Conference Agenda

History rhymes. At least the genealogy of audio-content recognition algorithms makes clear that tools for watermarking copyrighted music have, of late, returned as tools for watermarking AI-authored works. This paper traces this unlikely trajectory from a sociotechnical battle against P2P file-sharing circa century’s turn to a consumer-side solution to the problem of proliferating AI-assisted music today. Drawing on ethnographic fieldwork with software developers, engineers, and entrepreneurs alongside a now-fading digital archive of press releases, industry whitepapers, and trade publications, I follow Gillespie (2007) by showing how the first instantiation of watermarking algorithms depended upon the perceived promise of Digital Rights Management (DRM) hardware to control end-user behavior, the “free market” be damned. I proceed to argue that the latter instantiation instead depends upon the neoliberal notion that consumers, given a choice between “synthetic” and “authentic” products, will reward human authors. Trevor Pinch’s and Wiebe Bijker’s social construction of technology (SCOT) framework helps explain how similar core technologies serve contradictory ideological purposes.

Pinch and Bijker (1984) posit that the social construction of a given technology unfolds in two stages: “interpretative flexibility” and “closure.” During the era of audio-content recognition’s interpretative flexibility, in the 1990s, it was unclear whether “fingerprinting” technologies (which encode audio in easily retrievable mathematical representations) or “watermarking” technologies (which embed imperceptible data into media to “tag” it for recognition) would prevail. Via Propellerhead’s loop-slicing software ReCycle and the dissertations of engineers Tristan Jehan and Brian Whitman, I show how composition, not copyright, was the first application for audio content-analysis algorithms; enterprise solutions have since been commercialized, achieving “closure,” or the narrowing of the range of interpretative flexibility to a specific use case, with Google’s Content ID. Thus, fingerprinting’s most profitable application proved to be enforcing user behavior. Its utility, however, depends upon the very intellectual-property regime that it supplanted. I conclude by arguing that the return to watermarking, via Google DeepMind’s Synth ID, is a symptom of the privatization (Tang, 2023) and automation (Blackmar, 2025) of copyright administration. With Google’s capture of the public sector, creative technologies protected by legal infrastructure are covertly repurposed as technologies of control.

Between 2014 and 2024, the term “hyperpop” evolved from online jargon to a subcultural-music social network, to a tech company’s metadata-tagging tool, to a curated playlist on a streaming platform. Only then did “hyperpop” become a mainstream genre. In the 2010s, the rise of low-bar-of-entry DAW software inspired musicians such as PC Music’s founders A.G. Cook and Danny L Harle, as well as Dylan Brady and laura les of 100 gecs, to experiment with software synthesizers and audio effects. Collaborating with PC Music, artists such as SOPHIE and Charli XCX gained an audience amongst queer and trans subcultures, where listeners associated their non-normative personal experiences with the expansive sound of digitally produced music. Picking up on this development in electronic music’s history, Spotify identified the two most commercially successful styles, authored by PC Music and 100 gecs, and created the “hyperpop” playlist to monetize this growing trend.

This paper examines how hyperpop, as a subculture, came to be constructed as a mainstream musical genre through Spotify’s metadata indexing and playlist curation. I analyze the metadata from former Spotify employee Glen McDonald’s open-source website everynoise.com, several iterations of the hyperpop playlist (from 2019 to the present day), and marketing language by Lizzy Szabo, Spotify’s curator for the hyperpop playlist. I then contextualize this data with recent scholarship that combines Science and Technology Studies (STS) and musicology by investigating the ways in which dominant streaming platforms use algorithmic representations, rather than actual music communities, to categorize artists and create genres (Johnson 2020; Hodgson 2021; Pelly 2025). Using community-based marketing language, futuristic imagery in playlist cover images, and a variety of guest-curated hyperpop playlists (including those by artists Yaeji, Bladee, Rebecca Black, Ericdoa, and Glaive), Spotify employees mask the vast influence of their algorithms—what are in fact doing much of the work to “discover” seemingly hyperpop-affiliated musicians. Critical study of the data categorization that undergirds curated playlists is important because, in the case of “hyperpop,” such a playlist ultimately fails to give credit to the queer and trans music subcultures in which the genre’s experimental vocal processing and software-based music production originated.

The undisputed heavyweight in the global marketplace of AI music products is the Boston-based, venture capital-backed startup called Suno. AI music generators like Suno have inspired new ways of creating music, listening to music, and new paradigms for thinking about intellectual property. Like previous music technologies such as the turntable or the guitar amplifier, Suno’s developers may have specific ideas about their product’s usage, but the users themselves have their own agendas, reconfiguring the technology’s usage. Following the STS work of Latour (2005), Pinch (2013), and Akrich (1992), I understand Suno users as actors within AI music’s broader sociotechnical network, contributing new ways of using AI to make music with existing features. Through discourse analysis, digital ethnography in online Suno user communities, and user interface studies, this paper theorizes new forms of musicking on the front lines of music technology, subverting prevailing assumptions about the destructive potential of AI in the music industry.

Suno’s healthy marketing budget affords researchers a window into how they understand the value of their product. They consistently market their ‘covers’ feature as a tool for aspirational songwriters who would like to record a rough demo and use AI to bypass the time and cost of producing the song in a studio. “Covers” uses machine listening, de-mixing, timbre transfer, voice modeling, and music generation processes to produce a professional-sounding recording from a rudimentary recording. Suno’s “covers” ads feature young photogenic artists who appear to be at the early stages of their careers. Yet many of the most active users of the “covers” feature are older hobbyists who have discovered that their dead recordings can gain new life and morph into novel genres using AI as form of musical upscaling I call “reanimation,” following Piekut and Stanyek’s notion of “deadness” (2010). Reanimators use AI music generation as a way of listening to their own work, collapsing composer, performer, and listener, into one composite musicking experience. Reanimation not only challenges notions of creation, but it also represents a form of automation that increases the value of older intellectual property, challenging prevailing narratives that have accompanied the AI boom.