A Convincing Deepfake Demo Rattles The TED Audience

When Tom Graham, cofounder and CEO of AI video startup Metaphysic, gradually morphed the video image of his face into that of TED director Chris Anderson in real time, as Anderson was interviewing him live on stage at earlier this week at the TED2023 event in Vancouver, you could hear the gasps fill the room. The TED audience, who you can’t really characterize as a mob of luddite peasants with pitchforks, were as alarmed at the implications of this technology as they were impressed with the technical virtuosity of the application.

Graham’s company is commercializing the video deepfake technology behind the viral sensation “Deep Tom Cruise” as a VFX solution for the entertainment industry. It is part of a wave of new tools being developed using generative AI to create photorealistic imagery and motion video good enough to mimic live performance and blend seamlessly with actual photography. The technology debuted on “America’s Got Talent” and Graham said it will see its first wide-scale exposure in a feature film to de-age Tom Hanks and Robin Wright in the upcoming Robert Zemeckis project Here.

This particular application of AI for de-aging and creating convincing digital performances sits at the relatively benign end of the continuum of uses for deepfakes, notwithstanding the disruption it is likely to cause in the special effects industry. Other more sinister scenarios such as creating vividly lifelike political disinformation or non-consensual pornography, are typically the first things people think of when they see it. Is this cracking open Pandora’s box? asked Anderson.

“We’re not making our technology open so we don’t have an issue of people using it without supervision,” said Graham in an interview following his TED Talk. “Our customers have to sign a very long warrant that they have the rights to the images we use to train it. But in the long run, when the cost of this drops to zero and you have a billion people out there doing this, you definitely need a distributed system where people own and control their data.”

To help force IP law to grapple with those issues, Graham, who worked as a lawyer before getting into the tech industry, has filed for a copyright on an AI avatar he created of himself using his company’s technology, along with a brief explaining how the creation fits the criteria for copyright protection. The copyright office has not yet issued a ruling.

The murky legal status of deepfakes is not the only issue causing concern. The technology demonstrated by Graham at TED is capable of obliterating the distinction between real and digital in human perception. No more uncanny valley; just an uncanny landscape extending in all directions.

Graham argued that blurring the lines between real and synthetic imagery has some therapeutic benefits, as users would be able to engage in conversations with computer-generated versions of their younger selves or deceased loved ones as a way of coming to terms with grief. “The mind integrates stuff and it makes it more realistic. A Zoom call is a poor substitute for real life, but if you can make that Zoom something way closer to real life [through immersive, hyper realistic AI-generated imagery], you can make that into real life in your mind. The memory you form of it will make it seem real.”

Most of the TED audience members I spoke to found that supposed benefit to be nearly as disturbing as the obvious downsides. Graham said that response is largely generational, and that younger people who marinate constantly in digital media and culture do not prioritize distinguishing between real and digital experiences. “Many of them spend a lot of their lives in videogames, and those places are more real to them than the physical world.”

Another controversial aspect of Graham’s presentation to the extremely socially minded TED crowd was his posture on the inevitability of deepfakes becoming part of the information ecosystem. “It’s happening and it’s happening quickly,” he said on stage, appearing to suggest that we simply need to accommodate innovation for innovation’s sake regardless of legal, ethical and philosophical qualms.

Graham appeared nonplussed by that interpretation when we discussed it later. “My default is that if there’s nothing we can do to stop it then we absolutely have to regulate it. This idea that, ‘Well, it’s going to happen anyway, so we should just not do anything’ feels like a uniquely American response,” said Graham, who is Australian.

“At the very heart of the Internet is Section 230,” he continued. “That is designed to prevent people from rampantly abusing other people’s copyrighted material through the proliferation of social networking and user generated user posted content it. In this paradigm, maybe it would be better if you could just view anything, anytime, any place and there was no moderation. There are certainly people who might argue for that, but it’s definitely not better for society. I am very pro-regulation, at the very heart of who I am. I am an institutions person.”

While Graham may welcome regulation, it is unclear whether our institutions are agile, responsive or competent enough to handle what is coming as this kind of AI imagery gets better, faster and cheaper by the moment. Even if they are, the disruption to business, culture and society is likely to be more profound than we can currently imagine. Several speakers later in the TED program, including cognitive scientist Gary Marcus and researcher Eliezer Yudkowsky, advocated for immediate international initiatives to lock down the more concerning applications of AI, including deepfakes, until we can get more clarity and transparency into the systems.

Meanwhile, Graham said that demand for Metaphysic’s capabilities is booming in Hollywood, and the company, which was founded in 2021, expects to grow to nearly 200 employees worldwide by the end of the year.