Face/Off: "DeepFake" Face Swaps and Privacy Laws.

• Author: Gerstner, Erik

IN 2018, a curious trend spread rapidly across the Internet: people posting videos of Nicholas Cage's performances in various Hollywood films. (1) To the uninitiated viewer, these videos might appear to be nothing special, just various facets of Cage's prolific career. However, closer inspection would reveal a subtler thread running throughout these clips: none of these performances actually involved Cage. Rather, thanks to relatively new artificial intelligence (AI)-powered software programs colloquially known as DeepFakes, Internet users had seamlessly inserted Cage's face over the faces of the original actors in these scenes, making it appear as though Cage had always portrayed those characters. (2)

Nicholas Cage's central role in DeepFake videos is fitting, given his starring role alongside John

Travolta in 1997's Face/Off, a film in which his and Travolta's characters both end up wearing the other's faces throughout the film. Although it was only a fanciful Hollywood invention in 1997, face swapping technology entered the mainstream in 2017. In August of that year, University of Washington researchers released a video, seemingly of Barack Obama, discussing topics such as terrorism, fatherhood, and job creation, which had been created using machine learning algorithms. (3) By 2018, similar tools became publicly available, with the most popular, called FakeApp, available for free online. FakeApp was developed using Google's open-source deep learning software. In its first two months of being publicly available, it was downloaded more than 120,000 times. (4)

The many-faceted ramifications stemming from the widespread availability of this and other similar software are staggering. The unprecedented ability to create fabricated messages from politicians and other celebrities, or "fake news" in the parlance of our current political climate, is a major concern - the Pentagon alone has already spent tens of millions of dollars in an effort to research and combat DeepFakes. (5) However, although the political and cultural ramifications of DeepFakes are significant, and worthy of considerable attention across the spectrum of areas that they affect, this article will be limited to primarily examining the legal issues likely to arise from these programs, including privacy, the right to one's own likeness, and defamation/false light claims. The first section will discuss the software and the technology behind it, including a brief introduction to how it technically functions. Next, this article will discuss the state of the relevant law and examine how face swaps have and will continue to intersect with applicable statutory and case law. Finally, it will discuss potential judicial and legislative solutions to present and future problems arising from these sorts of AI technologies.

1. Fakeapp and Machine Learning

   The influx of fake videos stems largely from the widespread availability of simple yet powerful software tools such as FakeApp. Utilizing machine learning to train AI, it condenses what would be an exceedingly complex operation for even the most experienced digital artists into a single button press to create a face swapped video. While having a moderately powerful computer is a slight barrier to the effective usage of the program, it otherwise is relatively uncomplicated to create fake media with it. (6) In its most basic form, all a user needs is a "base" video and a number of source images of the face of the person being pasted into the video. The more source images input into the program, the more seamless the final video will appear. (7)

   After creating the datasets, FakeApp then trains the deep learning algorithm, a process that can take hours or even days, depending on how powerful a computer is used and the quality sought for the final video. Thereafter, the user needs only to click one more button to create the resulting video. A more experienced creator may be able to achieve a higher degree of realism through more involved interaction with the FakeApp software, but by following the basic steps, even a novice can fairly easily create a face swap using the program. (8)

   While this process is straightforward for the front-end user, it is anything but for the computer running FakeApp. (9) The software utilizes Google's open source TensorFlow machine learning algorithm to power its operations and requires considerable computing power from any hardware on which it runs. The final video quality is determined by a combination of factors, including the similarity of the faces and poses among the base video and the source images, and the amount of time spent and quality of the AI training. What is not a factor, however, is the software itself--computer-generated faces were once strictly the domain of bigbudget studios with deep pockets, proprietary software tools, and considerable amounts of time. For example, the much-discussed appearance of a computergenerated young Carrie Fisher in Star Wars: Rogue One in 2016 was the product of a $200 million production budget, and, according to the visual effects supervisor, "a super high-tech and labor-intensive version of doing makeup." (10) Now, private individuals are able to create videos equaling or even surpassing those created by these studios for a tiny fraction of the time and expense. (11) Machine learning is the great equalizer: thanks to the powerful AI algorithms powering FakeApp and similar software, those able to make full use of it have had the power at their fingertips enhanced exponentially, a process of growth that is likely to continue as this technology continues to progress. (12)

2. Privacy, Defamation, and Fake News: The Present State of the Law

   There are potential ramifications flowing from the creation and use of the resulting media that span the legal spectrum, including ramifications in election law, criminal law, evidence, and intellectual property. This article, however, will focus on potential privacy issues, including defamation, false light, and the right of publicity, also sometimes known as personality rights. (13) While case law and statutory law regarding deep fakes currently is scant or nonexistent, there are analogues which may provide some guidance as to how the law in the United States will address issues arising from these new technological advancements.

   One potential legal concern flowing from these fake images is defamation. A defamation cause of action could arise from an individual using FakeApp or similar software to create a fake video of an individual saying or doing something that would injure the individual's reputation if it were true. For example, in the aforementioned video the University of Washington created of President Obama, the audio could be any recording the creator wants to use, literally putting words of the creator's choosing into Obama's mouth, including statements that could be highly offensive to an unsuspecting viewer. (14) In states that recognize a difference between slander and libel, a face swapped video could easily give rise to both of these causes of action. For example, if someone creates a video purportedly showing Person A saying defamatory things about Person B, then Person B might have a claim for slander (as the defamatory statements were verbal), while Person A might have a cause of action for libel.

   While fake media of the sort described above would seem to satisfy the requirements for a defamation claim, there is no common law jurisprudence either way for a claim resulting from such a created video. There is some precedent, though: in some states, the tort of defamation explicitly applies to altered still images. (15) Generally speaking, videos are treated the same as still images under the law. (16) Thus, a defamatory video should be fully actionable if a plaintiff attempts to bring suit. However, there are powerful affirmative defenses for defamation claims that apply in many cases and make it difficult for plaintiffs to win these lawsuits. (17) The primary defense for a content creator is a claim that their face swapped video is parody, which...