Abstract

This paper contributes to the growing scholarship on the cultural impact of AI from the perspective of popular music studies, considering the human-machine relationship in music production through two case studies of 'sampling-based' AI music by drawing on the concept of liveness. With the rise of the ‘signal’ processing in AI-assisted music composition, which concentrates on the generation of sound signals rather than symbolic representation (Tanaka 2023), musicians increasingly utilise AI-generated sound material in a way that popular music producers have drawn on sampling. Although scholars have already suggested the connection between sampling and AI music (Audry 2021; Navas 2022), no one has provided a detailed analysis of the practice based on the empirical perspective. This paper situates relevant works in the tradition of sampling-based music and demonstrates how the introduction of AI challenges the human-centric view in the conventional articulation of the liveness of such music. Taking a material-centric approach in analysing patten’s Mirage FM (2023) and Nao Tokui’s Emergent Rhythm (2022), this paper argues that the AI system creates new forms of entanglement between a human, a system, and sampling material in historically, socially, and technologically informed ways.

Author Keywords

popular music studies, sampling, liveness, latent space, real-time

1. Introduction: ‘Signal’ Processing and Sampling-Based AI Music

In 1957, Hiller and Isaacson composed Illiac Suite, arguably the earliest computer music, by having a computer generate sequences of numbers that represented musical parameters such as musical notes, rhythm patterns, and articulations [1]. The intensive research on AI composition followed this in and after the 1960s [2], and works such as David Cope’s EMI and George Lewis’ Voyager established AI-assisted compositional methods by 2000, at least in the academic circle[3][4]. However, the 2010s saw a radical shift in the major technique: the rise of the ‘signal’ processing over the ‘symbolic’ one [5]. The compositional methods of the works mentioned above fall into the ‘symbolic’ one, which deals with symbols. Yet, the mutuality of neural network speech synthesis, represented by DeepMind’s WaveNet (2016), pushed forward the emergence of the ‘signal’ one, which deals with audio signals. For example, in 2019, Dadabots (CJ Carr and Zack Zukowski) trained a neural network model on the ‘raw’ audio recordings of a death-metal band Archspire to generate the technological extension of the recordings, which was then streamed online [6]. While the earliest electronic simulation of human speech appeared as early as the 1920s, its recent development with machine learning attracts musicians, letting them generate novel sounds beyond the reproductive speech synthesis [7]. Moreover, the technique expanded the domain of AI music beyond serious ‘art’ music to popular music because of its affinity to sampling.

Historically speaking, the impact of sampling, or musique concrète, enabled by recording technology predates the way that the signal processing extended the conventional AI-assisted composition that was overwhelmingly symbolic. For centuries, composers in Western art music history have relied on symbolic representation, notation in particular, which directs performers how to render the work. But sampling foregoes the score. Utilising recording technology instead, the technique enables musicians to directly manipulate and organise recorded sound material. Given the affinity between sampling and signal processing, it is not surprising that contemporary artists integrate the sampling technique into AI-assisted music composition. One could argue that that is obvious because sampling is now ubiquitous, being a basic element of popular music production [8]. However, some artists take this approach more conspicuously, situating their AI-assisted composition in the tradition of genres strongly characterised by sampling, such as DJ performance and hip-hop, which I call ‘sampling-based’ music.1

Existing scholarship has already suggested the relevance of sampling to AI, or more specifically, machine learning, in music production. Centring on the idea of ‘selectivity’, Eduardo Navas considers the principle of sampling a crucial element of machine learning because generative models are trained to extract (or ‘sample’) appropriate material from datasets [9]. Furthermore, Sofian Audry argues that machine learning ushers in ‘a new era of remix culture’ because a machine learning model enables artists to manipulate ‘not only content but content-generation processes’ [10]. However, these authors have not investigated AI’s impact on the practice of sampling-based music. From the perspective of popular music studies, this paper focuses on how AI technology affects musicians’ experience in the production of music that samples AI-generated material. I consider this question significant because the AI’s heightened agency, compared to the conventional media, notably records, seems to challenge the existing assumption about the role of a human in relation to technology. To consider this, this paper investigates two relevant cases by focusing and drawing on the concept of liveness.

2. Literature Review and Methodology: Liveness of Sampling-Based Music

Liveness is conventionally defined as the negation of recorded music, forming a binary opposition. However, this definition cannot satisfactorily explain the liveness of sampling-based music, particularly, the DJ performance that centres on the re-play of existing records. An alternative articulation of liveness should account for the recording technology’s role, reflecting on the changing nature of the human-machine relationship [11][12].

In discussing the liveness of such music, musicological scholarship has focused on the musician’s artistry. Rietveld details various techniques that DJs employ in interacting with other participants including audience and stage dancers. Although she discusses technological mediation that constructs the liveness (i.e., records giving the audience prior knowledge about the performance), her emphasis is put more on artists’ skills in manipulating electronic devices [13]. Clancy also holds a similar perspective when discussing technological shifts in the DJ’s equipment including the introduction of AI. While admitting the impact of such technological shifts on the music industry, he adamantly focuses on the human agency when it comes to the authenticity of performance [14]. It is understandable that scholars focus on the human craft in the tradition that emphasises the manipulation of musical instruments, responding to the public suspicion about the automatic function of electronic devices [15]. However, if the underlying technology shifts, the concept of music can shift, and there is no need to frame emerging performances within the old assumption. Thus, this paper takes the non-human agency into account.

The shift of attention to the non-human agency corresponds to the generalising understanding of AI beyond merely a tool. Notably, Holly Herndon emphasises that AI music composers including Cope and Lewis have admitted the subjectivity of the AI system and calls her system a collaborator in the production of Proto (2019) [16]. The collaborative view is also upheld by Sougwen Chung. Other artists prefer to characterize AI’s role in more instrumental terms while admitting its increasing degree of agency. For example, Rebecca Fiebrink calls her system ‘metainstrument’, meaning ‘an instrument to create an instrument [10]. Admitting the agency of AI does not necessarily mean either the anthropomorphisation of a system or the abandonment of the human agency. Rather, it aims to re-examine the human-machine relationship, instead of superimposing their unidirectional relationship—be it that a human has full control over a system or that a machine replaces a human creator.

In the context of sampling-based music, this requires what I call a ‘material-centric’ approach. Influenced by, but not complying with, the non-human turn in humanities, this approach focuses on sampling material and considers how both a human performer and a system participate in the manipulation of the material [17]. By shifting the attention away from merely human labour, it aims to explain the interactional relation between the performer and the system.

Taking this approach, the following case studies investigate patten’s album Mirage FM (2023) and Nao Tokui’s performance Emergent Rhythm (2022). Focusing on their core concepts, ‘crate digging in latent space’ and the ‘real-time’ material generation, the analyses demonstrate how AI subverts the established assumption about the relationship between a human performer and a machine. Each section has the same structure: the general introduction and technological description, followed by an in-depth analysis of the interaction between the performer and the system. They draw on published and personal interviews with the artists as well as their descriptions of the work.

3. Case Study 1: patten’s Mirage FM (2023)

3-1. ‘Crate Digging in Latent Space’: Exploring the Virtual Space

London-based artist Damien Roach, aka patten, released the album Mirage FM in April 2023. The artist’s official website claimed it to be 'the 1st album fully made from text-to-audio AI samples' [18]. The ‘samples’ were generated by Riffusion, a web-browser-based AI application that generated short audio clips, including voice, bassline, synthesizer’s harmony, melody, etc., from text prompts (Image 1). While the creator admits that the generated samples underwent considerable post-production processes (e.g., ‘chopping, sequencing, layering, EQ, repitching, and adding effects’), the listener of the album would rather feel various distinct textures of the materials [19]. That impression is intensified by the album’s organisation. It consists of 21 tracks given a short, abstract title, such as ‘Inhale’, ‘Say’, and ‘Forever’. They last only a few minutes (49 seconds at the shortest and three minutes and ten seconds at the longest), and none of them exhibits any dramatic development. As a music writer observes, as if modelled on radio or streaming playlists, the segments ‘flit and flicker’, unceasingly floating between and beyond various popular music genres [20]. That also sounds as if the album rehearses the material generation process by Riffusion as such, which generates unpredictable, miscellaneous outputs. In addition, the overall sonic quality adds another layer of noticeable material texture to the album. That timbral texture will be explained as ‘lo-fi’, a poor sound quality often associated with lo-fi hip-hop, a musical genre that incorporates the distorted and muffled sound like an old vinyl, radio, or low-quality video into its aesthetics [21][22]. The timbre is a characteristic of Riffusion’s output. The rugged sonic amalgamation covered in the lo-fi sonority adds up to the ‘uncanny’ impression that the above author had [20].

While the album thus foregrounds the materiality of the AI’s outputs, patten emphasises his role of collating the materials in the same manner as scholars found the liveness of DJ performance in the craft of a human performer. To promote the idea, he creates a manifesto, ‘crate digging in latent space’ to employ in several places [19]. As a neologism, it situates the conventional labour of DJs (‘crate digging’) within the new technological environment (‘latent space’). Whereas ‘crate digging’ means the act of discovering material to be sampled, unclear is how it applies to the technological notion of ‘latent space’. To understand this, and his labour in the production in general, I briefly describe the function and mechanism of Riffusion. The description below regards those at the time of the production, as the platform transformed in October 2023 in commercialising the initial ‘hobby project’ [23].

3-2. Riffusion and Latent Space

The application Riffusion is based on an open-source library of the same name released by Seth Forsgren and Hayk Martiros in December 2022.2 Although the application generates audio output, as they explain, the library itself is a ‘text-to-image generation model’ [24]. The application utilises the model to process audio internally as images. It functions so because the model is developed upon another open-source text-to-image model, Stable Diffusion (v1.5) released by Stability AI in August 2022.

Stable Diffusion consists of a text-encoder and an image generator. The text-encoder translates the text inputs into numbers (tokens) representing the words, and the image generator generates an output image by processing the numbers. The output image represented as numbers is generated as follows. Firstly, random noise is added to the array of input numbers. Then, another part of the system predicts what noise should be subtracted from there to restore the original data. That prediction is based on prior training. When the predicted noise is subtracted, there appears a new array of numbers close to the original but slightly different due to the gap between the actual noise and the predicted one. A text input, translated as numerical parameters, affects each stage of the generation process so that the output image reflects the input prompt [25].

The creators of Riffusion finetuned their model based on Stable Diffusion to process audio data internally as spectrograms. A spectrogram is a visual representation of sound. As shown in the image below (Image 2), the spectrogram documents the change of frequencies over the time with amplitude of each frequency: frequency and time are respectively represented by the y-axis and x-axis, and the amplitudes are expressed by the intensity of colour [26]. Importantly, a spectrogram, an image converted from sound, can be re-converted into sound in a reversed process.3

Riffusion, as an application, is capable of instantly generating a musical clip based on any combination of keywords, even those unfamiliar to the training dataset. However, its strength is neither instantaneity nor flexibility, but infinite variation. Because it generates audio from a small spectrogram image (512 x 512 pixels), it only emits very short clips (five seconds). But the clip is constantly looped; and with or without the user input, the looped clips keep changing, leaving an infinite sequence of variations. The creators explain: ‘As the user types in new prompts, the audio smoothly transitions to the new prompt. If there is no new prompt, the app will interpolate between different seeds [original data to generate new outputs] of the same prompt’ [26].

It is in this function that the application takes advantage of ‘latent space’. As an engineer briefly explains, latent space is 'a representation of compressed data' [27]. In machine learning, for the sake of efficiency, data are compressed into a set of numbers, and the numbers are mapped into a virtual space. That space is called the latent space. The ‘space’ is not necessarily three-dimensional: since data are analysed from numerous perspectives, it is usually an ultra-dimensional space that a human cannot perceive. Using latent space can be an 'advantage' for creative use because it allows the system to generate something between two distinct things. A TechCrunch writer suggests: ‘Like if you had an area of the model representing cats, and another representing dogs, what’s “between” them is latent space that … would be some kind of dogcat, or catdog’ [28]. For Riffusion, the ‘dogcat’ and ‘catdog’ are the virtually infinite number of spectrogram images that exist between seed images. Furthermore, the degree of variety increases, as Riffusion conjoins other algorithms to realise smoother transitions from one state to another [26].

3-3. Interpreting ‘Crate Digging’: the Infinitude-Finitude Duality

Given the description of latent space, patten’s ‘Crate Digging’ literally means his exploration of artistic possibilities in the virtual-logical space constructed by the combination of various distinct features. However, since he explores the space through the interface of Riffusion, it is more adequate to take the manifesto as his experimental engagement with the tool for learning correlational patterns between human input and the system's output. He explains:

When I was experimenting with [Riffusion], one thing I was interested to find out was what it knew about ... So I was kind of exploring. ‘Oh, does it know about this kind of music, that kind of music?’ … then after maybe a couple of hours, I was like, ‘Wait, there’s interesting potential here to use this stuff as samples, and to stitch them together to make music of some kind’. [19]

The experience is not limited solely to the interaction with Riffusion. Prior to the application, he also tinkered with a text-to-image AI tool, DALL-E [20]. These experiences gave him various ways of prompting in making Mirage FM, including the 'combination of sometimes very, very specific directions and then sometimes really vague' [19]. In summary patten’s ‘crate digging’ involves not only the search for sampling material in latent space but also the learning of the dispositions of the AI tools.

His coining of the manifesto reflects the contemporary scepticism towards AI. Around the end of 2022, the surge of hands-on generative AI tools lowered the technical barrier for the public in equipping AI for creative activity. Behind the fad, concerns around its use grow: aesthetical questions about its claimed creativity due to the derivativeness of AI; ethical and legal problems about the authorship of training dataset; ethical problems about its use, notably deepfakes. Riffusion is not free from the allegations. To generate image outputs aligning with text inputs, text-to-image generation models must be trained on a dataset comprised of images and their corresponding text descriptions. It is known that Stable Diffusion is trained on a public dataset, LAION-5B [29]. Problematically, however, the creators of Riffusion have not officially disclosed any information about their training dataset: opacity exists as to its audio and text data. It is fair to mention that the creator of Riffusion did not intend its commercial use; however, after the commercial release of patten’s album, Riffusion became entangled in the controversy.

Although patten is aware of these concerns, he circumvents them, instead attempting to lay the aesthetic foundation by emphasising the user’s skill of manipulating the software. He emphasises that users can express their artistry through work despite the tool’s derivative nature: ‘listening to it, it made it clear that anybody accessing these tools and deciding to use them could very much retain their identity, and their own ideas, and impulses, and aesthetics’. While admitting the tool’s agency, he subjects it to a human creator by maintaining, 'these systems kind of await activation by an active human mind' [20]. In these terms, patten’s coinage of the manifesto is strategic. Reflecting on the problematic opacity of AI, he deliberately combines the conventional practice and the technical terminology to publicly elucidate his productive role: 'I wanted to be really transparent about what was going on. ... That's why also I came up with that term “crate digging in latent space.” Just so that people could have a way in to [sic] understanding what was going on there' [20].

When regarded from a historical perspective, however, his emphasis on the creator’s role of organising material is interpreted not only as his personal volition but also within a larger technological shift. As Matthew Yee-King argues, since the introduction of the computer, artists have been faced with the 'over-abundance problem’. While taking advantage of the computer’s calculational power in generating outputs based on human input, avant-garde composers observed ‘a transition from a scarcity of ideas to an abundance’. With this ‘transition’, composers became involved more with selection [30]. Similarly, patten’s role as a ‘crate digger’ is situated in this context. A machine learning system contributes to the abundance in two ways: firstly, the sheer volume of data in the training dataset; and secondly, the ability of latent space to generate infinite variations. Given this, patten’s selection of material concerns not only pre-existing works of others but all the possible variations. While the creator cannot utilise these possible materials until they are actualised as system output, latent space exhibits the virtual infinity of the expressional palette.

In this framework, a system generates countless options, whereas a human creator has to select some from the options. If human selection is understood as the translation of infinitude into finitude, the machine and human respectively represent infinitude and finitude. However, this relationship can be subverted: the infinitude of a human is translated into the finitude by the function of the machine. While I explained the virtual infinity of a machine learning system above, there is another layer on the human side that engenders the 'over-abundance’: user input that can take any combination of words. However, this abundance is to be limited in the interaction with Riffusion because of its technical mechanism that interprets user input. While the system accepts any words, they are translated into a sequence of system-readable numbers (tokens) by the text-encoder. In that process, the encoder refers to a list of words called a tokenizer. The list contains ordered words with corresponding numbers. When the user input includes words that are not already listed, the encoder statistically approximates the most likely words [31]. By this process, the system determines what words are used and how they are organized in the output generation, making the infinity of human input finite. Then, it is both a human and a machine that not only possess infinity but also put limitations. The ‘crate digging in latent space’ exemplifies this duality.

4. Case Study 2: Nao Tokui’s Emergent Rhythm (2022)

4-1. The AI-Jocky: Manipulating AI Models Like Records

Tokui Nao’s performance titled Emergent Rhythm took place at MUTEK Japan, an annual digital art festival in December 2022. Proposing the concept of ‘AI jockey’ (AI-J), ‘who tames and rides an AI’, he combines the DJ performance and the new technology, employing several types of AI models [32]. Mainly two kinds of models play central roles. The first kind is StyleGAN2, a model based on GANs (Generative Adversarial Networks) [33]. While it is a model generating images, he employs it to generate sound material, just as Riffusion does, by internally processing spectrogram images. The performer used seven GANs models trained on different datasets, such as genre-specific rhythms, instruments, and environmental sounds. As a demonstration clip on YouTube shows, the model is controlled by a two-dimensional input: the output sound changes as the mouse pointer (a grey circle) moves in the space on the above-left (Image 3) [34]. Noticeably, the usability of its interface differs from Riffusion: his interface looks much like an instrument or a drum machine, unlike patten’s metaphor of ‘crate digging’. That difference is given by the limited possibility of input compared to Riffusion’s limitless text prompts. But the way the performer deals with the models nevertheless appears as if a DJ manipulates records, rather than an instrumental player, as the models’ outputs are reassembled and added some effects by a mixer [32].

The audio is altered by the other group of AI models, Rave (Realtime Audio Variational autoEncoder), originally developed by Antoine Caillon and Philippe Esling in 2021 [35]. As a timbral transfer, the model can change the timbre of input to another one. Tokui’s performance draws on the models trained on diverse cultural resources including ‘Buddhist chantings [sic], Christian church choirs, Bulgarian female choruses …, an African folk instrument, the kora, and so on’ [32]. Models are sometimes layered to cause a ‘feedback loop’ (whose effect he acclaims as ‘Heavenly’) [36]. Since individual models and their outputs have a distinctive character, they are, as sampling material, appropriated, combined, and altered.

4-2. Comparison of AI-J and DJ: its Implication to Liveness

As the naming suggests, the role of an AI-J performer resembles that of a DJ—both manipulate their material. However, there is a fundamental difference between them in terms of the performer’s relation to the material because of the nature of the material. For a DJ, the material is a record, a static medium which documents and re-plays the original performance, ideally as it was recorded. In contrast, the material for AI-J, an AI model is dynamic as it constantly generates variations of the original.

This nature of the AI-J’s material disturbs the assumption of liveness about the DJ performance. The conventional liveness has been conceptualised as the negation of recording. When considering the liveness of the DJ's performance under this understanding, a human performer is more advantageous, as a contributor to the liveness, than the material. As mentioned earlier, scholars regard that a human performer can nurture its liveness, although records produce its sound. In this view, only the human performer can activate the fixed media for making the sound ‘live’. This view implicitly assumes the contrasting natures of a human performer and recordings, respectively being active and static. However, AI models are dynamic, as maintained above. The only static about the AI-J’s material is the training dataset that the models learn. In the AI-J performance, a human performer modifies the data by manipulating the models. Yet, it is important to remember that the models also participate in modifying the data by its function. I stress the latter dimension of production where the model modifies because the concept of AI-J makes this dimension less apparent. The reason is as follows: the concept of AI-J treats AI models like records for a DJ; by this, it black-boxes the model’s inner workings; furthermore, the black-boxing brackets the fact that they generate variations, as taking the model’s output simply as ‘output’ rather than derivations from the original (a training dataset). In this respect, a human performer cannot assume superiority to the models in dynamizing the static. The adaptation of a machine learning system as sampling material subverts the hierarchy between a human and material established as to the DJ performance.

In the following section, I upset this hierarchy further by associating a DJ and an AI-J, focusing on the material distribution as their background. By explaining how the technological development of the machine learning algorithms enabled the live performance, in which the models generate sounds in ‘real-time’, I demonstrate that an AI model as material not only attains a subjective position comparable to a human player in organising the sound but also affects the mode of interaction between a human and a machine.

4-3. The Implications of Being ‘Real-Time’

Perhaps, it is easy to draw a parallel between the distribution of records, as a backdrop against the DJ culture, and the open-source culture since the 1980s, as that against Tokui’s performance. In introducing the idea of ‘Deep Remix’, Audry associates the proliferate use of machine learning by artists with the online distribution of various models as open and free source codes [10]. Likewise, Tokui draws on third-party models, GANs and RAVE for the core part. Although he modified some parts, it should have been harder to build the entire system without them. Like DJs, the AI-J appropriate the AI models of others, while he modifies and trains them on his datasets.

Tokui’s reliance on the public models also informs the quality of his performance as live in terms of the temporal aspect. As he admits, the performance became possible because of new models for audio generation that appeared around the beginning of 2022. In general, audio generation is a technically demanding task in machine learning, compared to symbolic processing, due to the necessary amount of data. Therefore, it was not feasible to implement this kind of machine learning in the live performance until such models appeared because Tokui wished to generate sounds in ‘real-time’ [32].

The term ‘real-time’ reflects Tokui’s ideal vision regarding the roles of a machine and a human in live performance. Here, to consider the implication of this term, I draw on Simon Emmerson’s discussion on terms, ‘real-time’ and ‘live’. According to him, these terms are sometimes confused with each other, but the term ‘real-time’ gradually became preferred to ‘live’ in electronic music production around 1980. He considers that the terminological shift occurs when the ‘real-time’ operation of a system becomes more integral to the unique quality of performance than other components including the human manipulation of sound [37]. Then, how does this apply to Tokui’s performance? Tokui defines being ‘real-time’ as follows: that status is accomplished when ‘sound is generated in the same or shorter time as the actual performance’ (i.e., generating a three-second clip within three seconds) [32]. As this suggests, the definition is technical, as Emmerson interprets this term. Given this, his performance is oriented to the system’s function. This orientation differs from the existing discussion about the liveness of DJ performance: as explained earlier, scholars attribute its liveness mainly to the craft of a human performer. Judging from the implication of the term ‘real-time’, the AI-J performance focuses more on the system than the performer, although Tokui derives its inspiration from the DJ performance.

However, the ‘real-time’ system operation also facilitates the human performance. The orientation to real-time interaction with the system drives his designing of other AI applications, too. According to him, the advantage of the real-time interactivity is that it allows a user to easily experiment with the system. If a user likes a system output, they can keep with the same input; if not, they can instantly try out another input. This interactivity is crucial for him because he believes in the creativity of unexpected outputs that are engendered in ‘tries-and-errors’ [38]. This assertion resounds with patten’s emphasis on the learning process. Given this, Tokui’s focus on the real-time sound generation does not exclude a human role, despite its orientation to the system. Rather, the condition vitalises the interaction between a human and a system. Indeed, he believes that the real-time interaction has been the feature of technological devices that have been creatively ‘misused’ by musicians, notably the turntable for DJs. In this respect, his preference for the system’s real-time operation reflects the tradition of the DJ performance, although AI technology possesses a heightened agency in manipulating material.

5. Concluding Remarks

Concerned with the cultural impact of AI from the perspective of popular music studies, this paper situated AI-assisted works involving the ‘signal’ processing, rather than the ‘symbolic’ one, in the tradition of ‘sampling-based’ music to consider how the introduction of AI challenges the human-centric notion of the human-machine relationship. Against the conventional discussion on the liveness of such music, the case studies in this paper adopted the ‘material-centric’ approach that accounted for the roles of both a human performer and an AI system in manipulating sampling material.

The first case study on patten’s Mirage FM focused on his manifesto ‘crate digging in latent space’. The analysis demonstrated that the term, closely related to his learning process, reflected both his deliberate emphasis on the human craft and the technological context which posed the 'over-abundance problem’ to composers. In addition, the investigation into the technological function of Riffusion showed that the system and the composer both had virtual infinity while simultaneously limiting each other in concretising the material. The second case study on Nao Tokui’s Emergent Rhythm focused on the quality of the ‘real-time’ material generation. The term, often confused with ‘live’, has implied the system-centric view in electronic music composition, as opposed to the human-centrality in the conventional articulation of liveness in the DJ performance. Tokui defined the term in a technical term, but the analysis demonstrated that he regarded the quality of being ‘real-time’ as necessary for facilitating the creative interaction between a human performer and a system.

To conclude, from these observations, I argue that AI technology, with its extended agency, not foregoes human performers, but creates new forms of entanglement between a human, a system, and sampling material in the production of sampling-based AI music. The shaping of these patterns is not determined solely by the artist’s will but informed historically, socially, and technologically. To consider the implication of this with regards specifically to AI, perhaps, the opacity of AI systems plays a key role. Although researchers consider the opacity of AI socially problematic [39], my case studies indicated that the opacity partly shaped the mode of creative interaction with an AI system: for patten, the opacity instigated his learning, and for Tokui, it was the source of creative unpredictability.4 Then, the intersection between social responsibility and artistic design can be the focus of future research, and an interdisciplinary approach will be required in this respect.

Ethics Statement

There are no potential conflicts of interest (financial or non-financial) involved. This study is part of my PhD research approved by the ethics committee of Goldsmiths, University of London. It does not have any potential societal, social, or environmental impact that is detrimental. While this paper draws on data attained in a personal interview with an artist, the interviewee consents to the use of the data for the research.