Don't Let Them Fake You Out: How Artificially Mastered Videos Are Becoming the Newest Threat in the Disinformation War and What Social Media Platforms Should Do About It.

• Date: 01 April 2021
• Author: Sylvester, Shannon
• Published date: 01 April 2021
• Author: Sylvester, Shannon

TABLE OF CONTENTS I. INTRODUCTION 370 II. DEEPFAKES DEFINED 371 A. Neural Networks and the GAN Approach 372 B. From Hollywood to Handhelds 373 III. DEEPFAKES AMPLIFY THE PROBLEM OF DISINFORMATION 376 A. Disinformation Campaigns and the Difficulty in Seeking out the Truth 376 1. Weaponizing Social Media 377 2. Fake Speech is (Mostly) Free Speech 378 B. What Social Media Companies are Doing About Deepfakes 383 1. Facebook 384 2. Twitter 385 3. Google/YouTube 386 IV. MITIGATING THE DEEPFAKE THREAT 388 A. Amending CDA Section 230 388 B. Stronger Deepfake Legislation 389 C. Fighting Technology with Technology 390 D. Knowledge is Power 391 V. CONCLUSION 392 I. INTRODUCTION

It looked like former President Barack Obama. It sounded like former President Barack Obama. And without a second glance it fooled the best of us into thinking it was former President Barack Obama. (1) The "it" was a deepfake, an artificially generated video that used images and audio cloning technology to imitate the former President, making it appear as though he was saying things that he, in fact, never said. (2)

"This is a dangerous time," the fake Obama warned as he claimed, "[w]e need to be more vigilant with what we trust from the Internet." (3) While the video convincingly portrayed the former president addressing the nation, it was only because of a lack of eloquence that the video's creator Jordan Peele gave to Obama that people questioned the truth of the video. (4)

But what if Peele refused to admit that he created the video? The video, hosted on YouTube, has over 8.3 million views. (5) BuzzFeed's title of the video, "You Won't Believe What Obama Says In This Video!" followed by an emoji wink, is an obvious attempt to get people to click on the video. (6) Many people, intrigued by the title of the video, might be tempted to click on the link, and find themselves believing it was indeed former President Barack Obama saying obscenities, rather than a fake.

Believing that the former President used profanity while addressing the nation, could at a basic level, harm the President's reputation, but at a higher level, stands to do much more damage. Beyond the President's reputation, the nation's reputation could be harmed abroad. Critics of Obama could be further inflamed by the former President's offensive remarks in the video. Peele's deepfake Obama video sought to warn us of the real possibilities of disruption that could be caused by this manipulating technology. It further attempted to show that deepfakes can impair our understanding of the truth through deception, and in the hands of bad actors, can contribute to our already toxic and uncivil political discourse. The technology used to create deepfakes is advancing, and in the future, realistic deepfakes that might not be so easily debunked threaten to disrupt our already fragile democratic infrastructure.

This Article will explore the manipulative effects of deepfakes and how their truths can spread if left unchecked, significantly disrupting democracy.

Part I of this Article will introduce the origins of deepfakes and explain the technology and techniques that make up a deepfake. This section will also describe how deepfakes emerged on the consumer scene, and how this can have some beneficial uses--but like any technology, many negative implications as well. Part II will focus on deepfakes as catalysts to the disinformation war. As trust in the media has waned over the years, especially in the era of "fake news," public faith in the media to deliver accurate, credible news has become increasingly important. First Amendment constraints add difficultly to legislators seeking to regulate deepfakes, especially on social media sites where companies currently enjoy immunity through Section 230 of the Communications Decency Act (CDA). Social media sites are thus a prime market for deepfakes to thrive. Realizing this, some social media companies have adopted policies banning deepfakes, but they do not go far enough.

Part III of this paper will prescribe guidance on further steps social media companies should take to combat deepfakes. Social media companies should look towards helping policymakers and legislators define more clearly what deepfakes are and develop standards aimed at addressing the manipulation issues caused by deepfakes. In addition, companies should look towards adopting technological solutions. However, because social media sites are set up and run differently, there is no "one size fits all" approach when it comes to regulation and enforcement. Therefore, this paper attempts to show the problems deepfakes can cause and offer best practices that social media companies can adopt to help prevent the onslaught of damage deepfakes threaten to do if left unguarded.

2. DEEPFAKES DEFINED

   Manipulated media encompasses a wide range of material, with deepfakes falling under that umbrella. (7) Deepfakes are a type of manipulated media created entirely through artificial intelligence (AI) processes. (8) "Deep" describes the "deep-learning" aspect of deepfakes, whereas "fake" refers to the fact that the video created often depicts people saying or doing things they never said or did. (9) Deepfakes should be distinguished from shallowfakes, which are also manipulated media, but manipulated through human intervention rather than artificial intelligence. (10)

   The technology that creates deepfakes is relatively simple to access and use. (11) But when that technology becomes more readily available on the consumer market, that raises concerns about the widespread use of deepfakes. As deepfakes enter the mainstream and grow in popularity, it is likely the technology used to create them will also become more advanced. If all it takes now is downloading an app to your phone to create a deepfake, imagine what a malicious actor could do with more sophisticated technology. (12) This section will explain the technology surrounding deepfakes and how its uses extend beyond its initial inception in Reddit threads.

   1. Neural Networks and the GAN Approach

      Deepfakes use deep learning, or neural network processes, known as "Generative Adversarial Networks" or GANs to function. (13) Deep learning dates back to the 1950s, when Frank Rosenblatt attempted to build a machine with a brain. (14) The idea of giving robots minds is why deep learning processes are often referred to as "neural networks." (15)

      The GAN neural network process involves two networks that work against each other to produce the outcome. (16) The first network, the generator, uses a sample dataset of images to create a new image based on the sample set. (17) The second network, the discriminator, receives the new "fake" image from the generator and determines how successful the generator was at creating a plausible image. (18) If the discriminator determines the new image is inadequate and does not match up against the subject (e.g., if the mouth does not line up when the subject speaks), the discriminator sends the image back to the generator so the generator can churn out a new and improved image. (19)

      The GAN method works with both images and audio clips. (20) Jose Sotelo of AI company Lyrebird, described his company's audio AI as pattern-matching. (21) The program runs by finding the uniqueness in a voice and then attempting to recreate that uniqueness. (22) While a fake audio message could be detrimental, a fake video is often times more damaging because it betrays both hearing and sight. (23)

      Another way to think of the GAN approach is as a game of trickery. (24) The first machine tries to trick its adversary, the second machine, into believing the image or audio clip is legitimate. (25) If the second machine can easily spot the fake, it sends it back to the first machine to try again. (26) The first machine tries repeatedly until it can successfully trick the second machine into believing the image or audio clip is real. (27)

      GANs have made their way into the consumer sphere. (28) Using the GAN approach, many companies work on their ability to create seamless deepfakes using just one video source or one photo source. (29) The results are impressive for the minimal effort it takes to create a convincing deepfake. (30) The ease of accessing and using deepfake technology for consumers has already resulted in a variety of entertaining purposes. Deepfakes have the potential to increase creative expression and even benefit the health industry, but they also have the potential to wreak havoc on individual liberty and democratic institutions.

   2. From Hollywood to Handhelds

      Deepfakes are relatively new to the consumer scene, but Hollywood's special effects teams have dabbled with the technology for years. The film Forrest Gump (1994) included an appearance by President John F. Kennedy, digitally recreated from archival video. (31) When Paul Walker died halfway through filming Furious 7, his brothers served as face templates to form a digital recreation of him used in the rest of the movie. (32) Even more recently, facial mapping and AI programming made actors look years younger in the

      Netflix film, The Irishman. (33) As technology improves, some hypothesize that actors will be able to license their likeness for use in television and movies without ever needing to read lines on camera. (34) Take the same technique, put it in the hands of the consumer, and suddenly the consumer becomes the director. (35) Moviegoers can make their ideal cast ensemble for their favorite movie possible with this new technology. (36) Ever wonder what Nicholas Cage would look like in Superman? (37) People on the Internet did, and one of the first trends of consumer-use deepfakes included putting Cage into as many movies as possible. (38) The Internet's obsession with Cage (maybe catapulted from his appearance in Face/Off) (39) shows the potential for consumers to embrace their creative sides as they start reimagining film.

      The fascination with swapping...