Divine Intervention

A week before Christmas, I went to the Berlin Cathedral for a concert of mostly Swedish songs performed by Stockholms Musikgymnasiums Kammarkör, a chamber choir of thirty high-school-aged singers. A centuries-old tradition in Sweden for the Feast of Saint Lucy, on December 13, the singers wore white robes, the girls with red ribbons around their waists and wreaths on their heads and the boys with tall, cone-shaped hats. Each of them held a candle in their hands, which supplied the only source of light during the performance in the massive, domed nave. They sang without microphones, their voices unmediated by any technology other than the acoustics of the cathedral’s architecture. The concert was striking, an impressive display of the pure power of the live human voice in harmony with others.

Two weeks later, I went to KW to see Holly Herndon and Mat Dryhurst’s exhibition Starmirror. I knew Herndon and Dryhurst’s work from Berlin’s music scene. Herndon is a singer, musician, and composer, and Dryhurst is a musician and researcher, and the pair are partners in work and life. I saw them perform Herndon’s 2019 album Proto that year at Volksbühne in Berlin and at Unsound in Kraków. The album features a “singing AI,” named Spawn, alongside Herndon, a human choir, and other collaborators. Spawn was trained on Herndon’s and Dryhurst’s voices as well as the choir’s singing.1 At the live performances, at which Herndon and the choir sang with Dryhurst attending to a table of technical gear, the audience was led in a singing exercise that was used to further train Spawn.

The album is pretty spectacular. At the time of its release, it felt novel. Herndon’s embrace of this still-developing technology didn’t feel like an empty gimmick; rather, the record was an exciting early experiment in the possibilities this technology could offer musical artists. Since then, AI technology has, of course, significantly advanced, fueled by obsessive investment that many are beginning to describe as a bubble.2 Either way, AI is becoming commonplace in daily life, often superfluously and despite persistent (human-made) biases and a lack of regulation that frequently result in troubling consequences.3

At KW, Herndon and Dryhurst presented a suite of new sculptural and sound works that were billed as extending the duo’s experiments with machine learning technology. At the ticket desk, I could faintly hear the singing emanating from the exhibition’s centerpiece, a sound installation, in KW’s hall. It sounded like a church choir, like the Swedish teens I had just listened to in the cathedral. Following the music, I missed the designated “first room,” which contained the exhibition’s introductory didactics, and ended up in the hall. Speakers were installed along the walls and simple wooden benches encircled a central sculpture, The Ladder (2025). Produced in collaboration with the design and architecture studio sub, the sculpture is a wooden ladder encased in a scaffolding of the same wood. It rose into an opening in a suspended ceiling of panels that produced soft, throbbing lighting that changed with each song.

As for the music playing, I prematurely assumed that what I was hearing was music generated by AI. It had the general characteristics of Christian choir music, but the songs weren’t immediately distinguishable to me. After scrutinizing the song list in the exhibition booklet, I realized that, of the ten distinct compositions listed, six of them were performed by human singers—three professional ensembles, a professional solo singer, and “an invited audience.” Four of these were compositions generated by AI models; the other two were human-made compositions, one that involved “AI-driven vocal processing” and another with no relation to AI. Three more human-made instrumental compositions were performed by The Hearth (2024), an acoustic organ constructed from the fans used to cool graphics processing units (GPUs), here with varying pitched whirrs. GPUs enable AI’s complex and energy-consuming computing, which was suggested here through a conspicuous mass of cables that emerged from the organ. As it played, I wondered about the environmental impact of the exhibition and of the duo’s practice generally, but the exhibition made no mention of AI’s tremendous resource needs.4

The remaining composition, Linked Diffusion (2024–2025), appeared in the song list three times. According to the description of this “choral AI output,” this was closer to what I had anticipated hearing in the exhibition: choral compositions generated by an AI model trained on the artists’ archive and recordings of more than fifteen community choirs performing musical exercises and hymns. There were almost no identifiable words or lyrics, just a series of notes that sounded as though sung through an electronic voice modulator.

At first glance, the installation recalled so many contemporary art installations, their sparse footprints and minimal architecture. However, echoing the hymnal quality of the music playing, the visual elements in the hall had a clear religious—that is, Christian—orientation. The exhibition booklet revealed that The Ladder, with its very literal ascent into lightness, materializes “references from religious iconography, architectural schemata, and computational models.” Specifically, it refers to The Ladder of Divine Ascent, a Christian treatise outlining the path to salvation; the exhibition’s muse, Hildegard von Bingen, a German Benedictine abbess and mystic from the Middle Ages whose visions included ladder motifs; and “ladders of abstraction,” a contemporary metaphor for the organization of data in machine learning’s neural networks. Von Bingen’s influence was also present in the music: three of the AI compositions were based on an AI model trained on her Ordo Virtutum from 1151, which the exhibition described simply as “a morality play in which a soul must choose between good and evil.” The songs are in Latin, and when the recorded voices swelled to a crescendo, my brain conjured the scent of frankincense. For what it’s worth, I most enjoyed the one song completely unmediated by AI.

More immediately identifiable religious iconography was apparent in The Hearth. Hanging at one end of the installation, The Hearth acted as the organ of the gallery-made-church. Arranged in two halves, each containing eight blocks of fans and between them a metal plate engraved with a biblical-looking baby, the construction also suggested an elaborate, multi-paneled altarpiece. On a number of Sundays—the Christian Lord’s Day, of course—the public was invited to “contribute their voices” in live trainings of the AI. The call-and-response exercise was led by the Starmirror Ensemble, supported by various Berlin choirs, dressed in white robes. The activation of the scenography bore an unignorable affinity to Christian ritual.

Herndon and Dryhurst see this connection to the divine as enlightening the public’s understanding of AI, writing, somewhat inscrutably: “We took inspiration from Hildegard von Bingen, who channeled visions from the divine to propose a new harmony between humans and the cosmos. We see analogies between her celestial hierarchy and the stack of influential protocols that determine our culture.” The exhibition further positions itself as addressing “a critical gap in the public perception of AI” by “offering a direct, participatory experience with the human contributions that define it.”

Unfortunately, the exhibition conferred the opposite impression. The artists’ decision to mediate AI through religious history and iconography functioned to further mysticize an already complex and opaque technology, and to situate it unequivocally in the worldview of Western Christianity. The exhibition presented AI as otherworldly, as beyond human and with agency of its own, a messiah. The opening line of the exhibition’s introductory text reads: “Starmirror proposes a scenario in which artificial intelligence (AI) models coordinate humans to generate intelligence for collective benefit.” Here, in painfully vague language that characterizes both art-speak and ChatGPT, AI is made to have the capacity to take action—to coordinate humans! Whatever that means—and to produce knowledge for a higher good, supposedly autonomously.

This is claimed without offering even a very simple description of how AI works, what kind of “intelligence” is generated, or what the supposed “collective benefit” might be. Like so much of the mainstream discourse about AI, the exhibition’s didactics blew a lot of hot air and offered little substance. The vast human, economic, hardware, and environmental infrastructures on which AI depends went unmentioned, as did the more common extractive, repressive, and violent applications of AI, most notably by police and military apparatuses. Instead of foregrounding human contributions to AI, the artists’ approach created, at times, a blindly affirmative image of the technology that obscured the role and agency of the humans involved.

The exhibition’s techno-optimism and the attendant erasures echo the rhetoric deployed by AI’s corporate gatekeepers (Anthropic, Google, Meta, Microsoft, Open AI, X, and many others), against whom Herndon and Dryhurst position their work. On view in the exhibition’s first room was Arboretum (2025), a skeletal wood architecture with 3D-printed rectangular plastic sheaths around its vertical studs. These plastic forms feature motifs generated by Public Diffusion, an image model trained by Herndon and Dryhurst on public domain data—according to the exhibition didactics, the largest public domain image dataset ever assembled, comprising forty million images. This is an impressive and, in some ways, noble feat. Yet, the artistic results were disappointing. The motifs are banal, emoji-esque forms: a key, a snail, a teacup, a bunch of grapes, a watch face, crystals, an hourglass, a gear, a chess piece. Taken seriously as an artwork, Arboretum is profoundly uninteresting, aesthetically and conceptually.

The artwork’s accompanying text asserts that the artists “position this work against the common fear that an AI-integrated future will reduce us to passive consumers of endless automated ‘slop.’” Are these 3D-printed motifs not also slop? What use is a forty-million-image dataset that produces forms that already exist in the world in their most generic iteration? What does it offer an artist, or anyone? Unfortunately, the text does not attempt answers—it makes no case for the use and benefits of the dataset or the public sphere. Instead, it continues that “For the artists, the problem lies less in AI itself than in the platform economy.” Yet, just a few lines earlier, visitors are invited to download the artists’ own app Starmirror through which they can add “images and concepts to the commons.” While I don’t deny that such engagement has the potential to “strengthen the public archive”—though, the false promises of the rhetoric of participation are well-established—their app is also a platform. Just because the dataset is public doesn’t mean it can’t also be exploited for profit. In effect, the artists are asking their audience to create data for a dataset that they, and theoretically others, use to produce work that is capitalized and proprietary—for instance, this very artwork. It’s an extractive logic not dissimilar to any private, corporate, for-profit platform where users pay for access with their data.

This extraction also plays out at the Sunday singing sessions. The artists use the recordings of the public singing in those live sessions as the foundation of a public choral dataset to train an “AI Berlin choir” that will debut at the exhibition’s presentation at the Kunstsammlung Nordrhein-Westfalen in Düsseldorf, opening in June. Herndon and Dryhurst are well aware that data is currency. But their project does not sufficiently distance itself from the powerful, exploitative forces they seek to subvert. Gathering visitors and leading them in song does not in itself forge meaningful social connection or mutual understanding—let alone generate intelligence for collective benefit. The aim of these works is data production and collection, not social production or solidarity. The exhibition’s impetus appears to be more a response to AI’s supposed novelty and imminent ubiquity than an interest in the production of resonant art.

My strong response to Starmirror was fueled in part by a frustration with the art industry’s attempt to stay relevant by promoting the intersection of art and technology, with an emphasis on AI, at any cost. For a few years now, exhibition programming, residencies and fellowships, research and project funding have been overwhelmingly focused on this topic. Starmirror introduced me to yet another initiative, the German Federal Cultural Foundation’s funding group “Art & AI,” which will spend up to 3.68 million euros between 2024 and 2028. The exhibition was one of eleven funded projects announced last year, each receiving somewhere between 100,000 and 240,000 euros.

This comes at the same moment that public cultural funding is being slashed across Europe, including in Germany. Artists are losing their studios, community-run art spaces are forced to operate on impossible budgets or shutter, humanities scholarship must accommodate itself to the trend or perish. It should raise a bright red flag that governments, and other funding bodies, are nevertheless eager to invest in AI art and related scholarship. AI is, ultimately and unavoidably, a technology of the ruling class and their capital. This trend compels art to serve their interests, to expand and advance the technology’s uses, and, more damningly, to enhance its attractiveness and cultural relevance. We should remain suspicious of artworks that integrate AI—become attuned to the ways they align with power, cooperate with oligarchic tech corporations, or reproduce the logics of neoliberalism, and ask about their environmental impact.

More importantly, we should hold that work to the same standards we do other art. Unfortunately, so much AI art is just digital content. In this way, Starmirror was generative in the other sense of the word. It prompted me to reflect on what my criteria are, what I think makes art interesting, and what it is that bothers me, formally and conceptually, about AI art. I thought of another immersive sound installation I had seen a few days earlier, Cornelia Parker’s Stolen Thunder (A Storm Gathering) (2025) at Kindl. Installed in the towering boiler house of the former brewery, the work is a collage of field recordings of storms from around the world. The recordings span one hundred years and are sourced from public archives and community repositories, including the BBC, Freesound, the Internet Archive, and radio aporee.5 Parker, with composer Graeme Miller, wove the recordings together into a cohesive ur-storm.

Parker and Miller’s “orchestration of found sound” unfolds over twenty minutes, beginning quietly, with birdsong and the patter of rain, and grows through gusts of wind and hurried footsteps to a booming, thunderous storm. A single, dim lightbulb suspended in the center of the vast space barely illuminates the speakers that hang from the ceiling and sit in the corners. Occasionally, a strobe light mimics lightning, revealing the white, in some places crumbling tiles that line the walls. The experience of it was visceral. I felt the bass of the thunderclaps in my body, the darkness and the sheer volume producing adrenaline. But there were also moments of comfort, memories of storms I waited out playing Scrabble by candlelight. Such uncontrollable terrors occur everywhere on the planet, across history. Civil society is, to some extent, so defenseless to their destruction that they can only be explained as acts of God. The storm is a powerful metaphor for the experience of being alive amid violence and chaos, submitted to forces of all kinds beyond one’s control.

Thinking about the production of a work like Stolen Thunder vis-à-vis AI-generated art, one of the key differences relates to process. AI technology favors—indeed requires—comprehensive, often exhaustive data to produce outputs that are similar, generalized, recombined versions of that data. Artists mine their world for inspiration—objects, materials, ideas, experiences—that they respond to, comment on, critique, intervene in, parody, appropriate, recontextualize. And they do so in highly subjective, discerning ways. AI’s human programmers undoubtedly retain control of what goes in, as with Herndon and Dryhurst choosing what data to include in any dataset, but the results are outsourced to the machinations of the algorithm.

Art-making, as I see it, is an idiosyncratic reorganization of “data.” A machine can only metabolize data in programmed ways. The affective potential of art—the awe, excitement, fear, sadness, pleasure it can produce—is an effect of artists organizing and engaging with their materials in strange, unexpected, unconventional, singular ways. Artists reframe, challenge, and restructure our everyday worlds, or take us out of them completely. AI eliminates the coincidence and subjectivity of being alive and mediating lived experience oneself. It produces content stripped of context, made unintelligible by the machine’s abstraction. I felt awe watching the Swedish teens fill the historical cathedral with their voices. I felt reverence taking in Parker’s choreographed storm. I felt mostly confused in Starmirror.

I think art—and, for that matter, any creative activity—is the realm we should most want to protect from the apparent impending “AI takeover.” Art doesn’t need to be optimized for or integrated into the tech economy. I want art—and writing—that doesn’t embrace this technology but defies it, critiques it, and demonstrates the value of human creativity.

The day I saw Starmirror, I went to Schaubühne to see Thomas Ostermeier’s production of Hendrik Ibsen’s The Wild Duck. Typical of German theater, Ostermeier, with Maja Zade, recontextualized the nineteenth-century play in the present. The main character, Hjalmar, is a lazy, seemingly talentless, and slightly alcoholic dad. He justifies daily naps on the couch as sources of inspiration for a “great invention” that will solve his family’s financial problems. In one of the many comedic moments in this otherwise deeply tragic play, Ostermeier and Zade’s rewriting has Hjalmar enthusiastically announce that his great invention is none other than: the integration of art and AI. The audience roared with laughter.

1 Emilie Friedlander, “How Holly Herndon and her AI baby spawned a new kind of folk music,” Fader, May 21, 2019, https://www.thefader.com/2019/05/21/holly-herndon-proto-ai-spawn-interview.

2 Ben Casselman and Sydney Ember, “The AI Boom is Driving the Economy. What Happens if It Falters?,” New York Times, November 22, 2025, https://www.nytimes.com/2025/11/22/business/the-ai-boom-economy.html.

3 See, for instance, Brady Brickner-Wood, “Grok and the AI Porn Problem,” New Yorker, January 14, 2026, https://www.newyorker.com/culture/infinite-scroll/grok-and-the-ai-porn-problem and Allie Funk, Adrian Shahbaz, Kian Vesteinsson, “The Repressive Power of Artificial Intelligence,” from Freedom on the Net 2023, Freedom House, https://freedomhouse.org/report/freedom-net/2023/repressive-power-artificial-intelligence.

4 For a good summary of this well-documented impact, see Adam Zewe, “Explained: Generative AI’s Environmental Impact,” MIT News, January 17, 2025, https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117.

5 A list of the recordings included can be accessed here: https://www.kindl-berlin.com/s/List-of-recordings_CorneliaParker_DE_EN.pdf

Genevieve Lipinsky de Orlov is an art historian, critic, and editor, and a cofounder and editor of The Public Review.

Installation images of Starmirror: photo: Frank Sperling

Image of Cornelia Parker, Stolen Thunder (A Storm Gathering): photo: Jens Ziehe