AI Has Entered Composition at the Weakest Point: Form
Procedure, decision, and responsibility after machine generation
AI becomes philosophically serious for composition at the point where generation begins to resemble decision.
At that point, the usual public debate around AI music no longer reaches the level at which the problem becomes compositionally and philosophically decisive. The public conversation has so far been shaped by synthetic songs, cloned voices, prompt-based production, copyright anxiety, streaming abundance, and the possibility that musical labour will be cheapened beyond recognition. Its force lies in the economies it will reshape, the legal categories it will strain, the authority of performers it will unsettle, and the already weakened conditions under which listeners are asked to encounter sound online. Yet my concern here belongs to a narrower and more severe region of the problem. I am writing as a composer and philosopher of music, from within the tradition of art music, where AI raises its most demanding questions through form, decision, responsibility, and the status of the musical work.
The problem first becomes visible where public spectacle around AI music gives way to compositional procedure, in the environments where composers test material, extend patterns, interact with performers, model stylistic behaviour, respond in real time, and encounter continuations that begin to resemble musical thought. AI has therefore entered serious music through research environments, improvisation systems, machine listening, corpus-based generation, composer-facing tools, and the longer institutional history of computer music.
IRCAM’s Somax2 gives this shift a precise institutional and technical form. IRCAM describes Somax2 as an application for musical improvisation and composition, implemented in Max, trained on a symbolic MIDI corpus, and designed to provide stylistically coherent improvisation while listening and adapting to a musician, audio source, or MIDI source in real time.¹ In the research literature around the same system, Somax2 is described as an AI-based multi-agent environment for human-machine co-improvisation, one that generates stylistically coherent streams while continuously listening and adapting to musicians or other agents.² ICMC Boston 2025 announced a special call for music employing Somax2, including a dedicated concert for works written with that environment and for ensemble performance with Somax2.³ Guildhall’s 2025 event The Odd Couple: Human & AI Making Music in the Moment presented Oded Ben-Tal’s live real-time system as a musical partner for improvisation with classical and jazz pianists.⁴ Tarik O’Regan’s BBC Radio 3 documentary The Artificial Composer, first broadcast on 5 October 2025, explored AI’s impact on classical music through machine-generated violin sonatas and the use of AI suggestions in the making of a new orchestral piece.⁵
These examples require a careful reading because their significance appears in the relocation of compositional agency into procedural environments, into systems that listen, adapt, continue, recombine, and return material before the musical work itself has been publicly redefined by machine generation. AI, in this sense, has entered serious composition through procedure before it has transformed the public identity of the work.
The distinction acquires its pressure once AI is placed within the older history of serious composition itself, a history already marked by procedures, constraints, delegated processes, and technological mediation. Serialism, chance operations, algorithmic composition, computer-assisted composition, stochastic processes, spectral analysis, live electronics, open forms, improvisation systems, notation as constraint, and process-based composition all belong to the longer history through which composers have delegated, formalised, restricted, interrupted, or displaced the immediacy of preference. Lejaren Hiller and Leonard Isaacson’s Illiac Suite had already placed the electronic computer within the production of compositional material in 1957;⁶ Xenakis’s Formalized Music gave one of the most influential theoretical articulations of stochastic and mathematical procedures in composition;⁷ and IRCAM itself has long stood at the point where research, software, analysis, and compositional practice meet. The history of composition already complicates any simple account of authorship as the uninterrupted transfer of private intention into sound, since intention has long operated alongside procedure, constraint, notation, instrumentality, technology, and forms of delegated decision.
AI enters this history with a different order of pressure. Its novelty lies in the scale, speed, fluency, adaptivity, and decision-like behaviour of generated continuation, especially when a model extends a stylistic field without knowing why that field should continue, or when a machine-listening system responds to a performer in real time without assuming responsibility for the formal consequences of that response. Older procedures could produce, displace, calculate, restrict, or search; AI adds a new difficulty by making continuation appear locally intelligent, even where the deeper logic of form remains unresolved. Continuation begins to imitate compositional thought at the very point where its categorical difference from thought must be held most firmly.
A procedure may generate material, a model may extend a pattern, and a system may return stylistically plausible behaviour; the threshold of the work, however, is reached only when these materials begin to sustain demands that exceed their own production. Material then undergoes a change of status. It ceases to appear merely as available sound and begins to function within a field of consequence, where density must justify its weight, fluency its continuation, and complexity its claim to necessity. A passage may resemble a compositional language while failing to establish why its events should follow, interrupt, answer, or alter one another. A sequence may continue persuasively for several moments and still leave the larger temporal field untouched. What appears locally plausible remains compositionally undecided until it can bear consequences beyond the immediate surface that produced it.
The technical literature reaches the same difficulty from another direction. The Music Transformer paper begins with a problem that every composer already knows in practice, though under different names, since music depends on recurrence, self-reference, return, and altered repetition across several temporal scales, from the local life of a motif or phrase to the more distant recognition of a section or transformed relation.⁸ For a generative system, this is the difference between producing a plausible next event and sustaining the conditions under which later events remain answerable to earlier ones. Recent surveys of AI-generated music evaluation reveal a related instability, since objective metrics may be easy to compute while failing to tell us whether a passage carries musical force, and subjective evaluation depends on human listening tests that are difficult to standardise or translate into stable criteria of judgement.⁹ Generation can become fluent before it becomes formally accountable. A system may continue, vary, repeat, or extend; the harder question is whether those continuations acquire enough consequence to be heard as more than successful local behaviour.
For composition, the decisive question reaches beyond the attribution of production and enters the harder terrain of consequence. What makes a passage hold together strongly enough to demand continuation, interruption, return, or completion? Which event alters the weight of another? Which duration earns its place? Which gesture changes the field? Which silence carries structural force? Which repetition gathers consequence rather than merely recurring? Which rupture alters the whole? Which ending stabilises the demands that preceded it?
AI enters composition through this vulnerable terrain. It exposes the places where procedural production has been allowed to stand in for formal necessity, and where a compositional language, once sufficiently repeatable, begins to behave less like thought than like a repertoire of recognisable operations.
The diagnosis becomes most severe when it turns from machines back toward composers. AI unsettles serious composition because it exposes the degree to which compositional practice can already become semi-automatic, especially where density accumulates without formal demand, seriousness becomes institutionally legible before it becomes structurally compelling, spectral or post-spectral colour hardens into ornament, complexity circulates as cultural sign, procedure takes the place of development, and notation appears difficult while asking too little of the work’s internal necessity. These tendencies are recognisable, but they require careful naming, since the diagnosis loses force once it slips into laziness, resentment, or stylistic policing. They remain part of the current condition of serious music, where a piece may acquire the signs of contemporaneity more easily than the demands that make it necessary.
The critique has to proceed through a stricter distinction between compositional difficulty and the mere appearance of difficulty. Density, fragmentation, discontinuity, resistance, and technical difficulty can belong to the highest demands of musical form when they intensify the work’s temporal, perceptual, and structural claims. Simplicity, under the same measure, offers no guarantee of clarity; it can remain empty, decorative, sentimental, or evasive. A work may be sparse, dense, fractured, continuous, even violently discontinuous, and still acquire necessity when its materials are made to bear the demands placed upon them. Complexity alone secures nothing. The question is whether the work can sustain its own demands under conditions of density, rupture, duration, and return.
AI makes that evasion more difficult because it weakens the authority of signs that once seemed to protect compositional thought. A generative system can produce stylistically legible material, extend a texture, respond to a performer, or simulate recognisable musical behaviour, while leaving untouched the harder question of what has been decided, intensified, refused, or made necessary. Plausibility no longer proves thought. Extension no longer proves development. Responsiveness no longer proves creativity. Style no longer protects authorship. The composer is left with a more severe question. What has compositional decision made necessary beyond the system’s capacity to make continuation plausible?
The question is dangerous because human beings also continue. A composer can become automatic through habit, fashion, institutional expectation, technical comfort, career pressure, inherited gesture, or a private syntax that has stopped resisting the hand. Automation is usually imagined as a technological force arriving from outside the human act, while composition has always carried the risk of its own internal automation. The hand remembers too quickly. The ear accepts too soon. The software encourages certain solutions. The institution rewards certain surfaces. The commission deadline stabilises the familiar. The composer returns to what already works, and the work begins to close before it has truly opened. AI unsettles composition by making this human automaticity newly audible.
The panic around replacement therefore gives way to a more exact problem of responsibility. In serious composition, authorship lies in what the composer allows to govern the work, what is accepted, refused, altered, intensified, suspended, or brought into consequence. The composer may work with sketches, algorithms, AI systems, improvising agents, software, models, machine-generated suggestions, performers’ interventions, notation systems, or constraints devised in advance. Each of these can alter the site of authorship without dissolving it. Authorship changes its location, moving from the mere origin of material toward the responsibility for what is allowed to stand, develop, return, or disappear.
In any AI-assisted compositional situation, responsibility rests with the human agent who allows material to enter the work. A generated passage may arrive with local fluency, stylistic familiarity, or even a momentary sense of direction, yet the compositional question begins only when the passage is tested against the work’s larger demands, against what it changes, what it permits, what it prevents, what it obliges, and whether it can survive beyond the immediate plausibility of its surface. The composer’s responsibility lies in the harder labour of hearing what should remain, what should be altered, what should be refused, and what must be reduced until the work no longer depends on the abundance of its possibilities.
The language of co-creativity enters this debate with a useful ambiguity, and its usefulness depends on refusing the comfort of the term itself. In the Somax2 literature, distributed creativity, mixed musical reality, co-improvisation, and cross-learning name a real alteration in the musical situation, because a system that listens, stores, adapts, recombines, and returns material changes the field in which performers and composers act. That alteration should be recognised without being allowed to settle the question too quickly. A responsive system can affect timing, expectation, continuation, and the immediate horizon of musical response, yet the redistribution of technical agency leaves the ethical and artistic distribution of responsibility unresolved. The decisive issue remains the situation in which events are framed, accepted, altered, refused, or allowed to acquire consequence.
AI therefore sharpens the problem of decision rather than removing it. Its danger for serious composition appears most clearly in its capacity to flatter habits that were already weak, multiplying material where material is already abundant, extending surfaces that already persuade too easily, accelerating production where speed has already damaged judgement, and offering continuation where the more difficult demand is consequence.
At this stage the work-concept becomes unavoidable. Lydia Goehr’s account of the musical work remains decisive because it locates the “work” within history, as a regulative concept that came to organise musical practice with particular force around 1800.¹⁰ Machine generation approaches that inherited concept from another direction. Generated music may appear as output, file, performance, stream, fragment, sketch, suggestion, or interactive event; yet the threshold of the work is crossed only when such material can sustain demands that exceed the fact of its production.
A generated object may be fluent, complex, even compelling, while still leaving unresolved the question of what it can sustain beyond its own appearance. Continuation begins to matter only when it carries consequence beyond extension. Material acquires force when it places demands on what follows. Complexity, however technically or stylistically persuasive, remains insufficient until it begins to bear formal weight. A generated passage survives its first plausibility only if it alters the conditions under which later events can be heard. The final question is one of decision. What allows output to stand as a work?
As generated fluency becomes easier to produce, that question will become less avoidable. The work after machine generation will survive, where it survives at all, through a renewed severity of judgement. Composers will need a harsher honesty about what their materials can bear; listeners, a more disciplined patience with duration, return, and consequence; institutions will have to distinguish necessity from legible seriousness; and criticism, if it is to remain serious, will need to recover the language of form without returning to inherited dogma.
AI enters composition by asking whether what we call form still possesses the authority to command our decisions.
Notes
¹ IRCAM, “Somax2,” IRCAM project page, accessed 10 May 2026.
² Gérard Assayag, Laurent Bonnasse-Gahot, and Joakim Borg, “Cocreative Interaction: Somax2 and the REACH Project,” Computer Music Journal 46, no. 4 (2022): 7–25, https://doi.org/10.1162/comj_a_00662.
³ IRCAM Forum, “ICMC 2025 - CALL for SOMAX2 MUSIC PIECES,” event page, 2025, accessed 10 May 2026.
⁴ Guildhall School of Music & Drama, The Odd Couple: Human & AI Making Music in the Moment, event page, 30 March 2025, accessed 10 May 2026.
⁵ Wise Music Classical, “Interim (from The Artificial Composer),” programme note for Tarik O’Regan, 2025, accessed 10 May 2026.
⁶ Lejaren A. Hiller, Jr., and Leonard M. Isaacson, Experimental Music: Composition with an Electronic Computer (New York: McGraw-Hill, 1959).
⁷ Iannis Xenakis, Formalized Music: Thought and Mathematics in Composition, rev. ed., additional material compiled and edited by Sharon Kanach, Harmonologia Series, no. 6 (Stuyvesant, NY: Pendragon Press, 1992).
⁸ Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Ian Simon, Curtis Hawthorne, Noam Shazeer, Andrew M. Dai, Matthew D. Hoffman, Monica Dinculescu, and Douglas Eck, “Music Transformer: Generating Music with Long-Term Structure,” paper presented at the International Conference on Learning Representations, 2019.
⁹ Zeyu Xiong, Weitao Wang, Jing Yu, Yue Lin, and Ziyan Wang, “A Comprehensive Survey for Evaluation Methodologies of AI-Generated Music,” arXiv:2308.13736, 2023, https://doi.org/10.48550/arXiv.2308.13736.
¹⁰ Lydia Goehr, The Imaginary Museum of Musical Works: An Essay in the Philosophy of Music (Oxford: Clarendon Press, 1992), 89–119.