In this paper, I discuss how voice models created with artificial intelligence (AI) can facilitate potential diffractions (Barad 2007) and radically challenge notions of identity based on individuality and originality, instead emphasizing heterogeneous formations of the voice. In recent years, there has been an exponential rise in unconsented so-called deepfakes of musical artists flooding the Internet. Recordings of voices of Drake, The Weeknd, Rihanna, Taylor Swift, and Frank Sinatra, to name only a few, have been used as training data on AI voice models, causing much debate around copyright law and intellectual property (IP). As a contrast, artists like Holly Herndon (with the voice models Holly+ and Spawn, developed by herself and Mat Dryhurst), Sevdaliza (with the voice model Dahlia, released via myvox.ai), and Grimes’s AI voice model (released via Elf.Tech), have embraced deepfake technologies, making available some of these models for public use for a fifty percent share of the royalties. Copyright and IP debates set to the side, this technology facilitates new material-discursive entanglements between voices, identities, individuals and collectives by letting a voice transform through one or multiple others in what can be described through the feminist concept of diffractions. Diffractions, in contrast to reflections or representations, maps the disruptive and transformative realisations of identities, processes, voices, etc. In studying the music made with these AI voice models, by the artists themselves or others, I ask: Are AI voice models replicating an archive of the voice, as reflections of another person’s vocal identity? Or can AI voice models facilitate diffractive potentials of the voice by creating new entanglements between the individual and the collective? How is agency in the creative process shifting between the actors involved in deepfake entanglements between people, voices, technology, algorithms, and data? What implications arise from such AI technologies directed at the voice?