that, they exploit two properties that engage people’s attention, novelty and negativity (Chesney & Citron, 2019). Sharing novel facts holds social value because it suggests that the sharer holds inside information, a marker of social status (Groh et al., 2021).
The tendency to attend to negative information is well known. People attend more to potential losses than to gains (Yechiam & Hochman, 2014), and this negativity bias may apply broadly. Sharing negative information has the veneer of nobility by warning other people of potential threats. Health professionals have been found to be more willing to retransmit false rumors to prevent negative repercussions (e.g., causing cancer) than to produce positive outcomes (e.g., curing cancer; Kirkpatrick, 2020). Negative events are more likely to be transmitted across partners than positive events (Bebbington et al., 2017). Therefore, negative and novel deepfakes also have the potential to sow more chaos and generate stronger emotional reactions (Vosoughi et al., 2018) than positive, familiar news.
People share information to make contact with other people, and those with whom they share a political ideology will be most receptive to their political missives (Sloman & Fernbach, 2017). Sharing with one’s ideological community not only satisfies a fundamental human motivation to strengthen one’s social attachments (Baumeister & Leary, 1995) but also confirms one’s identity as part of an ideological group (e.g., Tajfel, 1974). Protecting self-identity takes priority over judging accuracy (Zhou & Shen, 2022). Indeed, identity will color what people consider to be true. People will attribute greater believability to information that aligns with their nationality, religion, race, or political party (Liv & Greenbaum, 2020). But because people tend to live in information bubbles, we rarely encounter inconvenient information, so partisan belief differences generally demonstrate an ignorance of inconvenient truths rather than an acceptance of falsehoods (Sloman & Fernbach, 2017). Therefore, it is reasonable to expect that people may see and share a deepfake video that aligns with their beliefs while never coming across any reason to believe that it is, in fact, a deepfake. In one study showing subjects faked photographs, conservatives were more likely to “remember” Barack Obama shaking hands with the president of Iran, whereas liberals were more likely to “remember” George W. Bush on vacation with a celebrity during Hurricane Katrina (neither event actually happened; Frenda et al., 2013). More to the point, deepfakes have been shown to radicalize people against the opposition (Dobber et al., 2021). Like other information sources, deepfakes may be more likely to radicalize existing views than to change people’s opinions.
Some people are more likely to share deepfakes than others. Older people are more likely to be deceived by deepfakes (Caramancion, 2021). The same study showed that political ideology influences how deepfaked news is evaluated. Although Republicans and Democrats are equally inclined to share fake news (Harper & Baguley, 2022), low-conscientiousness conservatives (those least likely to follow societal norms for impulse control) are the most likely to share misinformation because of their desire for chaos (Lawson & Kakkar, 2022). Even when presented with fact-checking warnings, low-conscientiousness conservatives shared fake news more frequently than liberals and high-conscientiousness conservatives.
Consequences
The picture we present of deepfakes is summarized in Table 1. The greatest amount of psychological pain from deepfakes has surely been endured by women whose image had been stolen to produce pornographic content (Ajder et al., 2019). The personalized nature of deepfakes adds a new layer of emotional distress for victims, as they threaten women’s relationships, careers, mental health, and physical safety through blackmail, harassment, and extortion (Harris, 2019). Most pornographic deepfakes present celebrities whose reputations may provide a degree of shelter from being seen as the genuine subject of the video. They also possess public platforms, as well as legal and financial means to dispute the veracity of the videos. Ordinary, private citizens, seen in cases of revenge porn, do not have even these limited protections (Gieseke, 2020). Moreover, if a video is taken down from a particular platform, a deepfake may be copied and continue to circulate on other websites, damaging the subject’s well-being, reputation, and job opportunities (Gieseke, 2020).
The Deepfake Threat
Even when citizens do not believe the misinformation presented to them or are not concerned about truth, deepfakes can increase uncertainty about content and decrease trust in media, preventing the establishment of a shared reality and allowing misinformation to flourish (Vaccari & Chadwick, 2020). In the United States, fake news caused 50% of Republicans and 38% of Democrats to reduce the amount of news they consume (Mitchell et al., 2019). As COVID-19 exemplified, in times of crisis, this atmosphere of conspiracy and uncertainty can leave citizens vulnerable to misinformation (Ognyanova et al., 2020). Deepfakes are exacerbating the problem.
Deepfakes also pose a threat to our governing structures. Democratic discourse is strongest when based on proven facts and consensus. However, the uncertainty that deepfakes introduce allows people to live in their own subjective realities, enlarging social divisions and obstructing the democratic process (Chesney & Citron, 2019). This is especially dangerous during elections, when deepfakes are likely to be used by both foreign and domestic powers to manipulate outcomes by releasing compromising manipulated materials about politicians (Liv & Greenbaum, 2020). Antagonistic parties may be enticed to subject an electorate to deepfakes long before an election in order to prime future attitudes (Cialdini, 2018). Although many organizations try to flag manipulated content around elections, these efforts are insufficient.
The use of deepfakes against public individuals creates another behavioral dilemma—the liar’s dividend—in which individuals facing accusations can write off factual evidence as deepfakes (Chesney & Citron, 2019). Widespread deepfakes can prime individuals into doubting the authenticity of materials in a judicial, political, or public context. This is illustrated by the case of the Malaysian Minister of Economic Affairs who deflected evidence of his involvement in a sex tryst by proclaiming it as a deepfake despite no evidence of video tampering (Barari et al., 2021).
To be clear, we would not recommend banning deepfakes. Deepfakes have enabled new and intriguing art forms, served as excellent pedagogical tools, contributed to the economy, and been a benign source of pleasure and amusement. Nevertheless, in the wrong hands, deepfakes are a novel kind of social virus, and like all viruses, their future trajectory and consequences are hard to predict. On a societal level, their greatest threat is their ability to shape public discourse. When misinformation enters the public conversation, it becomes increasingly dangerous because it alters collective understanding and memory. Deepfakes will leave innocent victims and mistrust in their wake. Their increasing prevalence could also lead people to stop believing much of what they see in the media.
Solutions
Deepfakes are created to trick us; the human mind is not prepared to always accurately identify the outputs of sophisticated technologies. This fact can be deployed for social benefit, but so far the technology has been more misused than helpful. Although some tech giants have started flagging some content as misinformation, such flags are not a silver bullet. The shareability of fake news has been found to decrease when it is accompanied by warnings (Pennycook et al., 2018); however, their effect on its believability is unclear. Some studies have found that participants have no greater overall accuracy in judging fake news in the presence of tags (Ternovski et al., 2021). A body of evidence suggests that prior exposure to misinformation increases its perceived accuracy, possibly negating the effectiveness of tags (Pennycook et al., 2018). Detection systems integrating both human and model predictions have been found to be more accurate than both humans and automatic detection methods working alone (Groh et al., 2021). However, the ability of Internet users to provide objective assessments outside of experimental settings may be limited to the extent that the Internet is divided into homogeneous echo chambers.
Steps that might alleviate the problem include pre-exposure warnings that make people aware that information might be false before they see it. Warnings need to be specific; it is ineffective to merely mention that misinformation may be present (Lewandowsky et al., 2012). Also, warnings should come with an alternative causal account that explains both what happened and the reason for the misinformation, as it is important to fill the resulting narrative gap.
From a legal standpoint, we can create tighter restrictions that would hinder the publication and spread of deepfakes. There is currently little distributor liability for social media platforms circulating deepfakes (Kietzmann et al., 2020). In the United States, the legal debate is centered around Section 230 of the Communications Decency Act, which prevents companies from being held liable for the content on their platforms (Chesney & Citron, 2019). The justice system could specify civil liability for the creators and distributors of deepfakes while also increasing legal protection for victims of defamation (Kietzmann et al., 2020).
Individuals have little power to prevent deepfake attacks. When deepfakes threaten reputations, individuals can increase their ability to deny actions by recording their activities (Chesney & Citron, 2019), but this raises privacy concerns. Methods to disseminate facts can help protect communities if they are deeply informed by understanding of the public’s information landscape and means of navigating it. When it comes to correcting political and other forms of misinformation, individuals are more easily persuaded and corrected by someone they know. Therefore, societal norms and discourse on deepfakes should be nudged to create a social environment in which people not only are more skeptical about what they see but also are encouraged to challenge each other’s informational claims. Changing norms and expectations are the key to help our communities develop a sensitivity and resistance to deepfakes and other fabricated content.
In order to alter societal norms, change is needed at every level. People with the most influence on their communities—thought leaders and those most central in social networks—are key. Educational resources including digital literacy training are helpful tools, especially if directed at influencers. Videos explaining political deepfakes have been found to reduce uncertainty and, in so doing, can increase trust in media (Vaccari & Chadwick, 2020). But norms really change only through collective action.
Recommended Reading
Chesney, R., & Citron, D. K. (2019). (See References). Thoroughly covers, with references, both positive and negative potential consequences of deepfakes and discusses potential solutions.
Köbis, N. C., Doležalová, B., & Soraperra, I. (2021). (See References). Offers strong evidence that deepfakes are effective at fooling people.
Lee, J., & Shin, S. Y. (2022). (See References). Provides evidence that deepfakes can be more believable and shared more often than other forms of misinformation.
Sloman, S., & Fernbach, P. (2017). (See References). Offers a theoretical framework for understanding why deepfakes would be more likely to sow confusion and chaos than to change minds.
