by Annalisa Giacinti, studio photos by Mackenzie Walker // Nov. 25, 2025
Holly Herndon and Mat Dryhurst own a studio right next to the Schöneberg Kulturhaus, a state-owned complex of spacious, affordable artist studios which, however, they didn’t qualify for, being above the income threshold. The irony is that their space is much smaller, consisting of three compact rooms that serve mainly as a meeting place. Their GPUs live at home, 15 minutes away from here; they use another studio for sound and one in Wedding for 3D printing. “Our studio is weird. It’s a bit decentralized. We’re kind of spread out,” Dryhurst says with satisfaction as he smokes a cigarette at his doorstep, while we wait for Herndon to arrive.
Decentralization for the artist duo doesn’t just describe their spatial setup, but it informs the philosophy undergirding their practice, their approach to AI and the development of machine learning models, as well as the byproducts of these processes, such as the music albums and AI tools they released and performances and exhibitions they engineered. ‘Starmirror,’ currently on view at KW Institute for Contemporary Art, is a collaboration with the Kunstsammlung Nordrhein-Westfalen in Düsseldorf, where the second part of the exhibition project will be shown next Spring. The development of Choral AI presented in ‘Starmirror’ builds on ‘The Call,’ an R&D and exhibition partnership between Serpentine Arts Technologies, Herndon and Dryhurst, curated by Hans Ulrich Obrist and conceived as the culmination of nearly a decade of work at the junction of art, music and machine learning. ‘Starmirror’ is an immersive sound installation that draws on earlier group singing projects, and simultaneously gathers a live environment where new recordings take place as a continuation of their research into choral AI models through live community sessions, which started with Herndon’s album ‘PROTO’ (2019).
While in London they wrote a songbook rooted in English choral tradition, in Berlin Herndon and Dryhurst chose to steer from the traditional German canon, gravitating instead towards the vast repertoire of Hildegard von Bingen, the Benedictine abbess and polymath whose visions, mysticism and compositions made her a proto-experimentalist of the High Middle Ages. “She would receive these visions. She was kind of an early mystic, but also part of the church. She was a poet and dabbled in medicine. She was a very interesting character,” notes Herndon. Bingen also wrote the earliest known morality play, ‘Ordo Virtutum’ (1151), in which a soul must choose between good and evil—a genre that, according to the artists, aptly resonates with contemporary debates around AI’s ethical stakes. “We are kind of writing our own morality play around this moment.”
At the core of ‘Starmirror’ lies their view of AI as a collective achievement, paired with the belief that “machine learning models are an art medium in and of themselves, so every step in the machine learning model creation process is a moment for creative intervention.” To put this into practice, Herndon and Dryhurst turned the KW hall into a recording studio, where on several Sundays visitors can participate in a live AI training alongside a choir and the exhibition ensemble, all the while experiencing the physical installation they’ve built around it. “We’re creating art, but we want it to be legible, so there’s a kind of translation that’s happening.” The recordings will form a public choral dataset, which will be used to train a Berlin AI choir set to debut at the Kunstsammlung Nordrhein-Westfalen in June. Thus, the choral experience is the byproduct of model making; a tangible, embodied, aestheticized example of humans and machines cooperating artistically.
In their practice, the choir metaphor holds a twofold function. First, it bridges a gap in the public perception of AI—the paranoia that it will replace us all—attesting to the fact that machine training can be a human-sourced operation, arranged and coordinated by people, and built from real voices and stimuli. During the live sessions, between each sung chant, an AI-generated angelic voice amicably prompts the audience to “offer” and “provide” their contributions to the recording—a vocabulary that seems to deliberately foreground the audience’s agency, so as to counter the generalized anxiety surrounding knowledge extraction and appropriation.
Secondly, the choir structure mirrors the artists’ own methodology: “There’s so many different moving parts, basically everything we do is a very collaborative work, it’s almost like we’re directing a movie.” Each of their works is typically made of a software, a physical and a conceptual component, “which is largely us just losing our minds in our home,” Herndon half jokes. For the ‘Starmirror’ songbook, they collaborated with the University of Lille and medieval musicologist and composer Thomas Fournil, who translated the scores and adapted it for a machine-readable format. They also teamed up with their neighbors at Sub, a creative architecture and design studio that made their sketches architecturally sing; lighting designer Bianca Peruzzi; and engineers from Studio STREEV, who built the central wood ladder in KW’s hall.
The two artists describe their approach to AI as holistic, “where each step in the process is an art-making moment,” from making their models, building the datasets to inventing ways for people to interact with the models, down to the bureaucratic tagging of data. They insist that being transparent about data sourcing is integral to mastering “a beautiful way to make AI”: “There’s a wicked latency between the details of the machine learning conversation and the general public,” which they try to offset by ensuring that anyone contributing their voice remains part of the conversation. When participants were asked how they wanted their recordings to be handled, they chose to release them into the public domain, and Herndon and Dryhurst honored that decision: “The best idea always wins!”
By virtue of being public and made of many, the nature of their artwork becomes harder to grasp, its authorship necessarily blurry (Holly+, Herndon’s digital twin, represented an early example of this stance, an AI tool that allowed anyone to experiment with her AI voice and image). “These questions about authorship and IP, they’re less interesting than the idea of contributing to culture,” they maintain. They are adamant about continuing to challenge the very scaffold that once separated us from technology, and about dismantling anthropocentric myths upholding art-making, regarding art as a post-human endeavor instead—“which is why we always go back to a choir as a kind of model to illustrate this. It’s greater than the sum of its parts.”
Artist Info
Exhibition Info
KW Institute for Contemporary Art
Holly Herndon & Mat Dryhurst: ‘Starmirror’
Exhibition: Oct. 31, 2025–Jan. 18, 2026
kw-berlin.de
Auguststraße 69, 10117 Berlin, click here for map
