Conference Agenda

This study explores the intersection of artificial intelligence and music through AI-driven singing technologies, examining their cultural, ethical, and ontological implications. Drawing on posthuman theory (Hayles, 1999; Braidotti, 2013), it interrogates how AI challenges conventional notions of voice, agency, and embodiment by extending vocal presence beyond human performers. The emergence of AI-generated voices —such as the digital “revivals” of Teresa Teng and Michael Jackson—raises critical questions about artistic identity, memory, and the politics of musical reproduction in the digital age.

Through virtual ethnography and participant observation on the Chinese platform Bilibili, this study investigates “AI singing” communities centered around Sovits, a technology that enables AI-driven voice synthesis. It analyzes fan interactions, the co-creation of AI-generated performances, and the evolving role of machine learning in music production. Grounded in Latour’s actor-network theory (2005), AI-singing is conceptualized as a “hybrid assemblage” in which humans, algorithms, and digital voices interact—destabilizing the binary between human and non-human actors. This process reconfigures traditional frameworks of authorship, authenticity, and musical preservation, positioning AI as both an agent of technological reproduction and a participant in creative collaboration.

By examining AI-generated voices as socio-cultural artifacts, this research argues that they engender a new form of musicking, one that transcends temporal and spatial limitations while redefining artistic agency in digital spaces. In doing so, it expands understandings of AI’s role in contemporary music practices, illuminating the entanglement of human and technological agency in shaping today’s digital soundscape

Igor Stravinsky argued that performers should leave their agency behind, allowing a piece of historical music’s meaning to emerge unsullied by “subjective” performance styles. Beginning in the 1980s Richard Taruskin argued that Stravinsky’s ideal of the performer as self-distancing “non-agent” slipped into discourses about how to perform early music. The result, Taruskin (1995) claimed, was an affectless style of performance in which performers seemed to avoid “expression.” Taruskin, writing from a North American liberal perspective, sensed undemocratic, authoritarian politics were in play: the primacy of the text over the human stands for the power of totalitarian leaders operating under technocratic regimes. Those old enough to remember will know how high the stakes of this debate seemed to be.

The arrival of musical machines with agency, whose products can seem flat and without affect, reminds us that for many, performers (and now machines) are merely infrastructure in a media delivery system that brings music to audiences. This paper asks: are the technocrats back, in guise of computer scientists who assemble musical Large Language Models? Is being responsible for a performance like being responsible for driving a car? What of the rights of infrastructure to interpret? Does the (potential) hegemony of generative AI packaged and sold by platform companies conceal anti-democratic impulses? My argument seeks to bring organology (Dolan 2013), media theory (Rehding 2017), “critical musical infrastructure studies” (Devine 2019 and Devine and Boudreault-Fournier 2021), and Bruno Latour’s ideas about non-human actors (Latour 2005, in music studies see Piekut 2014) into productive dialogue with current thought on responsible AI, for example the “Edinburgh Declaration on Responsibility for Responsible AI” (2023). The declaration asks those who develop and use AI to consider four imperatives: accepting responsibility, seeing the ascription of responsibility as relational, moving away from the idea of responsibility as blame towards concern for the vulnerabilities of those impacted by AI, and imagining a responsible future for AI that prioritises sustainability over disputes about pace. To think carefully about performance is to think similarly. Maybe we have more to offer to debates about AI than we think.

After abolishing 38-year martial law in 1987, Taiwan entered a new period in which freedom of speech, publishing, and assembly were no longer restricted. Ever since, political participation has become a part of daily life: people talk about politics, go on the streets to fight for their rights, and exercise their civil rights by directly voting for their presidents, mayors, and legislators. In this flourishing era of civic participation, music plays a crucial role in mobilization and engagement with its feature of bonding people emotionally, expressing their identity, and promoting specific ideology.

With my internet-based ethnography and personal experience, this paper will focus on the role of music and social media in mobilizing the Blue Bird Movement—a significant civic protest in Taiwan that emerged in response to a controversial legislative reform proposal in March 2024. The study explores how participants used cyberspace, particularly social media platforms like Threads, to organize and amplify the movement. The focus is on the use of AI-generated music and the creative integration of digital symbols, such as hashtags and fan culture, to engage the public. These tools allowed for efficient mobilization both online and offline. Additionally, the paper discusses the controversial political statements made by celebrities, including Mayday and Jolin Tsai, during their China tour, which sparked debates within the movement. This research highlights the intersection of digital and physical spaces in modern social movements, illustrating how AI-generated music and online collaboration shape contemporary protest strategies and political discourse in Taiwan.

The Spice Girls are placed across a chess board in the order they're mentioned in the "Wannabe" rap: who's in the place of the white queen? An odd question for AMS, perhaps. It is, however, part of a “musicology” problem that is being used to test Large Language Models (LLMs) via Humanity’s Last Exam (HLE).

Styled as “the final closed-ended academic benchmark of its kind,” HLE is a “multimodal benchmark at the frontier of human knowledge” which tests the gumption and grit of LLMs by posing “2,700 challenging questions across over a hundred subjects” (Hendrycks et al. 2025). As a benchmark, the premise is simple: the better the AI does on the Exam, the better the AI. While most LLMs presently do poorly, the team behind the Exam—the Center for AI Safety and the startup Scale—hope that performance will inform “discussions about development trajectories.” In other words, the intent is for models to be modified according to what the Exam demands.

This may be cause for concern for musicologists, who will likely see little modern-day musicology in HLE. Few questions approach the "Wannabe" problem for strangeness, but many emphasize the rote appellation of chordal or scalic qualities—there is always a right answer—and consequently betray an ideological aspiration towards an objective ear. HLE thus “prioritizes the capture and certainty of information” (Robinson 2020) and so seems epistemologically misaligned with much post-New Musicological literature. Such a misalignment exemplifies the “technoscientific” lean of music AI and the resultant “asymmetry” that scholars increasingly seek to remedy (Gioti, Einbond, and Born 2023).

As Drott (2020) observes, scholars tend to emphasize AI outputs and overlook their inputs. I build on this intervention by considering how benchmarks steer AI development and the courses they chart. Specifically, I explore what kind of musicology is constructed by the HLE dataset and how an LLM weighted to best the Exam would “understand” the field’s priorities. Situating HLE’s role in shaping LLMs alongside the proliferation of AIs in educational contexts, I argue that we should be cultivating closer correspondences between Artificially-Intelligent and institutional understandings of what it means to be “musicological.”