When Sally Bowles in Cabaret first sang: "What good is sitting alone in your room? Come hear the music play," she probably didn't envisage a global pandemic locking us in our houses for a year. But no matter: Interactive virtual cabaret club the Zizi Show, which mixes drag performances with artificial intelligence and deepfakes, will come to you instead.
As the red velvet curtains twitch open and Zizi takes the stage to the opening bars of the iconic Wilkommen, Bienvenue, I realize that even as a fan of both cabaret (the theatrical medium) and Cabaret (the musical), this show won't be like any I've seen before.
Zizi, an installation commissioned by the Edinburgh International Festival and the Edinburgh Futures Institute, is a drag act with an uncanny twist. The show I'm watching on my computer screen features a performer lip-syncing and dancing as they might at any drag or cabaret event. But that performer is actually a deepfake -- one of 13 drag artists (and sometimes a combination of them all at once) who are in fact all real people from London's drag community.
I get to choose the song and the performer, although throughout the act I can cycle between them -- there's Me with her pink beehive and oversized lashes, then Oedopussi Rex with his braided, multicolored beard and Lily Snatchdragon with her magenta shoulder pads and purple bangs. No matter which performer I'm watching, their movements remain the same -- the bodies and faces a messy mash-up of projections anchored onto the underlying deepfake skeleton.
Zizi puts a creative spin on deepfakes and challenges the narrative they're all poised to ruin us. The technology, which easily and convincingly swaps in faces in videos, calls to mind malicious clips in which famous or important people are made to appear to say things they never really said or get inserted into graphic pornography. At their most benign, their uncanniness can leave us confused or elicit a chuckle. At their most harmful, they've been identified as a serious threat to democracy.
But while watching the Zizi Show, we're asked to reconsider how we understand them altogether.
Queering the dataset and the internet
There is something inherently fun, playful and tongue-in-cheek about drag that invites people to ask questions about established gender norms and can also encourage them to ask questions about technology, said artist Jake Elwes, the creator of the Zizi Show.
Elwes trained at Slade in London, a "hardcore traditional art school," and from a young age started to teach himself creative software, animation and coding. He's been working with AI for many years now, but until he introduced performers, it was resulting in "quite cold, conceptual art" -- something that can't be said for the Zizi Show.
Elwes first started bringing drag into his work by injecting 1,000 faces of drag performers into a standardized dataset collected by AI researchers of over 50,000 faces -- what he describes as "queering the dataset." The aim was to train neural networks to shift away from representing and understanding only normative identities, and the result was a video installation of seven giant morphing faces at the Edinburgh Festival in 2019.
The next part of the plan was to do something on stage at the 2020 Festival -- ideally to build a holographic deepfake performer that could perform side by side with a real drag artist in a musical theater double act. The idea was to make fun of and subvert people's fear of AI becoming human enough to take their jobs. But then along came COVID, resulting in the cancellation of the 2020 Festival and the wider shutdown of queer spaces altogether.
The pandemic hit at an already "precarious" time for London's LGBTQ venues, which have been squeezed by rising rent and business rates and a reduction in disposable income among customers, Alim Kheraj, author of the book Queer London, said over email. "Since 2006, nearly 60% of all LGBTQ venues in the city have closed down, which is a shocking number," he said.
For a minority group that's frequently discriminated against, not having access to queer venues -- especially during the pandemic -- is a bigger problem than simply not being able to go drinking and socialize. "They provide spaces for self-expression without fear and a space for community," Kheraj said.
With queer people denied access to community spaces and events that often serve as safe havens, it was more important than ever to Elwes to rethink his concept. What he created was the Zizi Show -- a virtual cabaret that attempted to re-create the feeling of a real theater -- hence the noisy buzz of the bar when you first enter the web portal and the curtains open to reveal the stage.
Between and beyond the binary
After an introductory sequence in which you see the bodies of all 13 performers meld and mesh and flicker across one dancing skeleton, you get to choose which performer to watch lip-sync to Shirley Bassey and George Michael hits, which were recorded by several of the participating drag artists.
A key consideration when selecting performers for the piece, said Elwes, was that they be a wide-ranging group of people to ensure he wasn't replicating any of the normativity that often goes into pulling together datasets of faces. The London drag community is diverse beyond the confines of the biologically male drag queens usually seen on RuPaul's Drag Race, and Elwes said he intentionally made sure to represent that diversity in the show.
It was also crucial that all the performers knew what they were agreeing to when they allowed their images to be captured for the show. "There's no consent normally with deepfakes," said Elwes, pointing to common uses -- manipulating politicians and people in power, as well as revenge porn. "It was so important to reclaim that, to do it consensually against the normal, exploitative ways."
Elwes filmed each of the performers from every angle and used a deepfake model to generate their body doing any movement. For each track, one artist lip-synced an entire performance and the AI mapped the other performers' bodies over it. As each track plays, the viewer can switch between the performers or watch the character of Zizi, which Elwes describes as "all of the 13 performers smushed into one amalgamation."
But as much as Elwes -- and the drag artists involved -- enjoy the surprising results of what the deepfake model spits out, he is wary of denying responsibility for its choices. "With this AI machine learning stuff, it's also really important not to give too much autonomy or agency over to the machine," he said.
Subverting bias and power
Elwes has been deeply inspired by MIT Media Lab's Joy Buolamwini and the Algorithmic Justice League, whose widely cited work has helped establish understanding of how algorithms discriminating based on gender and race reflects biases in wider society. Buolamwini has repeatedly called for more accountability in AI, and it's something Elwes seems to have taken to heart.
It would be easy to distance himself from the decisions made by the deepfake model, he said, but it's important to understand that the technology we create is a reflection of ourselves and of the data we feed in to train it.
"It's challenging ourselves to understand why it might be saying those sorts of things -- and that might apply to predictive policing, or it might apply like hiring a person for a job," he said. "We've kind of got to constantly be questioning what biases these things might have programmed into them."
A study by Microsoft research fellow Morgan Klaus Scheuerman published in 2019 found that facial analysis software "performed consistently worse on transgender individuals and were universally unable to classify non-binary genders." For trans people, walking down the street and being perpetually misgendered by facial recognition algorithms via CCTV can exacerbate gender dysphoria and have other potentially harmful consequences, Elwes said.
One of the things that stands out about the Zizi Show as you play around with the controls is that these aren't the flawless re-creations of human beings designed to dupe you, a key trait of deepfakes. They're purposefully messy and strangely constructed. As you switch between performers, you momentarily see the underlying skeleton onto which their bodies are mapped -- a decision that Elwes says was intentional, a way of "visually deconstructing" the illusion.
"I love seeing the failures," he said. When engineers are creating tech for real-world use, it has to be as good as it can possibly be at all times in order to prove it's valid and functional, but there is no such pressure for artists. As a result, you never know what the AI is going to do or which part of which performer it's going to pick at any one time.
But returning to accountability, everything about the Zizi Show, the algorithms and datasets included, is created by humans, and Elwes has no intention of trying to paint the technology as a conscious artist in its own right. Instead, he sees it as "a collaboration between the creativity of the human and this slightly sterile, monstrous technology."
Elwes is keen to lean further into the idea of collaboration and is hoping, when physical venues reopen, to finally get a human performer up on stage with a deepfake. He wants the serious questions about AI that the Zizi project poses to be accessible to people of all backgrounds, so he's aiming to maintain the mischief of performance as a gateway.
"Drag is a perfect way of doing that, because it has humor, and it has fun, and people can just engage with that," he said. "You just always need a sense of humor, no matter how dark it gets."