Skip to main content
SearchLoginLogin or Signup

The Phenomenology of Deconstructivist Aesthetics in Music: An Autoethnography of Errors, Erasures, Permutations, Discontinuities, Paradoxes and Artificial Intelligences

(long paper published at AIMC 2023)

Published onAug 29, 2023
The Phenomenology of Deconstructivist Aesthetics in Music: An Autoethnography of Errors, Erasures, Permutations, Discontinuities, Paradoxes and Artificial Intelligences


It may be surprising to some that sometimes a technique, more than a creative impetus, can yield the aesthetic of an artwork or a musical composition. We can be reminded of the controversy in early modern music between the structural and phenomenological aspects of music creation. We argue for a phenomenological analysis of a particular use of Artificial Intelligence (AI) and General Adversarial Networks (GAN) (Goodfellow et al. 2020) in musical composition that we have described in recent publications: more specifically, a double process of deconstruction (i.e. feature extraction, learning transformation matrices) and reconstruction (i.e. generating transformation matrices using deep learning and musical scores). In Heidegger’s premonitory phenomenological destruktion of technique (tekhnè) (Heidegger 1958), technique becomes a logic of its own that may not easily be accessed by human reason (Husserl had previously spoken of deconstruction in the context of normativity (Husserl 2020)). This announces the age of machine learning in the arts (recently advocated by critics such as Sofian Audry (Audry 2021)) where deep neural networks are often created as black boxes, beyond the usual grasp of human intelligence. We believe that our particular use of AI in music composition (i.e. the process of deconstruction and reconstruction) yields a deconstructivist aesthetic (in the sense of the deconstructivist aesthetic that can be seen in architecture since the 1980’s). We provide a description of the sensations (in the sense of Deleuze’s sensation block description of music) that emerge from the experience of our AI music and we attempt to give a context to our deconstructivist aesthetic of music through the philosophies of Martin Heidegger or Gilles Deleuze for example, the works of writers such as William Burroughs or Ronald Johnson, architects such as Daniel Libeskind or Peter Eisenman and composers such as John Cage or Helmuth Lachenmann.


This work stems from a research-creation project where the objective was to write an opera based on AI techniques. The opera is based on a remodernisation of The Cherry Orchard by Anton Chekhov that is now set in April 1975 Vietnam (the fall of Saigon) rather than in the Russia of the end of the 19th century. It parallels the expropriation of a South Vietnamese family by the Communists of the North (a routine story in the Vietnamese diaspora) to the expropriation of a landowning family in Russia at the end of the 19th century. Verbatim techniques are used to parallel both works. In a talk given at the Impossible Projects 2022 Symposium, the problematic of addressing the Vietnam War with authenticity was discussed by the authors and the question of whether or not authenticity was a result of technique (the virtuosity of the creator) was addressed. We know that the critique by Heideggerian phenomenology of the dyad artist/artwork (whose origin, the philosopher thought, was in a phenomenological paradox, or loop) irremediably leads to the problem of authenticity and truth in art (Heidegger 1962). However, the idea that Being can be determined by technique (tekhnè), that an aesthetic is determined by its material constitution, may seem in itself blasphematory to some and contradicts the more conventional notion of an artist’s genius. For Heidegger, the tool (i.e. the hammer) becomes one with the hand that wields it (Heidegger 1986). We can also mention Gadamer’s well-known position that artistic truth is inaccessible to the scientific method (Gadamer 1996). This is effectively the practical reality in which many artists live today.

Composers of the critical composition movement (e.g. Nicolaus A. Huber and Helmuth Lachenmann) further find authenticity by politicising the orchestra and introducing extended techniques that reverse the conventional epistemological and political order of performance and the orchestra as an institution. However, it can be said that these so-called extended techniques, with time, have become a standard of their own, yielding particular and expected aesthetics and ecosystems. As an example, Lachenmann wrote his piece Air (1968-1969) for percussionist and large orchestra as a metaphor of the struggle between the individual and the collective. The piece was written during the Mai 68 protests; the composer is also said to have thought of the implications of the theories of the terrorist group Red Army Faction (Bosseur 2019).

The idea that technique can generate aesthetics, if it may seem banal today, is not evident in the aesthetic theories of the 19th century or even of the 20th century (where, for example, one could rely on the idea of creative genius, itself a thinly veiled derivative of the Deus Creator Omnium). It will be remembered that in the Western historiography of the concept of aesthetics, the emergence of the creative artist and the independence of aesthetics were impossible before the Renaissance, as long as God alone had the power of creation (Jimenez 1997). Greek tragedy and the ability to carry the human voice through megaphones (albeit rudimentary ones) and the architecture of the Ancient theatre, stone cathedrals and Gregorian chant, the use of wood in Protestant churches and Bach's polyphony, the Shakespearean theatre, the 19th century orchestra and post-romantic music, electroacoustics and High Modernism, are all examples where technique determines aesthetics.

Also, in (Nguyen and Tsabary 2023a, 2023b) it was hypothesised that some particular use cases of AI and deep learning techniques (e.g. latent spaces or deep neural networks such as GANs, VAEs, CNNs and RNNs) lended themselves to the emergence of a deconstructivist aesthetic in music. Parallels between artefacts created through the use of AI techniques and deconstructivist architecture (a term popularized by the MOMA catalog of (Johnson and Wigley 1988)) were shown. More specifically, the composition process was assimilated to a deconstruction phase (where a corpus was deconstructed into features) and a reconstruction phase (where a work is reconstructed based on the features learned in the deconstruction phase). We are here in a situation where a layering of techniques generates an aesthetic. Actually, materials and sketches are generated by AI and assembled by the composer. The sketches generated were highly disjuncted and this was due to the combinatory aesthetic of the process. In (Nguyen and Tsabary 2023a, 2023b) the notion of reconstruction paradoxes was introduced: that is, we can notice that, when using such AI techniques, reconstructing the parts of a whole no longer yields the same properties as the whole. Rhythm, tonality and harmony are broken as discontinuities are introduced in the enchaînement of the corpus (its well-ordering). These paradoxes remind the Banach-Tarksi paradox in topology where a sphere is deconstructed into pieces and with the pieces, 2 spheres can be reconstructed rather than the one we started with.

This paper discusses diverse deconstructivist aesthetics in the history of music, poetry and architecture, that is, aesthetics that are disjunctive, combinatory, erasive, discontinuous and paradoxical. We speak of documentary poetry, the cut-ups and experiments of Brion Gysin and William Burroughs and the erasure poetry of Ronald Johnson. We parallel these with deconstructivist architecture and the music of John Cage, post-serialism and New Complexity composers. The paper ends with a case study of the use of AI in a deconstructive manner in an important scene in The Cherry Orchard remodernisation opera (a research-creation project we have undertaken since 2022). More specifically, we show how GAN’s were used on transition graphs computed over some corpus of existing musical works to generate New Music. The objective of the paper is not only to describe a technique (AI) for music composition but also to give a feel of the aesthetic that emerge from using the technique, phenomenologically bracketing it in a collection of precedents and approximations.

The Aesthetics of Deconstruction

Territory Denial and Information Overload

Figure 1: Examples of territory denial: (left) Facade of the Jewish Museum Berlin by Daniel Libeskind (Image Courtesy @ Daniel Libeskind; Digital Image © The Museum of Modern Art/Licensed by SCALA/Art Resource, NT); (center) (left) inside of the Jewish Museum Berlin by Daniel Libeskind (Image Courtesy @ Daniel Libeskind; Digital Image © The Museum of Modern Art/Licensed by SCALA/Art Resource, NT); (right) the Holocaust Memorial by Peter Eisenman (© The Museum of Modern Art/Licensed by SCALA/Art Resource, NT).

The phenomenological desktruktion (Heidegger 2014), or deconstruction (Johnson and Wigley 1988), of the notion of dwelling (oikos), a question Heidegger approached repeatedly in his work (Bonicco-Donato 2019; Nesbitt 1997), ultimately and terminally leads to its negation, its nothing, that is, territory denial. We can see this in Daniel Libeskind’s Jewish Museum Berlin (2001), that some architecture critics have explicitly associated with the notion of territory denial and that Libeskind himself associates with the notion of wall (in association with the notion of the Berlin Wall) (Graft Architekten and Libeskind 2018); the notion of territory denial is also apparent in his urbanist sketches (e.g. Site Plan in Open Line, Urban Competition, Potsdamer Platz Berlin, 1991 (Dorrian 2005)). Walls zigzag across space and completely reshape it. Perhaps another harrowing example of territory denial in deconstructivist architecture is Peter Eisenman’s Memorial to the Murdered Jews of Europe, or Holocaust Memorial, in Berlin (2003-2004). A large collection of concrete slabs occupy the space and deny the existence of anything else, physically and semantically.

In music, territory denial becomes the information overload in the New Complexity aesthetic of composers such as Brian Ferneyhough, James Dillon, Claus-Steffen Mahnkopf and Michael Finissy, where sound masses do not allow for interference, interaction or dialog. More specifically, we can think of the opening measures of La Terre est un Homme by Brian Ferneyhough where the density of the notation is akin to the territory denial of deconstructivist architecture. Music occupies and denies space through sound blocks or sensation blocks (in the sense of Gilles Deleuze). Also, visually, clear analogies can be made between the graphical scores of Sylvano Bussotti or John Cage among others (e.g. Siciliano (1962) or Rhizome (1970) by S. Bussotti, Fontana Mix (1958) or Hymns and Variations (1979) by John Cage) and deconstructivist sketches by architects such as Frank Gehry, Zaha Hadid, Peter Eisenman, Bernard Tschumi or Daniel Libeskind, to name a few (Johnson and Wigley 1988). Many of those are shown in (Nguyen and Tsabary 2023a, 2023b).

Territory denial, declined as an expropriation, is a recurrent theme in deconstructivist and memorial architecture as much as in the research-creation project The Cherry Orchard remodernisation. It is no coincidence that the aesthetic of memorials have found a clear expression in deconstructivist architecture (Parr 2008).

Errors and Broken Hammers

In a talk at the Pompidou Center in Paris about the then nascent field of computer music, to the question of what happens when a computer makes a mistake, composer Pierre Boulez is known to have answered that mistakes are sometimes more interesting than correct answers. Computer generated art excels at generating variations and candidates.

Steve Tomasula in his anthology Conceptualisms paraphrases Heidegger: a hammer that is used as a hammer is just a hammer, but a broken hammer is a poem. He then elaborates as follows:

In giving up its functionality it takes on the potential to be a rhetorical or philosophical object or conceptual art, just as no one today reads Galen’s medical advice except as history or literature (Tomasula 2022).

A broken rotation (or a broken algorithm in general) is a rotation algorithm that contains a programmatic error (i.e. a bug). Unexpected asymmetries are created that improve on the repetitive monotony of a computer algorithm. In previous works, we evolved random errors in algorithmic procedures (i.e. rotation algorithms) to generate surprising outcomes.

One could argue that most of the works of Johann Sebastian Bach are mechanical in nature, mechanistic, what one would qualify as algorithmic today and what was an artisanry in the days of the composer. However, in some of his works, one can clearly see an emergence phenomenon (in the sense of AI) from the set of generating rules, something only possible from the ineffability of the human touch, that one can only approximate as an unexpectedness. In some instances, Johann Sebastian Bach’s mechanical reproductions transcend their own generating process and transcend their tekhnè. Errors encourage surprise and the unexpected leads to emergence.

Combinatory Aesthetics

Combinatory aesthetics in the arts in the 20th century can be traced back to at least as far as Tristan Tzara’s  instruction in his How to Make a Dadaist Poem (1920) (Lewis 2007). Onfray (2022) traces this aesthetic back to the so-called anartistes mouvement of Alphonse Allais and the like while Nierhaus (2009) and Edwards (2011) find examples of combinatory aesthetics dating back to the music of the Medieval Ages. More recently, Charles Ives is also know to have superposed popular songs in his pieces (e.g. in Fourth of July (mm 100-102), the superposition of popular songs in a polyphony is meant to reproduce what could be heard on a 4th of July (Dragu 2022)). Ives was the son of a marching band leader and he would walk the streets with his dad loving how the sounds of multiple orchestras would sometimes meet at crossroads.

Brion Gysin, a poet who was expelled from the surrealist group by André Breton and a friend of William Burroughs (known to have randomized the order of the chapters of his novel The Naked Lunch - an idea he got from Gysin) was an important figure of the so-called combinatory poetry movement. The following is William Burroughs’ take on the “cut-up poem” technique:

The cut-up method brings to writers the collage, which has been used by painters for fifty years. And used by the moving and still camera. In fact all street shots from movie or still cameras are by the unpredictable factors of passers by and juxtaposition cut-ups And photographers will tell you that often their best shots are accidents... writers will tell you the same. The best writing seems to be done almost by accident but writers until the cut-up method was made explicit - all writing is in fact cut-ups. I will return to this point - had no way to produce the accident of spotaneity, You can not will spontaneity. But you can introduce the unpredictable spontaneous factor with a pair of scissors (Burroughs (2003) quoted in Rettberg (2019)).

Gilles Deleuze and Felix Guattari in Mille Plateaux speak of William Burroughs’ cut-up technique in the context of the rhizome as follows:

Consider Burroughs’ cut-up method: the folding of one text onto another, constituting multiple and even adventitious roots (it looks like a cutting) implies an additional dimension to that of the texts under consideration. It is in this additional dimension of folding that unity continues its spiritual work. It is in this sense that the most resolutely fragmented work can just as well be presented as the Total Work or the Grand Opus. Most of the modern methods for making series proliferate or for making a multiplicity grow are perfectly valid in a direction, for example linear, while the unit of totalization asserts itself all the more in another dimension, that of a circle or of a cycle. Each time a multiplicity finds itself caught in a structure, its growth is compensated by a reduction in the laws of combinations (Deleuze 2013, trans. by the author).

There is a clear relationship between the combinatory aesthetic of Brion Gysin and William Burroughs (the so-called cut-up technique) and computers. Early experimental poetry by Gysin included computer programs. Gysin’s most quoted poem is I am That I am (1960) based on a famous tautology from a text by Aldous Huxley; the permutations were programmed by Ian Somerville on a Honeywell computer (Burroughs and Gysin 1976). An excerpt is presented as follows:

I I THAT AM AM (Burroughs and Gysin 1976)

Erasure Aesthetics

Figure 2: An excerpt of O III from Ronald Johnson’s RADI OS I-IV (1977) (Johnson 2005). The fragmentation reminds Anne Carson’s translations of Sappho, NourbeSe Philip’s Zong!, but also Mallarmé’s Un coup de dé jamais n’abolira le hasard.

Another important combinatory work, that uses the particular technique of erasure, is Ronald Johnson’s RADI OS I-IV (Johnson 2005). The title is derived from an erasure on the title of John Milton’s Paradise Lost. Johnson took the first four books of Milton’s masterpiece and generated a new text through erasures. Johnson is known to have erased all references to God and Satan. Ross Hair does a genealogy of Johnson’s techniques, augmenting it with comments from Guy Davenport and William T. Dobson:

Davenport may be thinking of the Empress Eudoxia who wrote the history of Christ in verses from Homer, or of Proba Falconia who, in her Cento virgilianus, uses Virgil for the same purpose. In his book Literary Frivolities: Fancies, Follies, and Frolics, William T. Dobson refers to Falcona in his chapter, “Centoes or Mosaics.” According to Dobson: “A cento is properly a piece of patchwork and hence the term has been applied to a poem composed of selected verses or passages from an author, or from different authors, strung together in such a way as to present an entirely new reading.” Radi os clearly follows in this tradition of “patchwork” composition but, as Esdale stresses, “Radi os is not a collage of quotation that testifies to extensive reading; instead it is the result of intensive reading” (Hair 2010).

The relationship between Radi Os and Marcel Duchamp’s ready-mades has also been established (Hair 2010). Johnson also acknowledges the influence of Lukas Foss’ Baroque Variations which he discovered during a party in Washington. In the piece, Foss “cuts holes” in a Handel Larghetto from his Concerto Grosso Op. 6 No. 12, a Scarlatti sonata and a Bach prelude (Hair 2010). Johnson specifically writes:

Groups of instruments play the Larghetto but keep submerging into inaudibility (rather than pausing). Handel’s notes are always present but often inaudible. The inaudible moments leave holes in Handel’s music (I composed the holes) (Hair 2010).

The first documented act of erasure in the arts is often attributed to Robert Rauschenberg and his Erased De Kooning Drawing (1953): He wasn’t the author of the drawing and this is often considered crucial to the concept of erasure art (Hair 2010). The influence of John Cage and his chance procedures can also be felt. John Cage wrote his Hymns and Variations for 12 amplified voices in 1979 where he erased, using a chance procedure, parts of two hymns by William Billings: Old North and Heath (Cage 1979). Figure 3 shows the first page of Cage’s Hymns and Variations for 12 amplified voices and Figure 2 shows an excerpt of Radi Os I-IV. The main difference between Cage’s work and Johnson’s is probably the lack in intentionality (the use of chance procedures) in the first and the intentionality of the second (forceful removals of concepts such as God and Satan).

Most of the examples of combinatory aesthetics we have provided are linear, with the exception of Ives’ polyphonies and Philip’s Zong! which are layer-based (e.g. the use of palimpsests in Zong!).

Figure 3: First page of John Cage’s Hymns and Variations for 12 amplified voices (© 1979 by Henmar Press Inc. Reproduced by kind permission of C.F. Peters Corporation. All rights reserved.).

Experiments Using Deep Learning for Composition


Figure 4: Transformation matrices and nearest neighbour graphs computed from the MDS embedding of the matrices. The nearest-neighbour graphs resemble deconstructivist sketches: (Left) features generated from Schoenberg’s Klavierstücke Op.11; (center) features generated from an excerpt of Rachmaninov’s Second Concerto (Movement 2); (right) features generated from an excerpt of Simon Steen-Andersen’s Piano Concerto (the matrix is remarkably sparse while covering the space of pcs like white noise). The transformations in the Schoenberg and Rachmaninov excerpts have data concentrated in the upper-left corner (i.e. the subsets of pitch classes with a small number of values: single notes, two and three notes chords, etc.). The peculiar transformation of Simon Steen-Andersen’s Piano Concerto denotes the resolutely modern take of the piece at a structural level. From such graphs, a piece can be generated by performing a random walk on the graph and selecting nodes based on some rules. In our use of GAN’s, new instances of nearest-neighbour graphs are generated by neural networks and are used to generated combinatorial sketches. These graphs may look like deconstructivist architecture sketches and diagrams.

Richard Coyne describes the relational model as it emerged in modern linguistics as follows:

As linguists focus on the synchronicity of language it is the structures that become important, rather than individual elements. According to the developmental psychologist Jean Piaget (1896–1980), Structuralism ‘adopts from the start a relational perspective, according to which it is neither the elements nor a whole that comes about in a manner one knows not how, but the relations amongst elements that count’. The idea of ‘the relationship’ is very important in Saussure’s linguistics (Coyne 2011).

When deconstructing a corpus, to extract features, we extract the relationships between onsets (hashed as pitch class sets) in terms of the expected probability of their occurrence. Similar techniques include Hidden Markov Models (used as early as the work of Lejaren Hiller and Iannis Xenakis). We thus obtain a transformation matrix (or distance matrix) that can be represented as a nearest-neighbor graph using multidimensional scaling. This graph is exactly an equivalent of a neo-Riemannian Tonnetz in which the distance between chords has been scaled in function of a predefined corpus. More specifically, we use the following relation (i.e. decay function) to build our transformation matrices (Nguyen and Tsabary 2022):

δ(t)={   1,if  1et/n+c1   1et/n+c,if  1et/n+c<1\delta(t) =\left\{\begin{array}{ll}      1, & \text{if}\ \ 1-e^{-t/n} + c \geq 1 \\      1-e^{-t/n} + c, & \text{if}\ \ 1-e^{-t/n} + c < 1 \\\end{array}\right.

A discrete version has been developed recently to account for sparsity effects when using GAN’s or other deep learning networks (i.e. learning integers or classes instead of real numbers):

δ(t)={   Nk,if  k<=N   0,if  k>N\delta(t) =\left\{\begin{array}{ll}      N-k, & \text{if}\ \ k <= N \\      0, & \text{if}\ \ k > N \\\end{array}\right.

where k is the number of times a pitch class set impacts another and N some integer constant. The interest of obtaining a square matrix (where the columns and rows represent the 4096 subsets of the 12 pitch classes) is that it can be used as the input of deep learning neural networks such as Generative Adversarial Networks (GAN). Currently, GAN’s, with Stable Diffusion, are the backbone of many text-2-image systems such as DALL-E-2, Google Imagen and Midjourney (Ramesh et al. 2022; Saharia et al. 2022). GAN’s were introduced in (Goodfellow et al. 2020). The usual problem when using MIDI data is the irregularity of MIDI (i.e. variable length data). Algorithms such as GAN’s require fixed size input (which makes them practical for image and sound data, or transformation matrices such as the ones described in this paper).

Using GAN’s on a corpus of transformation matrices, we can then generate a multiplicity of transformation matrices based on the corpus, like text-2-image systems generate new images from training data. From a given generated transformation matrix, we can generate sketches that reorder the original corpus based on some random walk rules. Our technique is effectively generative (i.e. generating transformation matrices from trained GAN’s) and combinatorial (i.e. generating MIDI from transformation matrices). The deconstruction phase then consists of extracting transformation matrices and the reconstruction phase involves generating new transformation matrices and musical compositions (i.e. combinatorial sketches).

Figure 5: Transformation matrices computed using MDS. (Top) transformation matrices computed for each of the 3 movements of the 3 piano concertos of Serguei Rachmaninov; (bottom) transformation matrices generated using GAN’s trained on the 9 movements of the 3 piano concertos of Serguei Rachmaninov (Nguyen and Tsabary 2023b).

A Scene from The Cherry Orchard Remodernisation

Figure 6: “Scene 6: Giving” of The Cherry Orchard remodernisation (mm. 37-48). The family is enjoying the day outside when suddenly a peasant walks by. Grotesquely, he mentions that he was just hit by a mine and has been looking for the railway station ever since. At measure 38, an Am triad punctuates the transition from a more continuous orchestration (m. 37) to a passage generated by a GAN (i.e. a deep learning neural network) which is resolutely more deconstructivist aesthetically (mm. 40-48). Disjunction replaces continuity.

We describe an example of deconstruction (the process of deconstructing and reconstructing in the generation of music) using a sketch for “Scene 6: Giving” of The Cherry Orchard remodernisation opera (a work in progress that will be published and produced in late 2023-2024). The music of Scene 6 starts with a passage that was generated using automated orchestration on a sound file. After the first 2 minutes, which correspond to the time the family in the opera is enjoying the day in a picnic in the fields, by the old cherry orchard, a peasant suddenly appears on stage. He is walking by, limping and maimed by a stray mine. Grotesquely, he says that he walked on a mine and has been looking for the train station ever since. He asks for some change. An Am triad resounds at the moment the peasant appears (the motif used across the opera since the “Overture”) and is followed by a passage that was generated using AI (our process of deconstructing and reconstructing in the generation of music) and the first passage as corpus. This transition point is shown in Figure 6 (mm. 37-48 of Scene 6).

The aesthetic of this passage can have a clearly deconstructivist feeling attached to it just after a simple listening of the piece. Audio Example 1 (excerpt from the “Overture“) shows a passage that was composed by layering stochastically generated lines and that is suddenly cut by a chord progression starting in Am (the leitmotif of the opera). Audio Example 2 (excerpt from “Scene 6”) shows 2 passages separated by the Am leitmotif: the first passage serves as the corpus for the second passage composed using AI techniques. The layering of techniques and processes (e.g. neo-Riemannian chord progressions, stochastic music, deconstructivism using AI) creates particular textures. The second passage of Audio Example 2, the deconstructed passage, can also remind the disjunctive music generated by pointillist serialism or by a Buchla (i.e. an instrumental Buchla feel).

Audio Example 1: Excerpt from the “Overture“ of The Cherry Orchard remodernisation opera (digital mockup only as the opera is currently in the works). Multiple techniques are layered (e.g. stochastic music, neo-Riemannian chord progressions, automated orchestrations, broken rotation algorithms). A chord progression in Am is used as a motif for the whole opera.

Audio Example 2: Excerpt from “Scene 6: Giving“ (digital mockup only with no singers as the opera is currently in the works). A first passage composed using automated orchestration techniques on sound files is followed by a break using the Am motif of the opera. A second passage composed using a deconstruction-reconstruction process is then used. The passage clearly has a “deconstructed“ feel to it and coincides with the grotesque appearance of a peasant maimed by a mine on the stage.

The Phenomenology of Musical Deconstructivism

The process of deconstruction and reconstruction, from feature extraction to generation on a latent space, is attached to an aesthetic of deconstruction that is easily recognizable. Recombining works based on a well-ordering principle that is determined by transformation matrices rather than traditional musical rules and conventions is disjunctive. The process supports a combinatorial sketching approach (in the sense of deconstructivist architect Thom Mayne (Leach 2022)) where candidates are generated, modified and filtered by a composer.

Lewis Mumford believes that language (symbols) more than technology (tools) determines the human condition (Mumford 2010). The special place of AI in this dyadic opposition (i.e language/technology) is due to the fact that AI relies on symbolic constructions (programs, mathematical formalisms, data representation, corpus) in order to provide a utility (its toolness). For Jean Baudrillard (Baudrillard 2020), art is a system of objects (we will remember Marcel Duchamp's urinal or his Erratum Musical) while for Nelson Goodman, art is a "system" of languages or symbols (we will be reminded of Pierre Boulez's early integral serialism) (Goodman 1976). Our inception of AI art is effectively a system of languages (in the sense of Nelson Goodman (Goodman 1976)). We speak of artificial intelligences rather than artificial intelligence.

Also, one could correct Deleuze's take on music by speaking of music as the construction of sound-notation blocks rather than simply sound blocks to reflect the duality of (Western) music as we know it (even if music can effectively sometimes be exclusively sound block and sometimes sound-notation block intricacies). Music has two states (as light is both particle and wave): sound and notation. AI algorithms and CAC systems in music act either on the sound (e.g. OpenAI’s Jukebox or Google’s MusicLM), or on the notation aspect (e.g. ChatGPT, OpusModus or OpenMusic), and hybrid systems are also possible. The combination of the sound and the notation dimensions creates a particular territorialization of music in the hermeneutic space of music and theory. The symbolic aspect is important in our particular use of techniques coming from AI for the generation of musical content. It shows that there is still place for innovation in classically written music.

Mark Fisher in his book Ghosts of My Life: Writings on Depression, Hauntology and Lost Futures warns us:

Music culture is in many ways paradigmatic of the fate of culture in post-Fordist capitalism. At the level of form, music is locked into pastiche and repetition. But its infrastructure has been subject to massive unpredictable change: the old paradigms of consumption, retail and distribution are disintegrating, with downloading eclipsing the physical object, record shops closing and cover art disappearing (Fisher 2014).

This leads to the Auerbach-Said-Rancière notion of mimesis (Auerbach and Said 2013, Rancière 2018). The Auerbach-Said-Rancière mimesis is such that if literary realism is a mimesis on the episteme of reality, its interpretative systems, its codes and its codification, the equivalent in music, musical realism, is rather declined by the language of the so-called Common Practice Period, its formalisms, a-prioritisms and rules from the era of plainchant and Baroque to post-romantic chromaticism and the neo-Riemannian systems of Carol Krumhansl based on cognitive appraisals (Krumhansl 2005) or the mathematical formalisms of High Modernism (which can push the common language to the extreme and sometimes exceed it) (Jedrzejewski 2006, 2019).

Some representations favour some sets of solutions to defined musical problems: for example, the Tonnetz of Hugo Riemann favours consonant triads since these are placed closer to each other in the Riemannian representation of major and minor triads. Therefore, how music is represented carries significant weight in the production of some aesthetic field. Representations can be as trivial as numbering A as 1, C as 0 or identifying pith classes to a cyclical quotient group to more specific ways of organizing pitch, rhythm, timbre and harmony in some spatial metaphor, a space equipped with some distance. The diversity of pitch numberings is significant in post-tonal music: we know that Luigi Nono numbered his pitch classes from A = 0 since orchestras tune in A (Guerrero 2006) while Maderna seemed to have disliked 0 since the Ancient Greeks did not dispose of the zeroth (Neidhöfer 2023). Also, Forte and Rahn numbers were meant as ways to dissociate nomenclature from the tonal system. Interval class vectors are another popular representation of pitch and harmony in post-tonal music: they allow for quick resolution of common tone problems for example.

The deconstruction of musical codes by techniques borrowed from AI, what we call musical deconstructivism, is intended as a critique of aesthetic realism (in the sense of critical theory or critical composition by Nicolaus A. Huber or Helmuth Lachenmann for example). This argument returns to Heidegger’s notion of Verfallen (fallenness, inauthenticity) when the experience of being-in-the world is determined from the outside (i.e. traditions, culture, social pressure). Furthermore, it may only be coincidental that Heidegger writes that technology is a mode of understanding while it develops beyond our comprehension and control (Heidegger 1958). Alternatively, AI is a mode of understanding and the black box nature of deep neural networks is such that they evolve beyond our comprehension and control. They have constituted pocket worlds in the field of human reason.


We have attempted to give a context (an ambitus to use a musical term or a bracket in terms of phenomenology) for a deconstructivist aesthetic in music composition where a double process of deconstruction-reconstruction gives rise to a perceptible and particular sensation of musical experience. Such deconstructivist pieces are clearly combinatory and disjunctive in nature. Deconstruction can be felt in the artefacts of our deconstructivist process of composition (i.e. the nearest neighbour graphs generated by our technique resemble the works of deconstructivist architects). This combinatory sensation of disjunction was also exemplified in audio examples (cf. Audio Example 1 and 2) and score excerpts (cf. Figure 5). In this paper, our goal was to provide a (phenomenological) analysis of a particular technique we developed to compose music using AI. We provided examples of our technique based on a currently ongoing research-creation project to write an opera using AI techniques based on a remodernisation of The Cherry Orchard by Anton Chekhov.

Following Heidegger’s chicken-egg paradox on the artist and the artwork, we could then ask ourselves what comes first, the technique or the aesthetic. And finally, we could conclude by quoting a poem by Sung Dynasty poet Yang Wan-li:

Now, what is poetry?

If you say it is simply a matter of words,

I will say a good poet gets rid of words.

If you say it is simply a matter of meaning,

I will say a good poet gets rid of meaning.

“But,” you ask,“without words and without meaning, where is the poetry?”

To this I reply, “Get rid of words and get rid of meaning and still there is poetry.” (Hirshfield 1997)

If we take out the human from the poetry, is there still poetry?

No comments here
Why not start the discussion?