The uncanny valley is one of the most famous ideas from the field of digital aesthetics. First identified by Japanese roboticist Masahiro Mori (and coined in English by Mori’s translator Jessica Reichardt), the concept describes the tendency for people to experience an unease, bordering on revulsion, when confronting something that very closely resembles a human form. This effect is non-linear, in other words, it only kicks in after an object crosses a certain threshold of verisimilitude. Marge Simpson, for example, does not trigger an uncanny valley response for most people in the way that, say, Ameca, the AI robot produced by UK-based Engineered Arts, does. This is because Marge, with her exaggerated features, bright yellow skin, and blue hair, does not look like a (medically viable) human being; viewers can easily see that she isn’t real. On the other hand, Ameca, whose proportions and facial expressions are much more in line with those of an actual person, gives people an unsettled, creeped-out feeling. She is simultaneously both human and inhuman, familiar and strange; the experience of looking at her is uncanny, hence “uncanny valley.”
From the earliest automata to the latest deep learning models, the history of artificial intelligence is riddled with examples of creators and audiences wrestling with this strange phenomenon. In a sense, it’s baked into the project of artificial intelligence — in order to create a truly intelligent agent, we must necessarily scale down the banks of the uncanny valley before we can climb up the other side. Further complicating matters, the uncanny valley is not a fixed geography, but rather a moving target. Audiences tend to get more sophisticated the more they’re exposed to any particular artifice; their perception grows keener, and they learn what to look for to spot a fake. What passes the Turing Test today may well fail tomorrow. In the famous words of George W. Bush, “You fool me — can’t get fooled again.”
It’s not surprising that today’s technology brings us as close as we’ve ever come to that second ridge of the uncanny valley. The famous recent case of former Google engineer Blake Lemoine attributing sentience to the LaMDA chatbot shows that even the savviest technologists can have trouble distinguishing between robots and humans. Lemoine’s case was a text-based encounter with a machine, but the technology of LaMDA’s conversational AI is part of the same umbrella of deep learning models that are used to generate “deepfakes” — artificially produced images and videos that, some claim, look just like the real thing.
Whether or not you believe that deepfakes can be detected with the naked eye, almost everyone will admit that doing so requires a much higher degree of scrutiny than any previously existing method of artificial image generation. Some have put this technology to light-hearted uses, such as this video that morphs Bill Hader’s face to match his impressions. More unsavory uses of deepfake technology include the production of revenge porn, disinformation campaigns, or state propaganda.
One of my favorite deepfakes is “Seeing Is Believing” by Devin Ronnenberg and Kite. This project takes famous faces of tech and business luminaries and grafts them onto footage of more unambiguously infamous historical figures. Once you’ve seen Elon Musk speaking the words of Heaven’s Gate cult leader Marshall Applewhite, you’ll never again hear Musk’s paean’s to space colonization in quite the same way.
Francesca Panetta and Halsey Burgund at MIT used this technology to create a fake moon landing video. “We hope that our work will spark critical awareness among the public,” said Panetta. “We want them to be alert to what is possible with today’s technology, to explore their own susceptibility, and to be ready to question what they see and hear as we enter a future fraught with challenges over the question of truth.”
It’s also possible to use deepfake technology with audio. Musician Holly Herndon recently released a digital twin of her own voice — an AI-powered model that allows anyone to “sing in her voice.”
“If I can allow people to play with my IP, my digital identity, my intellectual property, what might they come up with?” Herndon said. “Could someone else go on tour as me with my permission? Could I be in a thousand different bands in a thousand languages, and what would that even sound like?”
The implications of deepfake technology are dizzying to consider. Naturally, some artists are grappling with these same issues of authenticity and artifice in their own ways. Some of my favorite examples include Wendy Redstar’s print series in which she spoofs stereotypes about indigenous peoples by making real-seeming photographs with anachronistic elements.
Cindy Sherman, in her latest work, also plays with the traditions of realism in photography and performance.
One of the most unsettling takes on the uncanny valley I saw this year was at the Venice Biennale: a video by Diego Marcon titled “The Parents Room”. From the Venice Biennale’s website:
Diego Marcon’s newest film/video work The Parents’ Room (2021) plays on a loop, with its beginning and end marked by a computer-generated blackbird swooping down onto a snowy windowsill. The scene moves into a room where a man sits on the edge of an unmade bed, a woman lying beside him. The man begins a choir-backed monologue detailing the story of his murder spree and suicide. The victims — his wife, daughter, and son — enter the scene one-by-one. The human element in this work — something one might find comfort in — is overcast by the dark shadow of what appears to be a chilling afterlife. The film is reminiscent of stop-motion animation, but Marcon creates this effect with live actors. Made from casts of the actors themselves, the synthetic prostheses they wear become eerie, inanimate doubles, such that there is something off about the animated figures — more corpse-like than living. Marcon’s works, like The Parents’ Room, Monelle (2017), and Ludwig (2018), enter the “uncanny valley,” where extreme representative realism evokes a response of repulsion and restlessness. Manipulating every element of the set, lighting, costume, sound, and script, Marcon creates characters that give the viewer goosebumps with simultaneous feelings of aversion and unsettling familiarity.
In different ways, all of these works confront the aesthetics of realism, a tradition to which deepfakes and other AI image generation technologies are only the latest additions. What does the real world look like? What are the aesthetics of no aesthetic? And if we can’t trust our eyes to discern the real from the hyper-real, what can we trust?
This is part of a 4 part series on visual trends and aesthetics of AI Art, if you liked this one, check out the first article in the series Visual Trends in AI Art.