The Zizi Show by Jake Elwes (2020)

The Zizi Show, by Jake Elwes, is an interactive artwork in which a Generative Adversarial Network (GAN) has been trained on video footage of thirteen diverse 'drag' performers, filmed at a London cabaret venue during the Covid-19 lockdown.

In the work, Elwes which explores the intersection of AI and drag performance, and performance and human identity, in the new real. Drag challenges gender and explores otherness, while AI is often mystified as a concept and tool, and is complicit in reproducing social bias. Zizi combines these themes through a deepfake, synthesised virtual drag avatar created using machine learning. Zizi empowers the drag and LGBTQ+ community by a positive application of deepfake technology, exploring what AI can teach us about drag, and what drag can teach us about A.I.

Video: Making of The Zizi Show, Jake Elwes

The work revolves around captivating, beguiling imagery of AI-generated drag avatars. An algorithmically generated compere asks the audience to select performers and songs. Each performer has a body blended from video capture of drag performers that morphs and changes as they perform each work. As they change and flow between personas and identities, they glitch and breakdown, exposing software artefacts and through those their artificial, constructed character. In the online version, the artist draws on the visual tropes and interaction forms of cabaret theatre to design the user journey online. The audience view the output in different settings, and are able to select from a menu to switch between AI generated personas of drag artists for different music tracks.

The Zizi Show by Jake Elwes, 2020, Still of deepfake drag artist close up. Image © Jake Elwes

Zizi is mercurial, defying categorisation. Trans, queer and other marginalised identities are shown visibly breaking down, illustrating how AI struggles with ambiguity. Zizi is lossy, the continuity of the original human performers is unrecoverable. It invites us to reflect on harmful bias in society today, yet is also a celebration of difference. Zizi gives us joy and rage at once, it allows us to see the injustice and to look beyond it towards a vision that is enriching. It has empowered the people who volunteered their data and, as a work of art, it is truly astonishing.

The Zizi Show was commissioned by The New Real in 2020 and presented at the eponymous The New Real exhibition at Edinburgh International Festival in 2021.

The Zizi Show by Jake Elwes, 2020, Web interface. Image © Jake Elwes

Art as explanation

Zizi exposes the latent space of a machine learning model, and highlights the way the model outputs are shaped by the training data. Where many generative works have been trained on opportunistically collected data, the purposeful curation of Zizi’s dataset explores the question of how human identity is represented within complex models. The Zizi Show develops this through digital avatars, that have been learned from real performers to create an interactive work that allows user control. Significantly, it connects low level technology to high level, social, cultural and political aspects of AI, such as ideas of cultural appropriation and machine bodies. It exposes the limits to machine intelligence, and inverts what is otherwise a deficiency in the technology, through a positive use of deep fake technology, in which a marginal identity is celebrated and embellished, rather than obscured or misrepresented.

The Zizi Show by Jake Elwes, 2020, Training process. Image © Jake Elwes

Zizi is an explanation of bias in ML and the power of the dataset through experiential means. Zizi highlights the way data and design choices shape what ML does. It shows how the model learns a representation of people, that is embedded social life. Zizi engaged a marginalised group, developing their literacy surrounding bias in ML, thereby supporting their agency in contesting its fairness and accountability. Zizi shows end users there is something to contest, even if that do not interact directly with the model themselves. Zizi specifically targets anthropomorphised misrepresentation of AI, by constructing an AI persona, and then deconstructing it, and exposing its construction in software by the human artist.

The Zizi Show by Jake Elwes, 2020, Deepfake generating process for drag queen Lilly Snatchdragon. Image © Jake Elwes

The Zizi Show generates imagery of non-binary bodies in order to bring attention to the underrepresentation of LGBTQ+ communities in ML training data. It is an explanation through experiential means of a dense clustering of issues: discriminatory design (see also Parry 2021), bias in ML, lack of representation, non-binary identities, the unclassifiable character of real bodies, anthropomorphism in AI. Zizi specifically targets anthropomorphised misrepresentation of AI, by constructing an AI persona, and then deconstructing it, and exposing its construction in software by the human artist. By highlighting correspondences between AI and drag at a surface level, it asks deeper questions about the character of statistical knowledge applied to shifting human identities.

Empowerment

A south London community of drag artists were engaged throughout, providing them with positive representation, safe spaces (an in-person venue and a secure server), paid employment, accreditation, agency over the way data is stored, during Covid lockdowns. The project engages a marginalised group, and develops their literacy surrounding bias in ML, thereby supporting their agency in contesting its fairness and accountability.

Technology

The project engages with the current wave of machine learning techniques, using a StyleGAN network architecture re-trained on a modified version of Flickr-Faces-HQ (FFHQ) dataset, to which an additional 1,000 portraits were added, alongside a custom video, sound and interactive web interface. In Zizi, the artist interacts with the model by manipulating data and weightings. Machine learning here allows the creation of a generative space that includes bodies and faces.

The Zizi Show by Jake Elwes, 2020, Luke deepfake training. Image © Jake Elwes

Footnotes

For a review of the work and discussion of audience experience see Owen G. Parry's review in Volupté (2021).

The Zizi Show on the artist's website

The Zizi Show on Edinburgh International Festival website

The Zizi Show is part of a wider body of work The Zizi Project by Jake Elwes. Zizi - Queering the Dataset, the first work in the series, and The Zizi Show were commissioned by The New Real in 2019 and 2020 respectively. The works were original creations by the artist, developed during the artist's participation in a cooperative research study with The New Real.