Abstract
The Zizi Project is a series of connected art and performance pieces created by artist Jake Elwes in collaboration with Me The Drag Queen and members of London’s drag performance scene. The works – currently Zizi: Queering the Dataset (2019), Zizi & Me (2020; ongoing), and The Zizi Show (2020) – sit at the intersection of drag performance and Artificial Intelligence (AI), playing with and queering facial recognition software, deepfake technologies, and Machine Learning algorithms. I consider The Zizi Project as an example of work at the vanguard of an emergent field of queer AI performance. The project intervenes in complex conversations surrounding AI and Machine Learning, including the lack of representation of diverse identities and communities in datasets used to train these systems and the complexity of creating datasets which include queer and trans bodies and identities. However, in aiming to use drag performance to expose and demystify these complex technological systems to audiences, I propose that queerer forms of art making and performance emerge that push at the boundaries of both drag and the technologies used. Ultimately, The Zizi Project articulates drag and queer futures where the digital and the actual interact in increasingly complex ways to explore notions of diversity, inclusion, and access that speak to fundamental questions of what counts as drag, what counts as queer, and, indeed, what count as human.
Introduction
In October 2020, during a pause in the UK COVID-19 lockdowns, I attended a performance at the Royal Vauxhall Tavern (RVT), a legendary LGBTQ+ venue in Vauxhall, London. It was a drag show with a twist. The performers were joined by deepfake versions of themselves created using Artificial Intelligence shown on a projector screen.Footnote1 Playing with their deepfake selves, they could change the view for the audience (full size or zoomed in on the face), change the song they were performing, and even change their body entirely to become one of the other performers. The work is an interactive cabaret called The Zizi Show created by artist Jake Elwes. It brings together Artificial Intelligence (AI) and Machine Learning with drag performance, pushing at the boundaries of drag and technology.
This article considers The Zizi Project as an example of work at the vanguard of an emergent field of queer AI performance. Through an examination of three different artworks, I consider how this work at the intersections of drag and cutting-edge technologies offers insights into potential futures of drag and queer performance. Firstly, the video art piece Queering the Dataset (originally made in 2019 and shown in multiple venues internationally) highlights drag aesthetics and forms beyond binary understandings of drag and gender. In doing so, it intervenes in debates surrounding the limits of the datasets, which are used to train facial recognition software. Maintaining these utopian explorations, I turn to video performance Zizi & Me (originally screened in 2020 online during the COVID-19 lockdown in the UK, but currently being developed into a live performance), considering the failures in the technology’s creation of a deepfake drag performer. These failures expose the inherent issues with AI systems, and act as moments of potentiality which contain glimpses of queer (performance) futures. This performance, whilst anachronistic in its musical theatre content, makes headways into the future of queer performance practices in relation to technology. Finally, I turn to the interactive digital cabaret The Zizi Show (originally screened in 2020 online during the COVID-19 lockdown, but which has now been shown in multiple venues internationally), containing deepfakes of thirteen drag performers who put on a show for audiences in their internet browser. Extending the utopian analyses which sustain my thinking, in aiming to ‘demystify’ AI through the use of drag performance, I propose something more exciting occurs as new horizons of queerly digital performance practices begin to emerge. These utopian considerations frame The Zizi Project as laden with potentiality for queer performance futures, where drag and AI interact and intersect in complex ways to sketch out new possibilities for queer performance.
Where queer performance practices, and in particular popular performance forms such as drag, are often low tech because they happen in venues without much provision beyond a few lights and a sound system (and, in rare circumstances, a projector and screen), I engage with The Zizi Project as an important point of development in queer popular performance practice as well as in digital art practices. Its importance as an area of enquiry for queer performance studies lies in the fact that it does not only put drag on the screen, leaving both drag and digital practices unaffected by one another, but also collides drag with digital AI processes, rendering both fields changed in the process. Importantly, much of the work in The Zizi Project is in development and, whilst some parts of the project are finalised video and art pieces, other aspects of the practice continue to shift as access to technologies and spaces for performance develop. Indeed, in the process of writing this article I have continually updated information about The Zizi Project as different areas have been developed into live performances. This makes it both a fertile and challenging site for exploration, but one which indicates its importance in the field of queer performance studies.
Artificial Intelligence
AI and Machine Learning surround our existence in both insidious and obvious ways, from voice-activated home assistants including Siri, Alexa, and Google Home to algorithms that determine what content and advertisements we see on social media. AI and Machine Learning systems are able to engage with vast amounts of data at unprecedented speeds. AI systems learn through processes simultaneously similar and different to human learning. One example of these systems, Generative Adversarial Networks (or GANs), can produce deepfakes, where networks generate and manipulate fake imagery of a person almost indistinguishable from reality.Footnote2 Journalist Ian Sample describes deepfakes as ‘[t]he 21st century’s answer to Photoshopping, [they] use a form of artificial intelligence called deep learning to make images of fake events, hence the name deepfake’.Footnote3 Many applications of deepfakes are seen as negative or dangerous, from the spreading of fake news to its increasingly ubiquitous use in pornography, where particularly images of women are being added to pornography scenes as a form of ‘revenge’. The Zizi Project demonstrates an application of these technologies in performance, which arguably resists (and, to a certain extent, exposes) these misogynist and democracy-threatening uses.
These increasingly popular and quotidian manifestations of AI are accompanied by catastrophising narratives, where popular representations of sentient robots and conscious computer systems wreak havoc or save the world in literature, television, and film. However, these high-octane representations often miss the more insidious issues that accompany these technologies. Fundamentally, AI is produced by humans. The datasets from which AI learns are often compiled by humans. Therefore, they contain within them the biases that exist within our societies. As Legacy Russell makes clear, ‘All technology reflects the society that produced it, including its power structures and prejudices. This is true all the way down to the level of the algorithm’.Footnote4
Algorithmic bias in relation to race, gender, sexuality, disability, and more has been taken up by researchers and practitioners within this field.Footnote5 As the advocacy and research organisation the Algorithmic Justice League (AJL) note
In today’s world, AI systems are used to decide who gets hired, the quality of medical treatment we receive, and whether we become a suspect in a police investigation. While these tools show great promise, they can also harm vulnerable and marginalized people, and threaten civil rights. Unchecked, unregulated and, at times, unwanted, AI systems can amplify racism, sexism, ableism, and other forms of discrimination.Footnote6
These debates inform AI research and practice, and are articulated evocatively in Joy Buolamwini’s video ‘AI, Ain’t I A Woman?’, which draws on Sojourner Truth’s (1797–1883) ‘Ain’t I a Woman?’ speech, to examine racial injustice and prejudice embedded in the heart of many AI systems, search engines, and social media websites.Footnote7 Buolamwini’s work highlights how these systems discriminate against Black women, noting that a lack of representation within datasets used to train Machine Learning models erases or actively discriminates against diverse communities. Buolamwini and the AJL move beyond representation, however, and advocate for equitable and accountable models for the development and implementation of AI.
For the AJL, ‘Equitability’ and ‘Accountability’ extend demands for ethical AI (which, instead of producing mandatory and unilateral legislation and regulation, allow organisations and governments to set individual standards) and inclusive AI (where, in the name of inclusion, data is collected and used without consent or in violation of privacy). The critique of inclusion is important, where only addressing who is included in datasets could lead to technically accurate systems more capable of discriminating against already marginalised communities. For the AJL, Accountable AI involves meaningful transparency and continuous overnight to ensure both the legibility and continued responsibility of these systems and those who use them, and direct ways to redress both historic and ongoing harm caused by AI systems. They propose that one discussion required in any conversation about an AI system is whether those systems should be used in the first place, if their capacity to harm (particularly marginalised communities) far outweighs the benefits that might be produced. They propose that governments and organisations must not only be transparent about what AI products are being used and how, but also open up the processes by which AI systems are developed and implemented, to ensure those who are likely to be most affected by these systems have a say in their full life cycle.
Ethics and inclusion, equitability and accountability, run through The Zizi Project, which at times directly addresses these issues through drag. As I explore below, at various points the work has been questioned for its ethics but also its appropriate-ness; ‘should this work happen?’ is a question Jake and the drag artists in these projects explore. Furthermore, the process of making the work has been openly explored in the artworks themselves, where the process of construction is built into the aesthetic quality and where talks, discussion, and explanatory videos have run alongside the project in various ways. I do not propose The Zizi Project as a utopian example of an AI project, but instead suggest that the issues and challenges proposed by the AJL are deeply embedded in the project’s core and run through its DNA in complex ways.
Finally, it is important to note that art making practices using these systems are proliferating.Footnote8 Artists working with cutting edge technologies is nothing new, and arguably those working with AI and Machine Learning are indebted to practitioners working historically in new media and video art. These artists, such as Nam June Paik (1932–2006) and Pipilotti Rist (b.1962) to name just two, demonstrate histories of using technology not just to present art but to make it. Fundamentally, this history of art and technology is also a history of art disrupting technology, where artists (and often artists with experiences of marginalisation) play with and hack technological forms to make their work.Footnote9
A Note on Process
In this article, rather than exploring The Zizi Project (and AI art practices) from a technological or artistic perspective, I instead consider what queer popular performance forms such as drag do to AI and what these technologies might do to drag. I explore The Zizi Project as someone connected to it from the position of a producer and academic who extensively works with drag performers. I also explore it as someone new to these technological forms who, through engaging with this project, has begun to discover the potential (and the challenge) of AI as something connected to drag studies as well as contemporary socio-political conversations about access, inclusion, and ethics. I am indebted in this research to my discussions with the artist Jake Elwes, both ongoing informal conversations as the project develops and a research interview which informs this thinking.Footnote10 This research benefits from my relationships with Jake (my partner and collaborator) and with Me The Drag Queen (my husband and long-term co-producer), my position in the drag performance scene as a producer and my years studying and writing about the form.
Drag in this work is not a simple term, and whilst any definition of drag is doomed to fail to account for the multiplicity of drag forms, I employ the title Kings, Queens Monsters, and Things as a way of indicating the types of drag practice under examination in this article. Many contemporary drag performers, including those who appear in the various incarnations of The Zizi Project, do not describe themselves as kings or queens but instead think of themselves as drag monsters who might explore animals, mythical beasts, and the more-than-human in their drag aesthetics and practices, and drag things who might be similar to drag monsters but could also be understood as those who present in gender fluid or androgynous ways. Here monster-ness, and the articulation of the monster in drag, might be a purposeful rubbing up against the representations of queer and trans people as ‘monsters’ or as dangerous both historically and in the present day. This use of monsters, then, implies a murky and broad set of performance and aesthetic forms as well as a political positionality. Drag things are perhaps harder to define, where things-ness might imply the inclusion of technological, alien, and more-than-human aesthetics and in general a rejection and simultaneous extension of normative humanness.
These terms are expansive and are not meant to limit or further delineate different forms of drag or suggest some forms are more radical or subversive than others. In particular, this is not to add a value judgement on forms of drag that fit into the category of drag king or queen, which can offer complex engagements with both gendered forms and normative human-ness. Indeed, for many performers, presenting as a drag king or queen might be a radical position to occupy, whether due to their identity out of drag, where they are performing, or the type of work they perform. This is important, as much of drag performance work included in The Zizi Project (and in Zizi & Me and The Zizi Show in particularly) is, on the surface, relatively mainstream lip sync performance practice (as mainstream as these performers and performance forms can be in the UK in 2022). However, even the most ‘traditional’ drag queen lip syncing a ballad can offer radical interventions into normative embodiments of gender, and even the most radical of drag monster or thing might uphold them, much like Judith Butler notes in Gender Trouble.Footnote11 These terms – kings, queen, monster, and things – are not meant to cement or fix any definitions of drag, and any calcification of these terms should be actively resisted by drag audiences and scholars. Instead, I offer these terms as useful place-holders to articulate the various and complex representations of drag present in The Zizi Project.
The Zizi Project sits at the intersection of AI art and drag performance. Using AI and Machine Learning alongside performances and aesthetics created by drag performers, Jake examines key issues such as the representation of queer and trans bodies in datasets used to generate facial recognition software, the politics of the inclusion of queer and trans lives within datasets, and the poetics of AI drag performance. The project is currently made up of three distinct but connected iterations which are explored below: Queering the Dataset, Zizi & Me (created in collaboration with Me The Drag Queen) and The Zizi Show (created with thirteen drag performers). Zizi, the overarching drag act or character that runs throughout the Project, is not being made to replace human drag performers. Rather, Jake uses drag as a sort of queer method, a mode of hacking AI and Machine Learning systems and intervening in normative frameworks and processes, to test the limits of what AI can be and do. In so doing, through Zizi, I propose it is possible to glimpse drag futures where the digital and the material interact in increasingly complex ways to explore notions of access, inclusion, and ethics that speak to fundamental questions of what counts as drag, what counts as queer, and, indeed, what counts as human.
Queering the Dataset
In the video installation Queering the Dataset, Jake took a dataset of 70,000 images of human faces and inserted into it 1000 images of drag performers.Footnote12 Through training a Machine Learning algorithm used to generate fake images on this augmented dataset, what it produced are an amalgam of drag faces, monsters, and things. Inserting the drag images – what Jake refers to as disrupting or even corrupting the dataset or injecting queerness into it – resulted in these monstrous, alluring, and queerly moving images, where the images move queerly and are queerly moving or affecting (see ).Footnote13 This notion of corruption as a tactic or strategy in Jake’s practice is particular to this project, where the aim is to reveal the limitations in normative datasets and in so doing propose fabulous and monstrous new formations of human-ness. It might be assumed that this sense of corruption might connect to queer negativity or queer anti-social politics, however, I link this practice to José Muñoz’s articulation of queer utopias where any articulation of a queer utopia contains within it an active critique and even refusal of a straight present.Footnote14 Positions of queer negativity and queer futurity are not, then, binary oppositions but instead complexly inter-related, and here this notion of corruption indicates a negative relationship to normative datasets that produces a future-oriented practice which I unpack further below.Footnote15
Image 1. 3 different manifestations of Zizi in Queering the Dataset. Image courtesy of Jake Elwes.
Queering the Dataset has been shown internationally in artistic and festival settings, as well as on Instagram (see ).Footnote16 The impulse to actively inject queerness into the normative framework of these datasets means the work remains unfixed but also speaks to the playful engagement with the technologies that characterises Jake’s work elsewhere, with pieces such as dada da ta (2016) and Machine Learning Porn (2016) offering purposefully silly yet provocative deployments of AI algorithms in artistic work.
Image 2. Screenshot of Zizi on Instagram, taken by the author.
In my interview with Jake, they describe how the project began from playing with the idea of the representation (or lack thereof) of queerness within the datasets used to train facial recognition software and to produce deepfake technologies:
What was interesting was subverting this idea of creating the most ‘realistic’ face, which often means the most ‘normative’ face. What I love are the mistakes, the failures, the times where you can see this is obviously a construction, a digital process that is making these faces, which sometimes completely breaks down. This is not thinking about building a deepfake body but instead about injecting queer bodies and images into an existing dataset.Footnote17
Here, I reflect on ‘real-ness’ as connected to certain drag histories, and in particular the ballroom scene popularised by the documentary Paris is Burning (1990) and more recently the television show Pose (2018–2021).Footnote18 In ballroom practices, striving for ‘realness’ indicates the performers’ ability to pass or be seen as legitimate within a category (for example to achieve ‘executive realness’ a performer is expected to pass within office settings). As Judith Butler notes in Bodies That Matter, ‘what determines the effect of realness is the ability to compel belief, to produce the naturalised effect. This effect is itself the result of an embodiment of norms’.Footnote19 For Butler, ‘realness’ relates to dominant modes of gendered (and racialised) performativity, where drag and ballroom cultures are understood to engage in a ‘performance’ deeply rooted in striving for the accomplishment of normativity. Notably, both Butler and bell hooks (amongst many others) articulate the limits of gendered and racialised notions of ‘realness’ which both enable and contain minoritarian subjects within injurious representational systems.Footnote20 This is connected to Lauren Berlant’s notion of ‘cruel optimism’, understood as that which is both sustaining and damaging. ‘Cruel optimism’ recognises those things to which subjects are attached, such as the promise of the good life or normative modes of being in the world, can both sustain individuals and communities and injure them.Footnote21
Performance and performativity, however, are distinct but connected terms. In their earlier thinking in Gender Trouble, Butler draws on drag as a key exemplar of the construction and performativity of gender, noting, ‘The effect of gender is produced through the stylization of the body and, hence, must be understood as the mundane ways in which bodily gestures, movements, and styles of various kinds constitute the illusion of an abiding gendered self’.Footnote22 Importantly, gender is not being described as a performance (or a costume which can be put on and taken off at will) as is often still misunderstood by critics of this work, but instead as a temporally and spatially contingent set of acts and actions that cohere into normative understandings of gender. Regardless, this notion of performativity has permeated the fields of queer studies, gender studies, and performance studies in countless ways. Whilst many theorists (including Butler themself) have critiqued and developed this early position, these ideas remain. Drag performance still comes to represent the idea of gender as ‘constructed’ by drag performers, and those who study and watch it. Through these debates, drag as a performance has been central to understanding gender in the Global North since the 1990s (if not before). However, what is often missing in these discussions is that drag might offer not a re-performance of gendered norms, but instead a space in which to play, experiment with and expose the very limits of these gendered ideals – something which The Zizi Project does. Furthermore, the use of drag as an exemplar of the performativity of gender often fails to account for drag as a performance form with a distinct and complex set of histories and politics connected to and beyond the politics of gender, as Stephen Farrier notes.Footnote23 However, returning to these notions opens up pertinent areas for discussion surrounding how these emergent AI technologies can be used to both subvert and uphold the status quo in much the same way Butler articulated of drag in relation to gendered norms.Footnote24
Importantly, I am not setting out to draw an overly simplified link between ballroom cultures primarily connected to Black and Latinx queer and trans subcultures in the USA in the 1980s and 1990s and this contemporary AI project. Instead, my focus is the link between ‘real’ and ‘fake’ as connecting tissue that offers interesting modes of comparison between historic and contemporary drag forms and standards and contemporary languages used to describe the successes and failures of AI and deepfake systems.Footnote25 This connection between realness and normativity is reflected in Jake’s assertion that the desire for ‘realness’ in deepfake technology is often synonymous with a desire for normativity. The languages of real and fake, then, offer inroads into considering how technologies can be mobilised in service of and in resistance to gendered (and other) norms.
In an informal conversation with Jake, when discussing these complex ideas of realness and deepfake technology, they made a link between notions of realness and passing in relation to gender and notions of computers or AI systems passing the Turing test (a test to show if a computer has a ‘mind’, which is achieved when a human speaks to both a computer and another human and cannot tell which responder is the human).Footnote26 Jake cites Yuval Noah Harari’s Homo Deus: a brief history of tomorrow and their assertion that ‘[t]he Turing Test is simply a replication of a mundane test every gay man had to undergo in 1950 Britain: can you pass for a straight man?’.Footnote27 This connection between passing the Turing test to pass as human and passing in relation to codified and hegemonic forms of gender and sexuality places The Zizi Project within wider technological and philosophical debates surrounding how norms and forms are reproduced and resisted by marginalised subjects and communities. However, I am resistant to too neatly finding confluence between these diverse areas of gender and technology and caution against any simple readings of how technology might both resonate with and challenge norms of gender and sexuality. Instead of getting lost in this (admittedly pleasurable) theorisation around what Queering the Dataset might represent, I return to the work in question to consider its effects and affects, or what it does.
What initially intrigued me about Queering the Dataset was how the ever-morphing drag faces slowly picked up on recognisable images or forms. In looking at the various versions of the artwork on Instagram or in a gallery, fleeting visual references to recognisable drag performers can be picked out, only to move on to another less recognisable (and less recognisably human) face in the next moment. Drag monsters and things populate the work as the AI moves through the potential faces it has learned to generate from Jake’s queered dataset, navigating the latent space or ‘the limitations, boundaries and space containing everything a neural network has learnt’.Footnote28 For Jake, this concept of the latent space is at the centre of their thinking around the potential of AI as well as its limitations. In a blog post for Digital Democracies, they state:
In many ways, I believe that there is already an inherent queerness to latent space. While an algorithm may have been given the task to distinguish between male and female faces – due to the labels inputted by humans – the neural network in fact reduces input faces to numbers or coordinates in a latent space. If we remove our human classifications and binaries from this system, it doesn’t read images of faces as male or female. Instead, everything falls into spectrums or points in this fluid and continuous space. Binaries and labels can be read back into these systems depending on where in this space an image is recognised to be. But we can also explore the unmediated mathematical latent space. Zizi – Queering the Dataset takes an unsupervised journey through this space. In the video, you can watch as it continuously and effortlessly moves through different identities and personas.Footnote29
As such, in Queering the Dataset utopian possibilities might well be seen. There are glimpses of what José Muñoz might describe as ‘a kernel of political possibility within a stultifying straight present’.Footnote30 However, beyond a political possibility, I also locate a potentiality within this work. As Muñoz theorises,
Possibilities exist, or more nearly, they exist within a logical real, the possible, which is within the present and is linked to presence. Potentialities are different in that although they are present, they do not exist in present things. Thus, potentialities have a temporality that is not in the present but, more nearly, in the horizon, which we can understand as futurity.Footnote31
Queering the Dataset, with its shifting, morphing, and unceasing litany of drag faces, offers glimpses of potentiality and futurity beyond the boundaries of what AI understands as human and beyond the boundaries of what can be readily understood as drag. Whilst there are clear references to specific drag performers and performance forms, there are also these monstrous moments that defy human logics (and, indeed, biology) to speak to the future of drag whilst still being connected to drag histories and presents. Ultimately, whilst body-less, the images created refer back to drag bodies that queered the normative dataset used to generate the work and simultaneously make utopian gestures towards technological and embodied forms which move beyond binarized recognitions of gender and drag forms.
So far, however, the drag explored remains aesthetic (it is about what drag looks like and what these images do to normative datasets) rather than performance-based. Following Farrier’s impulse to consider drag as a performance form and see ‘what drag does, enacts or brings about’ does not mean to ignore the impact of gender but instead to explore drag ‘without a singular focus on gender’.Footnote32 To attend to the potential of drag performance in relation to these technologies, I move to consider the second iteration of the project: Zizi & Me.
Zizi & Me
Zizi & Me is an ongoing collaboration between Jake and Me The Drag Queen, a drag performer known for her lip sync prowess and hosting on the London drag scene.Footnote33 The performance is currently a video piece made during the lockdown in 2020 but is being extended into a live performance. Early versions of this live performance have been shared at the Gazelli Art House in London (July 2021), which I discuss in the conclusion to this article, and the Zabludowicz Collection in London (June 2022) as part of wider events about drag and AI. In the video examined here, Me performs alongside a deepfake version of herself, lip syncing to ‘Anything You Can Do I Can Do Better’ from Annie Get Your Gun (1946) (see ).Footnote34
Image 3. Video still from Zizi & Me. Left: Zizi, Right: Me The Drag Queen. Image courtesy of Jake Elwes.
In Zizi & Me, Jake created a Machine Learning model of Me by filming her in drag walking around and moving. In this instance, the ‘dataset’ that was used to create the deepfake was all the filmed footage of Me. Then, from that dataset, they were able to get that machine learning model – or AI version – of Me to do anything. However, with a key tenet of The Zizi Project being concerned with the politics and ethics of engaging with these systems, Jake and Me worked in collaboration to explore what these technologies can be used for in performance rather than exploiting Me’s image through the deepfake process. This mode of collaboration with drag is something that extends through the project, but is epitomised in the relationship between Jake and Me, who continue to work closely, both being informed by one another’s artistic and political sensibilities.
Jake described the creative process during our interview
For Zizi & Me, once we had that model, we decided on a performance. We chose ‘Anything You Can Do I Can Do Better’ […] because it is about competition and gender. This competitive nature starts to say a lot when it is an AI deepfake performer and a human drag performer. The song becomes a metaphor for AI and society: the idea that AI is going to be able to do everything better than us in the near future, which I think is actually a lie. For this project, what made the most sense was for Me to be performing the movement which the deepfake would turn into the drag performer. Ben [Me’s name out of drag] would perform the ‘Annie’ part out of drag, and effectively the AI would be applying the drag make-up and costume to Ben’s body. That is really interesting: to see this drag transformation happening through a completely digital process; a strange augmented, digital deepfake process where someone gets put in their drag persona through a digital process.Footnote35
There are exciting moments in the video during which the AI fails to accurately make a ‘real’ image of Me (or anything recognisably human once again), showing the limit of the dataset and the implications of limited datasets. For example, there is one moment where Me out of drag has turned to the side and, because in the creation of the model Me did not turn to the side for long enough, the AI fails (see ). This failure demonstrates a fundamental point: if a dataset does not contain something, the AI cannot learn it. One reading of this moment in the performance is as a simple metaphor for the issue of diversity in datasets, as articulated and critiqued by Buolamwini and the AJL explored above. However, as noted above, only focussing on inclusion can fail to tackle the deeper structural issues that inhibit access to these forms of technology and the wider discriminatory practices they can produce. Whilst Zizi & Me offers a playful drag interpretation of an old-school musical theatre number, it also intervenes in these conversations about how AI can be used by offering insights into what queer performance might do with AI and how AI might be queered by drag.
Image 4. Screenshot from Zizi & Me demonstrating a moment where the deepfake fails to accurately replicate Me turning sideways. Taken by the author, shared courtesy of Jake Elwes.
What is most compelling in this work is how these failures reveal the cracks in technology, and in performance they also open up possibilities for something new and unexpected. Discourses surrounding failure are now ‘well and truly established in theatre and performance scholarship’, and queer failure as a method and area of enquiry has both drawn from and extended these ideas.Footnote36 The failures in Zizi & Me are connected to these wider performance-based and queer discussions.Footnote37
In the video performance Zizi as a deepfake of Me, who fades in and out of legibility as a drag performer (and, indeed, a human), offers an exploration of the futures of queer digital/AI performance practice where the deepfake drag performer is not being made to perfectly simulate human movements and performance forms. Instead, using the already hyperreal or hyperbolic performance of drag, Zizi pushes the limits of what drag performance can be and do. There are exciting moments of failure such as a note held longer than humanly possible, or Zizi’s face entirely disappearing as they sing ‘softly’. These and other failures to remain human (for any given hetero- and cis-normative value of what human might look like or what a ‘natural’ human might be) that emerge through Machine Learning processes, and are further manipulated through the video editing, start to move beyond a merely representational deepfake (trying to trick us into thinking it is a real drag queen) and into something much more exciting. If Me, as a drag queen, is already playing with the limits of gender with her clown-like and exaggerated make-up and aesthetics and is lip syncing a male voice in the act, then the limits of the ‘real-ness’ of the performance are already called into question. Therefore, the ethereal movement of the feathers on Zizi’s dress and the crumpling of the face and body through the technology’s failure do not detract from the performance but offer up new horizons of queer performance practice in relation to technology.
Zizi is not limited by human forms, but instead locates modes of performance practice that revel in and expose these limitations and push beyond them into new frontiers. These kernels of potentiality arise here, much as they did in Queering the Dataset, as small moments where alternative queer futures are glimpsed, and where the human and the technological interact in increasingly complex ways to push at the boundaries of drag and queer performance, and the human form itself.
The Zizi Show
The Zizi Show was a continuation and extension of Zizi & Me. During the ongoing COVID-19 restrictions in 2020, Jake and Me wanted to find ways to support queer performers and venues that had been adversely affected by the pandemic, using funding from Edinburgh Futures Institute (EFI) to pay performers and a live venue in which to film (The Apple Tree in Clerkenwell). What resulted was an interactive online cabaret where audiences watch deepfake versions of drag performers and switch between performances (and, within those performances, different performers) to create a bespoke drag experience (see ).Footnote38
Image 5. Screenshot from The Zizi Show website (www.zizi.ai). Image courtesy of Jake Elwes.
It was created using similar processes to Zizi & Me, with thirteen performers being filmed walking around to create the datasets for new deepfake drag acts. Of those, five performers then created lip sync performances which were filmed to be used as the movement to control the deepfake drag performers. The performers all came from a cross-section of the drag performance scene in London and were chosen as broadly representative of the different identities and performance forms on the scene.
As well as Me The Drag Queen, The Zizi Show included Bolly-Illusion, Cara Melle, Chiyo, Dakota, Lilly Snatchdragon, Luke Slyka, Mahatma Khandi, Mark Anthony, Oedipussi Rex, Ruby Wednesday, Sister Sister, and TeTe Bang. When watching the show the user can choose movement and lip sync performed by drag prinx Chiyo, with Chiyo controlling drag and burlesque legend Lilly Snatchdragon’s body, and you can see the whole body perform or a close-up of the face. You can also chose an amalgam performer, Zizi themself, who is made up of all the performers in the dataset. In this iteration Zizi merges between various iterations and combinations of all the performers as they perform (see ).
Image 6. Still from The Zizi Show with Zizi performing as an amalgam of all the deepfake drag performers. Image courtesy of Jake Elwes.
Image 7. Promotional image for The Zizi Show, showing a combination of different drag performers across six deepfake bodies. Image courtesy of Jake Elwes.
There was an ethical impetus in the process of creating The Zizi Show. These datasets were closed and the performers’ images could not be controlled by anyone. It was not possible to have a random member of the public using their body and movement to control Chiyo, for example. Instead, it was only movement made by other performers used in the process. This is important as, rather than engaging in existing datasets or practices of AI used more popularly, by Instagram filters for example, where anyone can use and distort the images provided, The Zizi Show explored the politics and poetics of creating bespoke datasets that contain queer and trans performers and bodies; bodies rarely present in the datasets used by AI engineers working on facial recognition, or bodies whose queerness is rendered invisible since it in unintelligible with (hetero)normative data collation and processing systems, for example. This is complex, however, since technology such as facial recognition software is being deployed ubiquitously in law enforcement and the employment sector, therefore, being seen or recognised by these systems is not necessarily desirable for queer and trans communities. As the AJL note, greater inclusion in datasets could lead to more effective discriminating systems.
When discussing this aspect of the performance, Jake noted
it was very controlled. At one point there was an interesting conversation around ‘should this be open? Should anyone be able to upload their persona and become part of the website?’. However, besides not having the computing power to train a deepfake – which takes multiple days for everyone who uploads images – I also wanted to be quite protective and think about consent in terms of the fact that all the people who are involved in the project are willing for their image to be distorted and altered and are aware of what is going on behind that […]. Trans and non-binary representation is a really important issue. This project is not only about trans identities in relation to AI but it does explore it in the sense that we are thinking about gender and queerness in relation to AI. There is a high likelihood that trans people are going to be even more heavily discriminated against when these systems are used by governments and corporations.Footnote39
These ethical questions inform the practice and, instead of addressing the balance of queer inclusion in mainstream settings, the project became about the pleasures of creating work by and for these communities. This pleasure is present in multiple ways, including in the pleasure of seeing the AI fail to make the deepfakes ‘real’ and the pleasure of seeing something so technologically complex be deployed for a performance form which is typically analogue. There is further pleasure in seeing the different bodies move with and through one another, particularly for those looking at it with a knowledge of London’s drag scene. At the RVT in 2020, during the show explored in the introduction, the performers took pleasure in having their bodies being controlled by one another and delighted in picking out the movements, gestures, and styles of the different performers. This pleasure of being seen is a simple one, but an important reminder of the pleasure, politics, and risks of being seen and being included. In a review of The Zizi Show in Volupté: Interdisciplinary Journal of Decadence Studies, Owen G. Parry suggests, ‘Zizi represents a potentially new icon of liberty and democracy for our time’.Footnote40 Whilst perhaps a grand statement, there is here a sense of the complex politics at play in this work both for queer performance communities and beyond.
The Zizi Show starts to sketch out new frontiers of queer performance practice that interact directly with AI and Machine Learning. Through the complex interplay of bodies in the show, and how the performances slide between virtuosic drag lip syncs and fantastic failures where bodies and architecture merge and create new morphological forms beyond human imagination, The Zizi Show articulates queer digital forms in more-than-human ways. In resisting anthropomorphising AI and by aiming to use drag to demystify AI (something which was part of the remit of the Edinburgh Futures Institute funding), something queerer has emerged. In the process of demystifying AI (and here I take the term ‘demystify’ literally, as blowing the mist away) other horizons of practice and performance at the intersections of drag and technology are starting to form at the periphery of what artists and researchers can see and imagine. Zizi allows performers and academics to take the first steps towards that terrain.
Conclusions
In July 2021, Zizi had its first live incarnation at the Gazelli Art House in London where, amongst other acts, Zizi and Me performed live, with a lip-sync performance to ‘Me and My Shadow’ by Frank Sinatra and Sammi Davis Jr. and Judy Garland and Barbra Streisand’s iconic ‘Get Happy/Happy Days Are Here Again’ duet. There was a complex interplay of temporalities onstage as Me live in the space lip-synced and, reflecting afterwards, was also aware of the AI projected body next to her which was generated on movement and lip sync she had performed (out of drag) a few weeks previously. The complex inter-temporality often present in drag, and lip sync performance in particular, is explored by writers including Farrier who articulates drag lip syncing as ‘serving a non-heteronormative heritable link with the past’.Footnote41 Elizabeth Freeman, in a re-reading of queer belonging, similarly offers the reflection that drag performance ‘seems to be a matter of not only performing but also enacting, summoning, even willing “sympathy, friendship, or love” between the dead and the living’.Footnote42 These temporal readings of drag articulate the co-presence of past, present, and, indeed, future in drag performance, with live bodies mediating voices from the past, sometimes the voices of those who are no longer with us. Within The Zizi Project’s live incarnations, this inter-temporality is only made more complex by the co-presence of the body of the performer in both live and digital incarnations, mediated by AI technology, a video screen, and the recorded voices of these songs from the past. The techno-temporality of this and future live iterations of Zizi is a useful lens to consider how queer performance practices might be further queered through the presence of technological forms and live bodies simultaneously.
Reflecting back on The Zizi Project from its inception in the video art piece Queering the Dataset to the video performance Zizi & Me and the interactive digital cabaret The Zizi Show, I am struck by the scope and scale of the work alongside its purposeful irreverence. Attempting to create deepfake drag performers to play with and expose the fallacies surrounding AI as well as challenge normativites and bias within Machine Learning datasets is a bold and ambitious project. The project is shot through with failure in various ways; both the AI’s failure to create a deepfake drag performer and the failure of Zizi to take itself too seriously. This is not a limitation of the project but its potentiality.
Failure and not being serious, as Jack Halberstam notes, is a vital part of queer theorising, art making and living:
Being taken seriously means missing out on the chance to be frivolous, promiscuous, and irrelevant. The desire to be taken seriously is precisely what compels people to follow the tried and true paths of knowledge production around which I would like to map a few detours. Indeed terms like serious and rigorous tend to be code words, in academia as well as other contexts, for disciplinary correctness; they signal a form of training and learning that confirms what is already known according to approved methods of knowing, but they do not allow for visionary insights or flights of fancy.Footnote43
In these discussions about The Zizi Project I have been struck by the investment in failure that runs through all the work as a productive artistic tool that also exposes the limits of normative frameworks inherent within AI and Machine Learning. These normative frameworks are part of the building blocks of many of these systems precisely because they are part of the building blocks of the world we inhabit. The Zizi Project works alongside academic and political interventions into the politics and ethics of deepfakes and the complexities of algorithmic bias explored in the introduction. It is one point amidst a constellation of resistive practices, and offers a playful and irreverent glance at these structures. However, these playful and silly interventions are also sensitive, critically aware of the precarious and complex positions inhabited by the performers and people represented. The Zizi Project offers not the flattery of legitimacy in an artistic project, but instead a chance to become part of the conversation through their presence in these highly complex art works.
The Zizi Project takes a queer glance at AI, Machine Learning, and the wider politics and practices of drag. It moves through systems of AI and the normativities it (re)produces and, with clacking heels and stomping boots, begins to drag (to exert a pull) on the constant acceleration of technologies that often leaves behind (or actively excludes) the diverse queer identities and positionalities present in LGBTQ+ communities. I have written elsewhere about the productive drag (or pull) drag can exert on future-oriented narratives, drawing on Freeman’s notion of temporal drag.Footnote44 Freeman in particular is considering the drag that seemingly historical feminist identities such as lesbian feminists can exert on queer fluidity not as a negative position but instead as something that can offer useful insights for present and future formations of queer identities, communities, and politics. I use this thinking to consider how drag performance can exert a drag on contemporary neoliberal forms which commodify drag. I re-deploy that here, suggesting that in bringing queer bodies and performers into direct contact with AI, The Zizi Project stages a moment of queer resistance to the normativites re-produced in these emergent technologies, as well as asking important questions about what it means for queer subjects to be included in these datasets in the first place.
Ultimately, Zizi engages with these complex ethical and political debates with a knowing wink, a playful and purposefully silly examination of the charged debates around AI and Machine Learning in the future of our societies. In dealing with grand narratives in local ways, The Zizi Project offers queer tactics of resistance to these technological forms (and norms) and in so doing it also platforms a diverse range of performers and performance forms in more mainstream artistic settings. Whilst we may never have the army of artificially intelligent drag terminators we all (or, at the very least, I) dream of, Zizi instead teaches us the value of failure as a queer tool and the possibility of resistance from both within and beyond the systems by which we are oppressed and which we may be complicit in preserving. In creating worlds of kings, queens, monsters, and things, dangerous and delightful queer performance futures are starting to be revealed.
Notes on Contributor
Joe Parslow is a queer researcher, writer, teacher, and producer. They are a Lecturer in Popular and Queer Performance at The Royal Central School of Speech and Drama, where they also support research ethics and integrity and practice research. They have worked extensively in queer nightlife as a producer of drag and cabaret events in LGBTQ+ spaces in London.
Notes
1. Ian Sample, ‘What are deepfakes – and how can you spot them?’, Guardian, January 13, 2020, https://www.theguardian.com/technology/2020/jan/13/what-are-deepfakes-and-how-can-you-spot-them (accessed July 13, 2022).
2. GANs work by a process of collaborating neural networks: a generator and a discriminator. In the creation of deepfake images, for example, the discriminator has learned and categorised a dataset (for example, images of human faces) and the generator starts to produce images. These fake images are judged by the discriminator as being images from within the original dataset or not. I am indebted to Jake Elwes for endless conversations explaining this. This language of discrimination (judging what images get to ‘pass’ as real, for example) is resonant of how gender and drag are policed.
3. Sample, ‘What are deepfakes’.
4. Legacy Russell, Glitch Feminism: A Manifesto (London: Verso, 2020), 23.
5. See, for example, AIArtists.org, ‘AIArtists.org: The world’s largest community of artists exploring Artificial Intelligence’, https://aiartists.org (accessed November 30, 2021); Algorithmic Justice League, ‘Algorithmic Justice League - Unmasking AI harms and biases’, https://www.ajl.org/ (accessed November 30, 2021); Catherine D’Ignazio and Laruen F. Klein, Data Feminism (Cambridge, Mass: The MIT Press, 2020); Kate Crawford and Trevor Paglen, ‘Excavating AI: The Politics of Training Sets for Machine Learning’, September 19, 2019, https://excavating.ai (accessed November 30, 2021); Safiya Umoja Noble, Algorithms of Oppression: How Search Engines Reinforce Racism (New York: New York University Press, 2018).
6. Algorithmic Justice League, ‘Algorithmic Justice League - Unmasking AI harms and biases’, https://www.ajl.org/ (accessed November 30, 2021).
7. Joy Buolamwini, AI, Ain’t I A Woman?, YouTube, June 28, 2018, https://www.youtube.com/watch?v=QxuyfWoVV98 (accessed November 30, 2021).
8. AIArtists.org, ‘Artificial Intelligence Timeline [UPDATED 2020]’, 2020, https://aiartists.org/ai-timeline-art (accessed November 30, 2021).
9. A BBC documentary called Kill Your TV (2019) presented by comedian Jim Moir explores this particular history, including Jake’s previous AI artworks.
10. Jake Elwes, interview with the author, London, January 25, 2021.
11. Judith Butler, Gender Trouble: Feminism and the subversion of identity (London: Routledge, 1990), 176-7.
12. Tero Karras, Samuli Laine, and Timo Aila, FFHQ Public Domain Dataset. Python. 2019. Reprint, NVIDIA Research Projects, 2021, https://github.com/NVlabs/ffhq-dataset (Accessed January 5, 2022).
13. Jake Elwes, ‘Jake Elwes - Zizi - Queering the Dataset’, 2019, https://www.jakeelwes.com/project-zizi-2019.html (Accessed January 5, 2022).
14. José Esteban Muñoz, Cruising Utopia: The Then and There of Queer Futurity (New York: New York University Press, 2009).
15. For more information on queer negativity or the anti-social turn in queer theory, and its relationship to queer futurity, see Lauren Berlant and Lee Edelman, Sex, Or The Unbearable (Durham, NC: Duke University Press, 2013).
16. Zizi (Deepfake AI Drag) (@zizidrag), ‘Instagram Photos and Videos’. https://www.instagram.com/zizidrag/ (accessed 9 December 9, 2021).
17. Elwes, Interview with the author.
18. Paris Is Burning, dir. Jennie Livingston, United States: Off White Productions Inc., 1990; Pose, created by Ryan Murphy, Brad Fulchuk, and Steven Canals, Netflix: 2019-2021.
19. Judith Butler, Bodies That Matter: On the Discursive Limits of ‘Sex’ (Abingdon: Routledge, 2011 (1993), 87.
20. Ibid., 81-97; bell hooks, Black Looks: Race and Representation (New York: Routledge, 1992), 145-56.
21. Lauren Berlant, Cruel Optimism (Durham, NC: Duke University Press, 2011).
22. Butler, Gender Trouble, 179.
23. Stephen Farrier, ‘That Lip-Syncing Feeling: Drag Performance as Digging in the Past’, in Queer Dramaturgies: International Perspectives on Where Performance Leads Queer, ed. Alyson Campbell and Stephen Farrier (Basingstoke: Palgrave Macmillan UK, 2016), 192-209 (192).
24. Butler, Gender Trouble, 176-7.
25. I am grateful to Stephen Farrier for helping me to make these links.
26. For further explorations of the Turing Test and its contemporary relevance see, for example, Eric Neufeld and Sonje Finnestad, ‘In Defense of the Turing Test’, AI & SOCIETY 35, no. 4 (December 2020): 819–27; Saygin Pinar, Ilyas Cicekli, and Varol Akman, ‘Turing Test: 50 Years Later’. Minds and Machines 10, no. 4 (2000): 463–518.
27. Yuval Noaḥ Harari, Homo Deus: A Brief History of Tomorrow (London: Harvill Secker, 2016), 265.
28. Jake Elwes, ‘Zizi: Queering Datasets and Latent Spaces’, Digital Democracies, December 14, 2021, https://digitaldemocracies.co.uk/zizi-queering-datasets-and-latent-spaces-jake-elwes/ (accessed January 4, 2022).
29. Ibid.
30. José Esteban Muñoz, Cruising Utopia: The Then and There of Queer Futurity (New York: New York University Press, 2009), 49.
31. Ibid., 99.
32. Farrier, ‘That Lip-Syncing Feeling’, 192
33. More information can be found about Me The Drag Queen on her Instagram: @methedragqueen. In academic work, Me has been discussed under her previous drag name, Meth, in Farrier, ‘That Lip-Syncing Feeling, 200-209
34. Jake Elwes, ‘Zizi & Me’, 2020, https://www.jakeelwes.com/project-zizi-and-me.html (accessed January 4, 2022).
35. Elwes, Interview with the author.
36. Tony Fisher and Eve Katsouraki, ‘Introduction: Knowing Failure’, in Beyond Failure: New Essays on the Cultural History of Failure in Theatre and Performance, ed. Tony Fisher and Eve Katsouraki (Abingdon: Routledge, 2019), 1-18 (3); Jack Halberstam, The Queer Art of Failure (Durham, NC: Duke University Press, 2011).
37. I return to Halberstam’s thinking in the conclusion.
38. Jake Elwes, ’The Zizi Show - A Deepfake Drag Cabaret’, https://zizi.ai/ (accessed January 4, 2022).
39. Elwes, Interview with the author.
40. Owen G. Parry, ‘Review: Jake Elwes, The Zizi Show – A Deepfake Drag Cabaret (Edinburgh: The New Real, Edinburgh Futures Institute, 2020)’. Volupté: Interdisciplinary Journal of Decadence Studies 4, no. 2 (2021): 203-207 (207).
41. Farrier, ‘That Lip-Syncing Feeling’, 198
42. Elizabeth Freeman, ‘Queer Belongings: Kinship Theory and Queer Theory’ in A Companion to Lesbian, Gay, Bisexual, Transgender, and Queer Studies, ed. George E. Haggerty and Molly McGarry (Oxford: Blackwell Publishing Ltd, 2008), 295-314 (309).
43. Halberstam, The Queer Art of Failure, 6.
44. Joe Parslow, ‘Not Another Drag Competition: From Amateur to Professional Drag Performance’, Performance Research 25, no. 1 (2020): 18–24; Elizabeth Freeman, Time Binds: Queer Temporalities, Queer Histories (Durham, NC: Duke University Press, 2010).