Exploring the Intersection of GANs Technology, and Right of Publicity Law: A New Paradigm for Publicity and a Potential Headache for Lawyers
June 21, 2022 · IDSN 428 · Overview of Intellectual Property Laws for Creatives and Entrepreneurs
On how generative adversarial networks are outrunning the Right of Publicity — and what a new legal paradigm might look like.
A generative adversarial network (GAN) is a machine learning model that utilizes a neural network to create generative outputs such as images, sounds and videos. The advancement of this technology has caused controversy and gained publicity in recent years, due to the ability of these machine learning models to generate images, videos and audio clips of celebrities that appeared realistic, but were entirely fictitious (Sample, 2020). The fidelity of AI generated images, videos and audio clips is continually advancing making it difficult for a reasonable person to determine the authenticity of any video clip (Metz, 2021).
The way GANs function is related to other machine learning models operate, where billions of data points from a given celebrity are input into the model through a process called “training.” The machine learning algorithm is trained to understand the characteristics of a celebrity’s facial features, facial expressions, mannerisms and auditory voice (Lewis, 2019). The training process relies on actual images, video footage and audio clips from the person or celebrity that the GANs model is trying to emulate. GANs relies on two neural networks, one is called the “generator” and the other is called the “discriminator.” The generator creates fake data and the discriminator distinguishes the fake data from the training samples. When a generator produces something implausible, it is penalized by the discriminator, which eventually trains the neural network to become more accurate (Wood, 2020).
In order for GANs to create accurate outputs, it requires a substantial amount of data. To create a realistic image, it requires 100,000 or more images (Johnson, 2020). These images must all be of high quality for the system to work effectively, as is said in computer science, “garbage in, garbage out.” Another factor surrounding GANs is that the process utilizes unsupervised learning to form the neural network. Supervision would mean that a human would have to label the images, informing the computer what they represent. In an unsupervised configuration, the computer interprets the data it is receiving on its own. As a result, the inner workings of a GANs neural network can be incredibly complicated and there’s very little visibility into the actual process of the computer’s decision-making, nor is their great control over the process. Largely, machine learning can be a black box.
In radiology, AI has been utilized to diagnose cancer in x-ray images and algorithms have been shown to be more effective than the world’s leading experts (Armitage, 2018). However, widespread application of AI for this purpose has not proceeded due to a lack of clarity on how AI works. An academic field of study has emerged around explainable AI (xAI) systems. This field seeks to unpack how AI systems make decisions (Dickson, 2019). With greater visibility and control over artificial intelligence, there is promise that the domain will grow and beneficial applications of AI will widen. The emergence of the field of xAI may also lead to new laws and regulations. This process has already begun, especially around applications such as deep fakes.
In November of 2020, Andrew Coumo signed bill S5959D (Savino, 2020), a law that expands right to publicity to protect against the unauthorized use of a “digital replica” of a person’s likeness. This NY state law protects against commercial use of the likeness of people who have been deceased for 40 years and it protects against non-consensual use of computer generated pornography (Ferrraro and Tompros, 2020). This law protects the most vulnerable to exploitation and may provide groundwork for future laws that expand regulation on artificial intelligence. Such regulations are in their inception and the federal government is still formalizing the policy implications of AI on national defense and civic law enforcement (Stanford HAI, 2021). The National Defense Authorization Act for Fiscal Year 2022 promised to report its findings on AI to Congress when its review is complete (Scott, 2021).
The Right of Publicity provides an individual with legal rights to protect their name, image likeness and voice. This right is governed by the “Naked Eye Test” in which a reasonable person is able to understand and determine who is being depicted in a likeness. The Right of Publicity is not explicitly defined in federal law, but it is present within 19 state statutes and is common law within 28 states. The federal Lanham act, section 43 forbids false designation of origin, including false endorsements which can be used to prohibit unauthorized use of a person’s name and likeness to generate financial benefit. There are exceptions and limitations to the Right of Publicity including artistic expression, parody, biography, and newsworthy material, is protected under the First Amendment.
Regardless of the infancy of AI regulations, the Right of Publicity provides a clear framework to protect misuse of one’s identity. So why have there not been any deep fake lawsuits? In the case of DeepTomCruise, the social media exposure has benefited Tom Cruise. The current viral videos which are created by Chris Umé and Miles Fisher, present a flattering, youthful image of Tom Cruise appearing next to other celebrities and in exclusive settings. Prior to the debut of Top Gun: Maverick, Tom Cruise’s image has been at the top of social media feeds for over a year and a half. This exposure has helped Tom Cruise achieve a new domestic box office record crossing $300 million in ticket sales. It would be hard to fathom that this viral deep fake phenomenon is merely a coincidence.
GANs technology should be seen as a new paradigm for publicity and entertainment. The possibilities of the technology are still being explored and new business models are still emerging. There will be a lot more experimentation and failures before regulations are firmly established around AI. If a case involving deep fakes were to be tried in court, a defendant might argue that the author of a deep fake image cannot be currently attributed due to the state of GANs technology and xAI. The defendant may also argue for protection of free speech under the First Amendment. Similarly, DeepTomCruise would be protected because it is a parody. If a deep fake is used to generate unauthorized commerce, a judge will most likely rule in favor of the plaintiff. Unfortunately, what is more likely to occur is a lawsuit surrounding defamation and that might lead to a very bizarre trial where someone may argue that intellectual property belongs to a computer.
—
Sources
Dickson, B. (2019) Explainable AI: Viewing the world through the eyes of neural networks, TechTalks.
Metz, R. (2021) How a deepfake Tom Cruise on TikTok turned into a very real AI company - CNN.
Sample, I. (2020) ‘What are deepfakes – and how can you spot them?’, The Guardian, 13 January.
Savino, D.J. (2020) NY State Senate Bill S5959D, NY State Senate.