A Violist Exploring AI: Reflections on Creation and Agency in The Uncanny Valley of Art

In this chapter, research assistant and classically trained viola player Jorge Enrique Lozano Diaz Granados reflects on the process of co-creating musical pieces in collaboration with the AI software ChatGPT. Through tinkering with this technology, Lozano Diaz Granados invites the reader to think about issues of creative agency, indeterminacy, classical music aesthetics, as well as the new skills - such as translation - that shape the creative role of the composer-performer. You are warmly invited to play the resulting score The Uncanny Valley of Art to feel and hear what a collaboration between musician and AI sounds like.

As a classically trained viola player who developed an interest in the sociocultural study of classical music and is a research assistant at the MCICM, neither composition nor technological literacy could be said to be my strengths. In fact, I only ever engaged in brief writing exercises in school projects within Music Theory courses during my conservatoire studies. That was until, as part of a university course about arts, literature, and technoscience, which I attended during my studies at Maastricht University, I encountered the first real opportunity to explore digital technologies, in particular AI, in combination with my musical practice. In the course, we were tasked with organising a festival around the course topics in which we could present our own works of art or literature. The special theme that year was “robots, cyborgs, and other creatures” (Hendriks, 2022, p. 42), and with my group we decided to create and present a multi-modal performance around the topic of AI-generated art. We combined visual arts, poetry, and music, and being the only musician in the group, it is not hard to guess what my task was in the project.

 

An AI-themed, AI-co-created Multimodal Performance

‍

On the lecture room’s wall, we projected a reel of visual artworks generated by an AI, while one of us recited an AI-generated poem and I performed music co-created with ChatGPT. The poem and the music pieces revolved around the implications of AI-generated art. For our performance we chose the title The Uncanny Valley of Art, drawing from the concept coined by roboticist Masahiro Mori in 1970 when describing the response to robots that look and act almost human. More humanlikeness elicits empathy in the viewer only up to a point, from which she descends into sensations of eeriness, strangeness, and unease, that is, into an uncanny valley (Mori, 1970/2012). However, uncanniness in the arts, and specifically in classical music composition, is often not necessarily to be avoided, but rather actively sought out by some artists (Laidlow, 2024).

‍

‍

The Poem

‍

From human hands, the art doth flow,  
A canvas, brush, and paint aglow,  
A symphony of colors, sounds and form,  
A reflection of the soul, a perfect storm.  

But what of art, born of machine,  
A product of code, a digital sheen,  
Can it truly be said to have a heart,  
Or is it just a clever facsimile of art?  
 
We marvel at its technical skill,  
At the way it can create and thrill,  
But can it ever truly capture the human soul, 
Or is it just a simulation, lacking in whole?

Yet in its creation, there is a spark,  
A flicker of something, shining in the dark,  
For even though its origin may be of a different kind,  
Its art is authentic, true and refined.  
 
So let us embrace the art of the age,  
Whether born of human or digital stage,  
For in its beauty, we find a common bond,  
A reflection of humanity, both new and old, beyond.

Each artwork in our performance used relatively common technical and aesthetic standards from each medium. The music performance, in particular, consisted of a series of six short pieces for solo viola played by a musician (me) standing in front of an audience. Nothing out of the ordinary there. The musical material itself was also nothing unprecedented: tonalism and atonalism without getting into serialism, contrasts between melodic and noise-based textures, structured arcs of tension and release, 20^th-century extended techniques on the viola… And yet, based on conversations with audience members and the video recording we made of the performance, something felt off about it, something uncanny.

‍

Filled with what we could call relatable material, that is, moments where it is possible to recognise a form of practice of classical music, each piece as a whole is somehow less familiar, the clarity of the little building blocks blurred when seen from a distance. Could this have been still perceived had we not told the audience in advance about the role that AI had played in the creation process? Or maybe it was the multimodality that shed a strange light on the individual media? Perhaps, in a different context—like a conventional recital in a concert hall—and without revealing their semi-artificial nature, these pieces might have passed for human-made, regardless of their perceived quality. All in all, it was precisely the disclaimer revealing their origin that in my view set off the descent into the valley in our performance.

‍

At the core of this uncanniness lies one of the dualities Robert Laidlow (2024) identifies in the intersection between classical music and AI— namely, between what is fake and what is real. That the music is real is seemingly straightforward: the pieces are made of organised sonic material that at first glance looks and sounds like music. And yet, despite this realness, such music can still feel unauthentic to its listeners. Authenticity remains then an open and subjective category, something to be felt rather than factually observed (Laidlow, 2024).

‍

As one of the central, centuries-old tenets of Western classical music, authenticity has been usually associated with unquestionable originality, a clear artistic vision or intention from the composer, and a degree of autonomy from external factors such as commodification. This appears inherently at odds with AI technologies and their tendency toward statistical generalisation and repetition. If authenticity in classical music is tied to an expressive individual voice, then it is debatable whether AI—at least in its current form—can ever truly produce music that meets this standard. This tension is not just theoretical; it materialises in performances like The Uncanny Valley of Art, where the music may be real, yet still feels somehow hollow or estranged.

‍

The question of authenticity in Uncanny Valley of Art is further complicated precisely because the pieces were not entirely produced by the AI. Had I used a model capable of score-writing, all compositional decisions beyond the initial prompt could be clearly attributed to the AI’s agency. However, only partially delegating agency and authority to the machine blurs the line between authentic and inauthentic because the feeling remains that something distinctly human lies somewhere in the music. Since by listening alone it is impossible to discern where the human is and where the machine, doubts on the piece’s authenticity begin to merge with—or even give rise to—the uncanny.

‍

‍

Co-writing Uncanny Music with an AI

‍

Even in its current version, ChatGPT is not capable of generating music notation, which is why the music for The Uncanny Valley of Art had to be co-written with me. Despite this obvious limitation, the chatbot is perfectly capable of articulating in words what can be later translated into a music score, albeit through a slow and clumsy process. To fully engage with the theme of our performance, I asked the chatbot to provide me with academic studies about the artistic or aesthetic value of AI-generated art. In 2022, the available free version of ChatGPT was known for making up what it could not provide an answer for, which was especially evident in the context of academic literature. Upon request, the bot could produce an entire bibliography about a topic, and yet most entries, if not all, would be made up. I was aware of this. Rather than a limitation, I took it as an intrinsic, defining element of my co-author, and embraced it as a part of the reflection that our performance sought to invite. 

It took a few prompts to get it to provide me with the questions and answers from the survey of the supposed study, but together, we ultimately defined six questions that discussed topics ranging from people’s awareness of when art is generated by an AI, their perception of its authenticity, its quality, potential drawbacks, and benefits (see Excerpt 1). 

I then proposed to write a short piece for solo viola sonifying each of the survey questions and their responses. Sonification was understood here as a process in which “data – which in and of themselves usually have nothing to do with sound – are translated into an audible signal, for the purposes of facilitating communication or interpretation" (Supper, 2019, p. 12). Based on each question and its answers, I prompted the bot to define the pitch, melody, texture, rhythm, and harmony for each piece. ChatGPT would at first provide vague answers, untranslatable without having to get involved in compositional decisions myself (see Excerpt 2).

Since I needed the model to create output as detailed as possible, in a next step, I asked for the specific materials for each of the elements, which it would then produce in words (see Excerpt 3).

In a last round, details were even more specified (see Excerpt 3).

Eventually, I realised that this process of refinement could go on indefinitely, and that I had to decide when to stop prompting and begin translating into scores. Rather than continuing until the pieces resembled fully-fledged, traditional sheet music (i.e., with as much detailed information as possible), the transcription process focused on making clear, legible scores with varying levels of detail depending on the qualities of each of the pieces and my needs as a performer. I chose to include only the minimum of material required to perform each piece at the festival. For instance, Piece No. 3 contains only pitch names in sequence (e.g., F–G–Ab–Bb…) and general instructions on how patterns should evolve. In contrast, Piece No. 6 has not only clearly defined duration and pitch for all notes, but also articulation, dynamics, and tempo, more closely resembling a typical printed score.

‍

‍

Which is the style? Or better, how is the style?

‍

Because the music was intended to accompany the poem and the images, and to have more control over the duration of the pieces and be able to end a piece when necessary or extend it to match the timings of the other media, I intentionally left open and indeterminate certain elements that could be decided during performance, such as specific number of repetitions of a motive, the rhythmic value or specific pitch of certain notes, etc. This approach, albeit chosen here for practical convenience, resonates with a 20^th-century compositional approach called indeterminate music, in which the composer relinquishes control by intentionally leaving certain elements of the music to chance or randomness (Griffiths & Whittall, 2011) and whose main representatives are composers like John Cage and Karlheinz Stockhausen.

‍

But are the Uncanny Valley pieces actually indeterminate music, given that randomness was introduced out of a practical need rather than an aesthetic intention? Moreover, do they belong in one of those categories we call “style”, like Modernism, Experimentalism, Minimalism, and so on? And if so, how can music co-created with AI be categorised into a style? One option could be to make it explicit through prompting, like my request for the Uncanny Valley once the study had been defined: “Now, pick an art-music style, school or composer that for you resonates with the study, its results and implications”. To which the model replied, “One composer that resonates with the study, its results, and implications is Brian Eno, who is known for his work in ambient music and has experimented with generative music techniques”. Eno is often considered stylistically to be a post-minimalist with strong ties to experimentalism (Reich, 2018).

‍

But is The Uncanny Valley of Art post-minimalism proper? Indeed, it seems to meet the criteria: all pieces use repetition in some form, many are tonal/modal and all use somewhat rich harmonic progressions, all have a steady pulse, often prioritise texture, and allow for performer interpretation and flexibility (Reich, 2018). Having little experience in composition, I don’t claim to have the authority to label the pieces as post-minimalism or otherwise. But perhaps a more interesting question than what style(s) do these pieces embody is how they might do so in a context of co-creation with AI. We have seen that, being asked explicitly, ChatGPT can propose a style and generate input within its conventions. Is it enough to examine the resulting piece in search for those conventions (use of repetition, steady pulse, complex harmonic progressions, etc.) to be able to call a style? And again, what are the implications of the form of intentionality that is at play, one where decisions are made not to follow an aesthetic desire or need, but to comply with a prompt?

‍

This brings to mind another of Laidlow’s (2024) duality in the relationship between classical music and AI, namely that between system and secret. On one side there is the view that music is constituted only by data—that is, purely musical material—, and on the other, that there is an elusive secret ingredient that is irreducible to sounds and silences (Laidlow, 2024). This duality applies to style as well. Is a piece post-minimalist if it reflects all of the style’s “data”, that is, if it ticks all the boxes in the checklist of conventions? Or is there a secret something that truly grounds it in one category or the other?

‍

Perhaps Laidlow’s own answer on the level of ontology can be also useful for the question of style: people continue to attend orchestra concerts in a post-COVID age of digital reproduction not because they want to experience the creation of sound, but “as an act of community shared between musicians and audience” (2024, p. 370). In other words, a music style is more than the sum of its sonic and purely musical parts. It exists also in the social world and cannot be excised from its contexts. All this suggests to me that perhaps the music in Uncanny Valley of Art is only partially stylistic in that it follows the musical conventions of things like Post-Minimalism, but is a sanitised abstraction, removed from the social and human contexts that once informed the style.

‍

Beyond the Valley

‍

These reflections, while resulting from a seemingly niche co-creative experiment, touch on broader discussions about the evolving relationship between classical music and artificial intelligence. In composing the music for The Uncanny Valley of Art, I was not merely writing pieces but also navigating a new and unfamiliar creative logic—where compositional decisions, style and meaning were all co-shaped by a system with no musical desire, no artistic intention, no social embeddedness, and yet, uncanny fluency in the language of music. This opens not only creative possibilities but also philosophical disruptions.

‍

My experience co-writing music with ChatGPT revealed that when composing with AI, questions of authorship, style, and authenticity shift from aesthetic judgements to deeper questions about how we construct and interpret meaning in music. In other words, instead of only whats and ifs (“who is the author?”, “which is the style?”, “is it authentic?”) we are compelled to consider the hows (“how is this the author?”, “how is it in this style?”, “how is it authentic?”).

‍

The Uncanny Valley of Art was not just a performance but an exploration of what makes music feel human, intentional, or sincere. What emerged was neither fully artificial nor fully human, neither stylistically grounded nor stylistically void, but rather something liminal that invites, or even forces us to reconsider not just the hows, but ultimately, whether musical meaning can be co-created through dialogue with a system that possesses no aesthetic desire of its own. At the same time, The Uncanny Valley of Art unintentionally echoed what Laidlow (2024) describes as the potential for compositional practice to speak back to the technology— by foregrounding its limitations, exposing its assumptions, and bending its generative logic into something reflective and critical, rather than treating it as an invisible, neutral tool. Perhaps, then, the uncanny in this project was not just a feeling—it was a method.