Paper deep dive

The Aura in the Machine: Genealogy and the Status of the Work of Art in the Generative Era

Giorgio Presti

Year: 2026Venue: arXiv preprintArea: cs.CYType: PreprintEmbeddings: 85

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%

Last extracted: 7/21/2026, 5:57:58 AM

Summary

This paper argues that Generative AI is an industrial-scale manifestation of historical generative art processes rather than a technological rupture. It proposes a taxonomy of generative systems as medium, artwork, or instrument, and introduces 'Manifestation' as a third ontological status for generative works, transcending the original/copy dichotomy. The author redefines the artist's role from craftsman to systems designer and entropic agent, suggesting that the Benjaminian aura condenses upon the productive system rather than dissolving.

Entities (13)

Generative Artificial Intelligence · technology · 98%Walter Benjamin · person · 95%Artwork · category · 93%Instrument · category · 93%Medium · category · 93%Manifestation · concept · 92%Algorithmic Repetition · concept · 90%Model Collapse · risk · 90%Environmental Enrichment · solution · 88%Johann Sebastian Bach · person · 88%John Cage · person · 88%Brian Eno · person · 85%Jean Tinguely · person · 85%

Relation Signals (12)

Generative Artificial Intelligence → isclassifiedas → Industrial-scale manifestation

confidence 95% · This paper frames Generative Artificial Intelligence (AI) not as an unprecedented technological rupture, but as an industrial-scale manifestation of a deeply rooted historical process.

Generative Systems → hasfunctionalcategory → Instrument

confidence 93% · A taxonomy of generative systems is proposed across three functional categories (medium, artwork, instrument)

Generative Systems → hasfunctionalcategory → Medium

confidence 93% · A taxonomy of generative systems is proposed across three functional categories (medium, artwork, instrument)

Generative Systems → hasfunctionalcategory → Artwork

confidence 93% · A taxonomy of generative systems is proposed across three functional categories (medium, artwork, instrument)

Benjaminian Aura → condensesupon → Productive System

confidence 92% · the Benjaminian aura does not dissolve in the generative era but condenses upon the productive system.

Generative Artificial Intelligence → hashistoricalprecedent → Generative Arts

confidence 92% · Through a genealogy of generative arts, it shows how AI's questions on authorship and creativity have precise historical precedents.

Jean Tinguely → created → Meta-matics

confidence 90% · One must cite Jean Tinguely’s Meta-matics. These are machines powered by an electric motor... They are machines that draw or rather, works of art that draw.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:This paper frames Generative Artificial Intelligence (AI) not as an unprecedented technological rupture, but as an industrial-scale manifestation of a deeply rooted historical process. Through a genealogy of generative arts, it shows how AI's questions on authorship and creativity have precise historical precedents. A taxonomy of generative systems is proposed across three functional categories (medium, artwork, instrument), the attribution of which is editorial rather than ontological. From individual cognitive atrophy to Model Collapse, the systemic risks of creative automation are identified; environmental enrichment is proposed as an antidote. The role of the artist undergoes a radical metamorphosis: from craftsman of the object to entropic agent, systems designer, explorer, and negentropic curator. This pipeline-based taxonomy rests on a specific premise: the algorithmic system remains medium, instrument, or artwork, while creative agency resides in the humans distributed along it. Algorithmic Repetition is introduced as the aesthetic degeneration of aligned generative systems; the Benjaminian aura does not dissolve in the generative era but condenses upon the productive system. Manifestation is proposed as a third ontological status for generative works, transcending the dichotomy between original and copy. To support the proposed theses, two complementary aspects are examined: the radicalization of distributed authorship; and the reevaluation of older generative models, whose instability constitutes an aesthetic degree of freedom lost by recent ones.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

83,959 characters extracted from source content.

Expand or collapse full text

The Aura in the Machine: Genealogy and the Status of the Work of Art in the Generative Era Giorgio Presti 1* 1* Department of Computer Science, Universit`a degli Studi diMilano, Via G. Celoria, Milan, 20133, Italy. Corresponding author(s). E-mail(s): giorgio.presti@unimi.it; Abstract This paper frames Generative Artificial Intelligence (AI) not as an unprecedented technological rupture, but as an industrial-scale manifestation of a deeply rooted historical process. Through a genealogy of generative arts, it shows how AI’s questions onauthorship and creativity have precise historical precedents. A taxonomy of generative systems is proposed across three functional categories (medium,artwork, instrument), the attribution of which is editorial rather than ontological. From individual cognitive atrophy toModel Collapse, the systemic risks of creative automation are identified;environmental enrichmentis proposed as an antidote. The role of the artist undergoes a radical metamorphosis: fromcraftsman of the objecttoentropic agent,systems designer,explorer, andnegentropic curator. This pipeline-based taxonomy rests on a specific premise: the algorithmic system remains medium, instrument, or artwork, while creative agency resides in the humans distributed along it. Algorithmic Repetitionis introduced as the aesthetic degeneration of aligned generative systems; the Benjaminian aura does not dissolve in the generative erabut condenses upon the productive system.Manifestationis proposed as a third ontological status for generative works, transcending the dichotomy betweenoriginalandcopy. To support the proposed theses, two complementary aspects are examined: the radicalization of distributed authorship; and the reevaluation of older generative models, whose instability constitutes an aesthetic degree of freedom lost by recent ones. Keywords:Generative AI, Generative art, Computational creativity, Artistic authorship 1 Preface Navigating the space of art and Generative AI requires overcoming what Charles Percy Snow called the “Two Cultures” problem [ 1]: the histor- ical disconnect between humanists and technolo- gists. Generative AI is precisely the point where these two cultures are forced to intersect: one cannot effectively use a tool that they do not understand, nor can one understand the tool with- out the cultural background necessary to interpret its implications. This contribution therefore addresses the issue without any discontinuity between technical and humanistic aspects, aiming to provide an inte- grated framework that is useful both to those 1 arXiv:2607.17940v1 [cs.CY] 20 Jul 2026 working in art and those in technology, and espe- cially to those operating at the intersection of these two cultures. While recent literature applies the Ben- jaminian framework to Generative AI [2–4] and proposes taxonomies of distributed authorship [5], and analytical philosophy addresses the onto- logical status of generative works in terms of functional continuity [6], this contribution distin- guishes itself by: a historical genealogy rooting current questions in pre-digital artistic practices; a taxonomy of roles mapped onto the Generative AI technical pipeline; and the concept ofmani- festationas a third ontological status that tran- scends (rather than redefines) the original/copy dichotomy. Finally, a necessary terminological clarifica- tion: When this paper attributes creativity or authorship to human beings and denies them (at present) to algorithmic systems, the criterion is not ideological but epistemic: the capacity for embodied experience of the world is considered here a fundamental requirement for the emergence of creativity [7–9]. This capacity is, in principle, extendable to other biological agents and, hypo- thetically, to future artificial systems endowed with a body or capable of experiencing emotions. The limitations attributed to AI herein are there- fore contingent upon current architectures, rather than absolute properties of any possible artificial system. In instances where the text uses “human” as an adjective, it should be read in this broader sense. 2 Data, Rules, and Knowledge To understand the dynamics of Generative AI and its capacity to produce results of such complexity, it is necessary to take a step back and analyze the logical and computational elements underpinning it. In fact, AI systems do not operate through intu- ition or by virtue of an unfathomable principle, but rather rest on extremely concrete founda- tions: they process immense masses of informa- tional fragments which, interacting on macro- scopic scales, give rise to unexpected emergent phenomena [ 10]. Before exploring the mechanisms of neural networks or their artistic applications, it is essential to understand how, starting from an inert mass of data, one can arrive at an active rep- resentation of knowledge; one so complex that it succeeds in mimicking certain traits of biological creativity. In the field of computer science,knowledge (information usable by a machine) can be struc- tured in various ways. One method istrees, such as the encyclopedia of Diderot and d’Alembert [11], which first attempted to represent all human knowledge as a tree. A more refined, ontologically robust alternative is graph-based representations likeknowledge graphs: entities linked by relation- ships that can be traversed to retrieve desired information. Knowledge may also be represented specifically through aset of rules, an implicit rep- resentation similar to flowcharts or software where knowledge is crystallized into rules executed to produce responses. Finally, it can be represented via theparameters of a machine learning system (such as the parameters of a regression function or the weights of a neural network). Trees and graphs are inert: they are orga- nized information that does not represent any- thing executable. The rules and parameters of a machine learning system, by contrast, are active, executable representations. Specifically, the dis- tinction between these latter two models (the one based on sets of rules and the one based on machine learning systems) represents two very dif- ferent approaches to software development. In the first case, there istraditionalsoftware, in which the developer analyzes data and defines an algo- rithm that makes decisions. In the second case, through machine learning, the process works in reverse: data and decisions are fed into a system that, via statistical techniques, outputs rules (to be subsequently applied to new data to generate new decisions). These principles, rarely brought into focus out- side a technical context, form the foundation of this discussion, in which the inquiry into “how, where, andwhatknowledge is encoded” guides reflections on authorship, the role of the artist, and the nature of the work. 3 Programming as a Creative Act and Philosophical Exercise To frame all of this within the context of the rela- tionship between art and generative systems, it 2 is necessary to revisit at least a core set of con- cepts from art history and philosophy through the modern lenses of computer science and AI (for a more in-depth historical overview, see future publications). The way traditional software is written is rooted in a pre-digital culture. In an evoca- tive metaphor, Matt Butcher [12] draws a parallel to the paradigms proposed by Plato and Aristotle. Plato defines earthly entities as imperfect instances of immutable, abstract con- cepts, contrasting with the Aristotelian view that there exists only a material, earthly reality shaped by specific processes. Reinterpreting these two philosophers today, one can associate Pla- tonic thought withobject-oriented programming (abstract classes, instantiated as objects), whereas Aristotelian thought can be seen as an archetype offunctional programming(data transformed by functions). If it is true that philosophy consists of a creative act aimed at modeling reality, then pro- gramming must also inherit these characteristics [13]: Programming consists of creating an ontol- ogy, ametaphysical systemregarding a specific problem. Programming is the creation of new worlds, leaving them free to evolve and to manifest their own emergent properties through execution on a computer. Programming should be the pre- ferred activity of philosophers because it allows them to test the models of the world they propose, provided that these models are unambiguous, complete, and finite [12]. Programming is thus a creative action and a philosophical exercise; soft- ware, therefore, cannot be considered merely as a tool, but can by extension also be considered an art object. Not only the output of a program, but the program itself, thevery process, can be art [13–15]. Software can, therefore, be situated within a broader framework: that of theGenerative Arts [15]. As early as 1920, Naum Gabo wrote that a work of art is not an object, but a process: something that manifests in an object without, however, coinciding with it [16]. The system is not a transparent medium for authorial intent, but the very body of the work itself. 1 1 Although Galanter, in [15], defines the system as the methodthrough which the work is realized, he also recog- nizes cases where the system itself coincides with the artwork, Consequently, the figure of the artist under- goes a profound transformation. Ceasing to be the Demiurge, the craftsman of the finished object, they evolve instead into asystems designer: Pla- tonic in modeling the Hyperuranion from which a Demiurge will derive forms, and Aristotelian in defining the processes that will shape matter. 4 Genealogy of Generative Creativity Once software has been established as a legitimate artistic outcome, it becomes possible to engage in a more specific discourse on Generative AI. To do so, it is useful to frame AI within the broader historical context of Generative Arts; not treat- ing it as an anomaly, but rather reading it as an industrial-scale manifestation of a deeply rooted historical and cultural process. This discussion traces several stages of this process (among many others) that help frame the artist as anegen- tropic curator(one whoreducesentropy) and an explorerof latent spaces (understood both as spe- cific spaces within the machine learning cotext and as generic spaces of the possibilities inherent in a system). An essential stage of this journey is the analy- sis of fugues. A fugue is a musical form in which, starting from a small thematic fragment, an entire composition can be developed. This inherently possesses a generative nature, but there is one composer in particular worth mentioning. Around 1750, Johann Sebastian Bach undertook a monu- mental feat by studying all the fugues written up to that point. After studying this corpus, he syn- thesizedThe Art of Fugue[ 17], which is regarded as one of the most profound representations of the concept of the fugue. What Bach achieved is not far removed from what occurs with AI: training a neural network (biological, in his case) to generate new works. Also in the 18th century, theMusical Dice Gameemerged. This game consists of writing small musical fragments, associating them with numbers, and then using dice rolls to assemble these fragments into a single sequence. The qual- ity of the result depends entirely on the quality, demonstrating a degree of variability that will be revisited later in this paper. 3 compatibility, andmodularityof the written frag- ments. This is not entirely different from what occurs when using Generative AI systems to cre- ate music. The difference lies in the fact that, in AI, the stochasticity of the process is heavily influenced by the training data and the prompt (though not always in a deterministic manner). However, the importance of the initial musical cor- pus remains central to both: even if elements are not simply copied and pasted but instead encoded within the neural network weights, the quality of the training data is a crucial factor for the quality of the output. A significant version of the game was formalized by Johann Kirnberger in 1757 [18], who worked extensively not only to select frag- ments but also to define some rules, leaving the player with the task of rolling the dice and judging the results. Continuing the parallel between history and modern AI, one must consider L ́aszl ́o Moholy- Nagy’s workTelephone Pictures[19], a Bauhaus professor who in 1923 attempted to use telephone mediation as an artistic experiment. Moholy-Nagy calls an enamel company and dictates the work he intends to create over the phone. The worker, likely unaware of the artistic purpose, attempts to execute the instructions as best as possible, thereby creating the piece. This serves as anante litteramarchetype of prompt engineering. Indeed, Moholy-Nagy’s intention was to emphasize the role of the artist as aproducer of conceptsrather than as a craftsman physically involved in the realization of the work. Also in the twentieth century, Bruno Munari wroteRegola e Caso(Rule and Chance) [20], in which he essentially praises the balance between rule and randomness, between determinism and complete aleatoricness. He argues that determin- ism is boring, complete randomness is too unset- tling, and that art must find the equilibrium between the two. Balancing determinism and ran- domness is indeed what all artists do, but it is also what neural networks do: there is a strong deter- ministic component emerging from the data, yet the final output is deliberately rendered slightly aleatoric via atemperatureparameter (cf. Sec.5) that ensures a certain degree of variability and unpredictability [21]. It is worth noting that Munari had already radicalized this insight in his 1952Manifesto del Macchinismo(Manifesto of Machinism) [22], in which he explicitly urged artists to “distract machines by making them function irregularly.” This vision foreshadows the practice of forcing sys- tems out of their statistical equilibrium to achieve aesthetically interesting outcomes, and (alongside the discourse on aleatoric processes) anticipates Galanter’s reflections oncomplex systems(sys- tems poised between perfect order and complete disorder) and their exploration [15]. When discussing aleatoricism, one cannot avoid mentioning John Cage, a twentieth-century composer who held a radical view regarding the importance of the composer. Cage rejected the cult of Western myth; according to him, the human being was not the author of a work: for Cage, the human being is aliberator of music[23] that already exists in nature. Through his compo- sitions, he sought to completely expel the concept of choice from the creative process, dismantling the European idea of music based on the centrality of the composer. Cage’s perspective, at first glance, seems to resonate with Alfred Kroeber’s superorganicism [24], which posits that the role of the individual is nearly negligible compared to the influence of soci- ety. Being shaped by a preceding culture, the act of creation becomes merely serving as a conduit for an idea already in the air; the individual act- ing as a manifestation and inevitable expression of the entire culture. In other words: had Mozart not been born, music similar to his would likely have emerged from another figure within that specific historical period and cultural milieu. AI, with its massive volume of human training data, can be read as a technical and circum- scribed manifestation of superorganism. However, there is a fundamental distinction between Cage and Kroeber worth addressing. Unlike dice, the I Ching, or any other aleatory device, society (and AI) does not operateonlyby pure chance. Cage sought the silence of the ego; he aimed to nul- lify human intention to let nature speak. Modern language models do almost the opposite: they gen- erate the white noise of thousands of human egos crystallized within the training data. Cage wanted art without memory; AI is, by definition, pure memory [6]. Two things that appear similar on the surface but are profoundly antithetical, converg- ing only in the idea of the dissolution of thesingle author. 4 4.1 Machines: A Taxonomy of Generative Systems Another relevant example can be found in Jean Tinguely’sM ́eta-matics[ 25]. These are machines powered by an electric motor that, through nearly random movements dictated by the shape of rotating cams (knowledge crystallized into a delib- erately imprecise mechanical “software”), draw upon sheets of paper. They aremachines that drawor rather,works of art that draw. Beyond the obvious parallel with image-generating AI, what warrants closer examination is the fact that Jean Tinguely distributed these machines not as instru- ments, but as works of art. He left the viewer to grapple with questions such as: “Are the draw- ings it produces standalone works of art? Are they parts of the same work, like detached metastases that people can keep at home? Or are they merely a byproduct of the work?” Continuing with the discussion of machines that create art, one must cite the work of the Barron spouses, who in 1956 created what is considered the archetype of science fiction film soundtracks. For theForbidden Planetsoundtrack [26], they created small electronic circuits that emitted sounds; they then recorded these sounds whiletorturingthe small electronic circuit, caus- ing it to degenerate and self-destruct. Through careful editing of the recorded materials, they sub- sequently compiled the soundtrack. This is an example of machines creating art, used not as the work itself, but as a medium for creating a fixed work, specifically based on theaesthetics of fail- ureand destruction (anticipating the concept of Circuit Bending). Up to this point, computers have not yet been discussed. An early instance relevant to this study can be found between 1956 and 1957, when the chemist Lejaren Hiller decided to aban- don chemistry to devote himself to computer- generated music. Together with Leonard Isaac- son, Hiller created theIlliac Suite[27,28], a four-movement composition considered the first machine-generated composition. They imple- mented software designed to compose an original score. The software combines sets of rules, statis- tical models, and mathematical models (thereby crystallizing knowledge across various modalities). This provides another example of a process that is not the artwork itself (or rather, is not marketed as such) but merely a medium through which a specific artwork was created (albeit one that is potentially reusable). By the 1960s, it became evident that the artis- tic use of machines had become influential across all forms of human expression. The so-called3N group (Frieder Nake, Georg Nees, and Michael Noll, considered a group more for chronological than geographical reasons) [29] began to popular- ize the idea of computer-based Generative Art, laying the foundations for moderncreative coding. At the same time, Nanni Balestrini was working on hisTristano[30]: a literary work composed of fragments from various textual sources, ran- domly recombined and printed (at least in the author’s original vision, which only materialized in 2007 [31]) as unique copies, each the result of different executions of combinatorial software. Both works are examples ofprocess as artwork, whose results (always different yet always similar) raise the same questions as M ́eta-Matics. These artifacts are neither originals nor copies: they sug- gest the necessity of a third ontological status (cf. Sec.10). Returning to the realm of music, but (unlike theIlliac Suite) citing a case where the process isthe work, one can discuss the work of Brian Eno and how he popularized the concept of gen- erative music. To createMusic for Airports[32], Eno designed loops of varying lengths to be played simultaneously, so that they interlock differently with each repetition, creating a harmonic tex- ture that only repeats after very long intervals. Eno would have liked to deploy a system that allowed for the experience of this generative quasi- infinity (analogous to Balestrini, who intended for every copy to be unique), but he had to settle for a fixed duration of 48 minutes. The original artwork is meant to be experienced in its full com- binatorial and mutable nature; listening to only one pre-recorded version is perceived as a rup- ture, betraying the fundamental centrality of the process that generates it. Across all these examples, beyond the paral- lels with AI, three primary ways of interpreting generative systems emerge: • they can serve asmediafor producing a specific, fixed work, such as theIlliac Suite; 5 • they can be parts of the entireartwork, such asMusic for Airports, where the process is inseparable from the experience of the work; • or they can function asinstruments, distributed as permanent generators of content, such as the Musical Dice Game(which does not produce a single artwork, but rather serves as a tool to produce many). A work does not unequivocally fall into one category or anothera priori. The distinction lies in the mode of engagement and publication, rather than in the system’s structure. It is the artists (as well as critics and audiences) who determine which of these three categories the object under exami- nation belongs to. This suggests that any general definition of generative art (e.g., [ 15]) should move away from framing the system as either a tool or an end in itself, leaving such a classification open to each individual work. This taxonomy applies as much to Gener- ative Art in general as it does to Generative AI in particular. For example, the researchers behind ChatGPT, Suno, or Stable Diffusion could have conceived of them asworks of art; gen- erative works that produce an infinite array of textual, musical, or graphic manifestations upon a user’s request. Instead, they chose to market them asinstruments. Had they made the former choice, public perception of the Generative AI phenomenon might be very different. 5 On Statistical Prediction It is useful now to briefly recall how AI works in order to highlight certain aspects relevant to the present discussion. As illustrated by XKCD comic strip no. 1838 [ 33], for a computer scientist, AI is a system capable of providingapparentlyintel- ligent responses. They do not need to be truly intelligent; they only need to appear so. A Large Language Model, in particular, has no way of distinguishing truth from falsehood; it merely pro- duces extremely plausible statements; it is up to the user (or an automated grounding system) to verify them. A Neural Network (a specific type of machine learning system) is composed of input nodes that feed intohidden layers. Numerical values flow between the nodes through weighted connections (the model’sparameters), ultimately producing an output. It is a simple structure from which, thanks to its enormous scale (and the presence of non-linear activation functions within the nodes), extraordinarily complex behavior emerges. A Large Language Model (LLM) is a spe- cific type of Neural Network capable of associat- ing each word fragment (token) with a series of numbers (embeddings). The numerical values of these words are updated based on other words in the input (for example, through theTransformer architecture andattentionmechanisms [34]) to predict the most probable next word in a context- dependent manner. For example, when chatting with such a system, the most likely word follow- ing the end of a question will be the beginning of the response. An example: Consider the input The quick brown fox jumps over the lazy This sentence, used to display fonts because it contains every letter of the alphabet, should be completed withdog. A trained LLM will have read it thousands of times and will almost cer- tainly respond correctly. In reality, however, the model does not choose the most probable word: it samples randomly according to a probability dis- tribution across a range of possibilities; a distribu- tion regulated by a parameter calledtemperature [21]. With very low temperature, the behavior is deterministic (always selecting the most proba- ble word) but boring; by raising the temperature, other words gain the chance of being sampled:cat, bunny,?(Cf.Regola e casoby Bruno Munari). This margin of randomness, thissemantic dither- ing, can cause abutterfly effectcapable of driving interactions with LLMs in unpredictable direc- tions, and constitutes a sort ofintentionality gap. What would happen if the model chosecat instead ofdog? Word by word, seeking combina- tions statistically coherent with what has already been written, it could produce: The quick brown fox jumps over the lazy cat. I wrote "cat" because I was getting bored with this absolutely overused sentence. It seems as though the network possesses con- sciousness and a sense of humor, but in reality, 6 it has merely stumbled due to high temperature, continuing to generate statistically plausible text from what it had already produced. This is the behavior of a pure transformer, without the corrective systems and optimizations of modern models. But the underlying mecha- nism remains: most LLMs generate the next unit of information, feed everything it has generated back into the input, and repeat the process until the token indicating the end of the response is extracted. Everything is based on syntax: on the fre- quency with which certain words appear near others. However, it is a syntactic statistics so vast and profound that it creates nearly the illusion of semantics 2 . A neural network does not know that an apple can be red or green: it has only seen the words “red” and “green” near “apple” countless times, yet it possesses no concept of any of the three. So much has been said and written about the world that, by training net- works on this vast amount of material through sheer brute force, they can even solve problems that ostensibly require logic and semantics. Yet occasionally (more often than one might think), syntactically perfect, highly plausible, yet false sentences emerge. This is the phenomenon of hal- lucinations, which betray the purely statistical nature of these systems [38] (along with other related issues, such as poor training or inadequate attention windows). However, hallucinations can also be viewed as mutations: unexpected deviations, secret passages into territories that no human being would have deliberately explored. A generative system is, in potential, an extraordinarySerendipity Machine [39] (much like Balestrini’sTristano, or Brian Eno’sOblique Strategies[40], those cards used in recording studios to break creative deadlocks): even the statistical irregularity of a neural network (devoid of any intentionality) can open pathways that an artist would tend to dismissa priori. 2 Within the framework of distributional semantics [35,36], the question remains as to whether semantics itself is reducible to large-scale syntactic structure. A literary example is pro- vided by Roland Barthes in “The Death of the Author” [ 37], where he argues that writing is an impersonal process; one in which it is not the subject who speaks, but language itself expressing and articulating itself. Nevertheless, the core issue remains: an LLM lacks experiential knowledge of what it discusses. The condition is that someone must be ready to recognize their value. 3 The imagery of “statistical blenders” often used to describe these systems, while useful as a first approximation, runs the risk of being mis- leading and fails to capture their sophistication. It would, in fact, be more accurate to speak of statistical resonators. This distinction is techni- cally precise for architectures based on State- Space Models [41], whose mathematical definition essentially constitutes an adaptive filter bank 4 . As a text is read, resonances shift: the system tunes itself to certain linguistic structures. It does notblend, but ratherresonates. For Transformer architectures [34], the attention mechanism per- forms an analogous tuning, albeit without the same direct justification in terms of systems the- ory. The philosophical implications are significant: if the output were merely an interpolation, it would be a degraded copy of something already existing. Instead, it is something more specific: not a copy, nor an original, but the manifestation of the model’s dynamic behavior excited by that context, subject to all constraints regarding filter order, as well as the fundamental limitation of fil- ters:it cannot amplify a signal that does not exist. One might object that in a non-linear context, the system is nonetheless capable of generating signals not present in the training set; however, non- linearities can generate components not present in the input only as deterministic combinations of the incoming signals (harmonics, intermodula- tion products); the reachable space, however vast, remains a subset of the possible. 6 Computational Creativity: Possibilities and Structural Limits Generative models based on statistical learning tend to resonate with-, or gravitate toward-, the most heavily populated regions of their training space. The model is trained to produce the sta- tistically most representative output for a given 3 It should be noted that this holds true only within the realm of art. In medical, legal, informational, or scientific contexts, hallucinations remain flaws to be minimized. 4 The eigenvalues ofAare the poles.BandCdetermine the zeros.Adetermines the conceptual structures upon which the system resonates;Bdecides which modes are excited by the input;Cdecides which are observed to produce the output. 7 context [42]: when asking Suno for a rock track, it produces something that istypicallyrock; if asked for jazz, it delivers something that istypically jazz; if asked for industrial noise experimental, it will produce something that istypicallyindustrial noise experimental (because at this point, it isn’t even all thatexperimental; these things have been done for fifty years now). 5 This resonance with “typical” output is the default mode of Generative AI models, rather than an insurmountable boundary. Prompt engineering can be used to drive the model toward the less- traversed regions of its learning space, countering its tendency to respond with something reassur- ing instead of something that provokes aesthetic friction. The challenge for the artist becomes learning to navigate these low-density zones with awareness, and recognizing what occurs when the system begins to exhibit unexpected behaviors. In this regard, it may be useful to introduce a theoretical distinction derived from studies on computational creativity. Margaret Boden distin- guishes betweenCombinational Creativity(the ability to recombine known elements to create new instances),Exploratory Creativity(the ability to explore a given conceptual space, generating new instances while respecting its rules), andTrans- formational Creativity(the ability to break the very rules of the conceptual space, inventing cat- egories that did not previously exist) [45]. The latter is, at a higher level, a form of the second: exploring the space of possible conceptual spaces themselves [46]. AI tends to excel significantly in the former; it struggles more with the other two, which are structurally related. Though, the debate remains open: some recent models exhibit emer- gent behaviors that blur this boundary [47,48]. It is precisely where the model exhibits unexpected behaviors, straddling that very threshold, that the most interesting space for the artist resides. But how can a deterministic system exhibit unexpected behaviors? This question originates in 5 Anyone who feels threatened by generative AI music sys- tems because they are encroaching upon “their creative niche” is likely operating within a creative space that is already exten- sively represented; conversely, those who create novelty (at least for now) feel more secure. This reminds of the “head in the sand” Turing Objection [ 43], which in this context could be paraphrased as: The prospect of a machine being capable of creativity is so destabilizing that one finds it more reassur- ing to convince oneself that it cannot be. For further reading on the subject, see [ 44]. the notes of Ada Lovelace [49] and was later revis- ited by Turing [43], who, however, provided an answer using generic arguments. To bolster Tur- ing’s argument more robustly, one could argue that it is a matter of perspective: the point is not determinism itself, but rather how practi- cal it is to determine behavior within a useful timeframe. The emergent properties of a system are technically computable, yet the process is impractical (reminiscent of a well-known issue in cryptography, where the most secure encryption is that which cannot be broken within a useful timeframe). It is worth adding an observation regarding the nature of aesthetic value in this context. Human aesthetic reception is largely independent of the intentions (or lack thereof) of the entity that mate- rially generated a work [50]. Beauty emerges in the eye of the beholder, not necessarily within the “heart” of the creator [51,52]. In computa- tional aesthetics, what matters is not whether the machine experiences emotions, but whether it is capable of evoking them in the user. AI func- tions as a mirror to human expressive structures, trained to simulate their syntactic configurations; however, the machine’s lack of emotion precludes it from being able to properlyevaluateaccording to human criteria (a specific form of alignment dif- ficulty [53,54]). 6 As Italo Calvino emphasizes in Cybernetics and Ghosts[55], for syntactic gener- ation (thecybernetics) to transform into true art, human lived experience, choice, and responsibility (theghosts) remain irreplaceable [7]. It is precisely these responsibilities that define the artist’s role as anegentropic curatorand an explorerof latent spaces. From the perspective of information thermodynamics, the models’ latent space is a system of extremely high entropy. In this scenario, the artist ceases to be one who “cre- ates matter” out of nothing, instead assuming the role (in the case of thenegentropic curator) of a modern Maxwell’s Demon [56]: an entity that performs continuous selections at the cost of an expenditure of intelligence, separating the “sig- nal” (meaning, emotional value, novelty) from the statistical background “noise.” There is, how- ever, a subtle irony in all of this: for the neural 6 Even assuming a form of embodiment for the artificial cre- ative agent, such emotions would nonetheless not necessarily be attributable to human ones. 8 network, “noise” is precisely what the training process sought to eliminate: statistical deviations, anomalies, and hallucinations. These are the very elements that, as we have seen, constitute the most fertile material for the artist. The artist’s signal is the machine’s noise, and vice versa; the artistic Maxwell’s Demon works, so to speak, in a direc- tion exactly opposite to the system it attempts to govern [22]. To abdicate this critical judgment would inevitably lead to the “heat death” of art. 7 The Risks of Creative Automation The intensive use of generative tools exposes us to a highly insidious risk: a desensitization tocogni- tive effort. This is a specific manifestation of the cognitive atrophy induced by chatbots [57] (docu- mented more broadly asautomation complacency [58]); it does not involve a decline in faculties in the strict sense, but rather a progressive reduction in the tolerance for effort; a diminishing will- ingness to invest time and labor into a complex process [ 59] that is a necessary condition for any form of mastery. The musician generating music with AI, the writer composing with an LLM, or the programmer delegating code to an agent can easily achieve a result that lacks certain qualities [60–62], qualities that are difficult to recover with- out friction with the process. The problem is that this result arrives insecondsrather thandays, and the subject tends to settle for the available result [59]. Over time, what is missing may even cease to seem relevant [63]. These tools can and should be used, but astools, not asdelegations. The final word (selection, correction, critical com- pletion) must remain human; not as an ideological constraint, but because it is through friction with the process that the capacity to create (or even simply to judge) the result is maintained. There is another phenomenon closely linked to this risk that is worth noting: Model Collapse [64]. It occurs when an AI model is trained on synthetic outputs generated by other models (of which the Internet is becoming increasingly saturated [ 65] due to so-calledAI-Slop). The system ends up resampling and degrading existing entropy in a sort of statistical inbreeding, losing all the vari- ance that made it useful. The combined risk of Model Collapse and cognitive atrophy foreshad- ows a dystopian loop of progressive degeneration of creativity itself, both biological and artificial. The antidote has a name:environmental enrichment[66], borrowed from neurobiology and applied to the technological domain. For both its own well-being and that of the models, the human must be incentivized to actively inject three things into the system that, at present [7], cannot be automated: 1. extra-ordinary data, rare experiences, unprece- dented configurations that the machine has never processed; everything that cannot be rep- resented through the statistical distribution on which it was trained. 2. embodied experience, Calvino’sghosts, that which is rooted in the body and physical con- text, which no purely symbolic model can ever possess by definition [8,9]. 3. critical judgment, which is the only operation that transforms a statistical generation into something meaningful [42]. These are the things humanity must safeguard, yet they are also those at risk of being lost if we allow ourselves to be seduced by the convenience of the tool. Therefore, there is a call for poli- cies that promote the production of human data: human creativity must be regarded as a common good to be protected. Returning to the debate onTransformational Creativity, Model Collapse demonstrates that there is indeed a limit to what a model can do. If AI were capable of true novelty, synthetic data would enrich the space; instead, it erodes it. AI, however sophisticated, operates within a manifold [67] with degrees of freedom fixed by the training data: it can navigate it with extraordinary skill and extend it by pushing into its less frequented regions, butit cannot add a dimension that does not exist. That capacity (to invent not just a new point in space, but a new axis, by operating at a meta-level [7]) remains, at least for now, the prerogative of biological agents. 7 7 However, even assuming that a future AI might be archi- tecturally capable of producing new degrees of freedom, such elements would not necessarily be attributable to human expe- rience, unless the new architecture replicates the human one exactly(in both body and mind). For the modeling of the ner- vous systems of biological organisms, see the experiments in [ 68]. 9 8 Co-creativity and the Roles of Distributed Authorship The taxonomy of systems outlined in Sec. 4.1 frames the status of the generative system; the taxonomy proposed here, instead, addresseswho contributes creatively, and at which stage. To understand exactly where and how creativ- ity and knowledge are injected into a Generative Art system, it is necessary to decompose it into a pipeline of primary components and processes: 1. creation and selection of the data upon which the system is based (in terms of training or an explicit corpus used by the system), 2. design of the architecture that defines its behavior, 3. conditioning of the system through prompts (including multimodal ones) or other channels of interaction with the resulting system, 4. the set of manipulations, but also curatorial and editorial choices applied to the produced material duringpost-generation. What fuels each of these elements is always, albeit to varying degrees,human creativity, and within each, one can recognize a variation of the roles that the artist is called to assume within the ecosystem of Generative Art. 8 The idea that the artist’s role is transformed by generative systems is not new. As early as 1973, Cornock and Edmonds described the artist as“a catalyst of creative activity”within the context of computational art [74]. This concept was later revisited by Norouzi and Prinz through the notion ofcatalytic collaboration[6]: a collaboration in which the machine is neither transparent nor pre- dictable, and where the artist triggers processes that limit their direct control over the output (a limitation also expressed by [15]). However, the present contribution rests on two different premises: 1. the partner in the creative chain is not the algorithmic system (which, at present, remains medium, instrument, or artwork), but rather the collective of humans distributed along the Gener- ative AI pipeline. 2. the degree of control varies 8 Note that the proposed taxonomy focuses on the role of theartistthroughout the (potentially asynchronous) stages of the pipeline; it should therefore not be confused with similar taxonomies and classifications that focus on the role of the system[ 69], the interaction with it [70–72], or what it produces [ 73]. radically depending on the stage; this is not a generalized relinquishment, but rather forms of qualitatively distinct mastery, some of which (such as architectural design or post-production) involve fine-grained control thatgovernsthe algorithm. The proposed taxonomy distributes these inter- ventions across the four stages of the pipeline, assigning each its own name and specific creative responsibility. Entropic agentThe first and most influential vec- tor of creativity is thetraining corpus. The con- tents comprising a model’s training set (texts, images, scores, recordings) are largely the product of human beings, who have embedded their own worldview into shared culture. If every output res- onates with and depends upon the totality of that corpus, then those who contributed to that cor- pus are co-authors in a non-metaphorical sense. This constitutes the most significant injection of creativity that a Generative AI system receives, yet it is also the most invisible: those who created the content are rarely aware they are contribut- ing to the training of a machine. It is precisely here that the role of the artist as anentropic agent takes root. The richness, originality, and embodied experience from which the data originates are not marginal details: they determine the quality and breadth of the entire latent space available to the system. 9 An interesting corollary to this point is thatanyone, the moment they create something, becomes a potentialentropic agentin service of future works. Systems designerThedeveloperswhohave designed a model’s architecture have, in turn, crystallized within it a set of non-neutral technical and aesthetic choices. The structure of a model is the encoding of a point of view on the world; it is not merely a technical act. As previously argued, it can be considered a philosophical and cre- ative exercise. The model architect chooses which structures of the world to make representable, which relationships between concepts will become codifiable, and what form the space in which the system moves will take, along with its degrees 9 It is no coincidence that one of the most interesting trends in training personalized models is the use of small and highly curated datasets (even deliberatelybiased), in which a narrow and coherent corpus is used to steer a model’s behavior much more precisely and creatively than a vast, heterogeneous, and generalist dataset [ 75,76]. 10 of freedom. They are asystems designer. Out- side the context of Generative AI, this role may take on less computational contours (not all Generative Art is digital [15]), but they remain essentially identical in substance. ExplorerThen there is prompt engineering, often regarded as a second-order practice relative to the preceding stages. Yet, it involves its own form of mastery that should not be underestimated. Understanding a model (its space, its tenden- cies and blind spots, the emergent behaviors that characterize it) is a skill acquired through expe- rience, leading to the ability to push the system toward regions where the unexpected resides. The prompt engineer explores a vast and partially obscured territory, learning to recognize its topog- raphy and to utilize its most peculiar recesses. They are, properly speaking, anexplorerof latent spaces. The same applies to generative art sys- tems not based on AI: interactive multimedia installations that allow for user control can be experienced by the viewer as instruments through which to express intentionality, transforming the viewer into an element of the work or a co- author of it [77,78]. Conversely, works that do not involve forms of conditioning by the viewer concentrate the role of the explorer onto the sys- tems designer, who also designs a mechanism for automatic exploration. Negentropic curatorFinally, there is the phase ofpost-generation(as a generalization of “post- production”): not merely selection [79], retouch- ing, or editing, but the entire act offraming: the decision of what a work is, what it means, and how it is presented to the world. It is a fully realized creative act, known in the artistic field ascurator- ship(a principle also suggested by Goodfellow [5]). As previously observed in Sec.4.1, it is the agent governing the interface between the machine and the viewer who establishes the status of the system and its output. This governance is far from pas- sive; it consists of choices that constitute creative acts in the fullest sense. Whoever operates in this phase performs exactly the function of thenegen- tropic curator: filtering the entropy of generative output through selection, correcting and refining results during post-production, and transforming a statistical distribution (or the entire system) into new meaning through curatorship. But what happens when the artist does not manage the entire chain? Goodfellow proposes viewing the authorship of AI works as a spec- trum, ranging from exclusively human creation to completely algorithmic [5]; here, however, the tax- onomy of roles mapped onto the pipeline always refers to human authorship. 10 A generative system is the device through which, in fact, an asyn- chronous creative dialogue is realized with its designer and with everyone who has contributed (consciously or otherwise) to its creation. It is a collaboration that transcends spatio-temporal boundaries: when reread through the previously discussed examples, it reveals historical prece- dents of remarkable relevance. It is what Bach achieved by synthesizing the entire contrapuntal tradition inThe Art of Fugue; it is the logic of Kirnberger’s Musical Dice Game, where he wrote the fragments and defined the rules, providing a system for players to roll dice and evaluate the outcomes; it is what Moholy-Nagy attempted by dictating instructions over Meucci’s telephone to a worker unaware of the artistic purpose of their execution. In all these cases, collaboration occurred between individuals distant in space, time, and intention: the “co-author” was unaware of their role, yet they were performing it. Genera- tive AI industrializes and radicalizes this pattern: the invisible co-author is no longer a worker in an enamel factory, but a multitude of human beings (contributing through the mediation of a neu- ral network) whose presence in the creative chain remains, in most cases, entirely implicit. In this light, the myth of the solitary “romantic genius” (already challenged by Cage through his radical dissolution of the creative ego, as well as by Barthes, Kroeber, Foucault, and many others [24,37,80]) definitively gives way to a concept of distributed authorship. The work is no longer the fruit of a single individual, but rather the result of a chain of asynchronous collaboration that spans space and time, involving the original creators of the dataset, the architects of the model, those who formulate the prompt, and those who perform the final selection. Each of these stages incorporates 10 In this proposed view, Goodfellow’s gradient can be inter- preted as thedegree of delegationor thedegree of plurality. It could be interesting to measure the artist’s contribution and theperceived weightof each of the four stages of the pipeline, in order to quantify this gradation. 11 a distinct and human creative act; the creativ- ity of the output is an emergent property of the entire system (including humans). In this regard, the machine, in its current configuration, can only be a medium or a work in itself. 8.1 Ethical and Regulatory Implications This perspective, however, reveals an ethical issue that cannot be ignored. For collabora- tion to be legitimate, all parties involved must at least be aware of their involvement. Those who contributed training data have the right to be consulted, recognized, and compensated for this contribution [ 81]. If this does not occur, the model of distributed co-creativity fractures: what could have been a collaboration transforms into unilateral appropriation, further compounded by widespread economic exploitation. A creative chain that fails to recognize all its links is not a form of collaboration; it is a form of exploitation [ 81]. This is a real tension that has led to signif- icant litigation and driven a necessary rethinking of the current regulatory framework. On the regulatory front, several directions are emerging [82,83]: the introduction of manda- tory licensing for datasets used in training; the requirement to label generated content (“created with AI”); the extension of protections to voice, image, and personal style as forms of intellec- tual property; and copyright reform to address the large-scale generation of derivative content. This remains an ongoing process: the speed of technological progress has outpaced that of the regulatory response, as evidenced by the degree of impunity enjoyed (and, in part, continue to enjoy) by major industry players. In particular, the issue of liability for derivative content remains the least legally mature and the most ethically urgent. What is certain is that true distributed authorship requires, to be such, the explicit recognition of all its contributors. 9 The Aesthetics of Failure: A Reevaluation of Old Neural Models The Barron’s practice of “torturing circuits” (which their contemporaries did not consider “music”) is an example of the deliberate misuse (or tampering) of an instrument, driven by artistic inquiry and oriented toward anaesthetics of fail- ure[ 84]. In the recent history of Generative AI, these practices are explored by a very small niche relative to the rest of AI research [85]. Notable is Eug ́enie Desmedt’sThe Syntactic Synthesizer [86], an interactive installation that treats LLMs as analog instruments to be played (using physical knobs to control temperature, memory horizon, and semantic coherence). It is worth tracing a brief arc around this theme as well, because it is around this very concept that a potentially highly productive tension resides. In 2015, Google conducted an experiment to understand which features were captured by indi- vidual neurons in an artificial neural network trained to distinguish faces and other patterns [87]. The problem, as is always the case when working with neural networks, is understanding what is actually being learned; which representa- tions are consolidating within the hidden layers, in its latent space. The solution was to reverse the computational flow: rather than propagat- ing information from input toward output, it was propagated backward, reconstructing the pixel configuration that would maximize the activa- tion of a specific neuron. Neuron values thus became controllable parameters, knobs that could be adjusted to modify the characteristics of an input image. The result wasDeepDream: images in which the network, much like a computational pareidolia, amplifies everything it perceives as familiar, projecting its own internal representa- tions onto external forms. A system trained to accurately recognize the world was forced (ordis- tracted, to use Munari’s term [22]) to deform it according to its own internal logic. What makes DeepDream artistically interest- ing is not merely its visual output (which is particularly lysergic), but the nature of the ges- ture that produces it. Forcing a neural network to manifest its own internal structures rather than describe the world is, structurally, the same oper- ation performed byCircuit Benders, and before them, the Barrons, on electronic devices: one does not workwiththe tool, butagainstit, in the zone of friction between intended function and the behavior emerging from abuse. In this mode, latent space is conceived in a radically different way: it is not merely navigated, 12 but altered by operating at a meta-level, modi- fying its architecture to exponentially extend its possibilities. In certain respects, it is transformed into a sort of Library of Babel [88]: an immense archive, largely dark and random, in which works that the system could not have generated spon- taneously also exist in potentiality. The artist therefore assumes all roles simultaneously in this mode, including the traditional ones:craftsman of the object,entropic agent,explorer,systems designer, andnegentropic curator. The problem is that (for understandable rea- sons) research has systematically drifted away from this territory. Starting from aesthetically radicalmodels (imperfect, unpredictable, and original precisely in their imperfection), much effort has been directed toward achieving some- thing robust, realistic, andwell-tempered. Con- temporary generative models are optimized to resist error: even when provided with an arbitrary input sequence, they will nonetheless produce a coherent, plausible output. This robustness, pre- sented as progress, is also a loss: the model no longer produces anomalous outputs, no longer exhibits peculiar stereotypes, and no longer yields the unexpected. As technically extraordinary as it may be, it tends toward something we already know how to do: a photograph, a rendering, or a commissioned illustration. The constraint, para- doxically, stems not from the imperfection of the instrument, but from its perfection. The older models should not be thought of asworse, but asdifferent, characterized by a behavior that contemporary engineering seeks to avoid: the production of anomalies. Drawing upon the distinction betweencombinatorialand transformationalcreativity, these models prove paradoxically closer to the latter, not because they autonomously produce radical novelty, but because their specific mode of failure is a dimen- sion (a degree of freedom) within which the artist can operate, and one that was not present before their advent. 10 Noise, Power, and Aura In 1977, Jacques Attali proposed a radical the- sis: music does not reflect society, but precedes it. Whoever controls sound controls social order. Attali identifies four stages in the relationship between sound and power (Sacrificing,Represent- ing,Repeating,Composing) and describes moder- nity as the era in which music becomes a stan- dardized and accumulable commodity, envisioning within theComposingstage (in which anyone cre- ates for the pure pleasure of doing so) the only possible subversion of the commodified order. [ 89] In these pages, it has been shown how the tendency of modern AI to gravitate toward the center of the distribution, to excite its most rep- resented modes behind a promise ofComposition, has actually accelerated an ongoing process of cul- tural homogenization (the publishing world was already saturated withMuzakandHuman-Slop before generative models arrived), leading to what could be defined asAlgorithmic Repetition. How- ever, a clarification is urgently needed: the issue is notAI yes or AI no, nor is itold models vs. new models, but ratherhow and with what inten- tions they are used. The decline in diversity [60] is not an intrinsic property of language models, but a specific effect of alignment processes (such as RLHF), oftuning, that renders them safer and more predictable [61]; it is up to theexplorerto push them into contexts far from their zone of statistical comfort. Furthermore, it should be noted that the effects ofAlgorithmic Repetitionare also the causes of the Model Collapse discussed in Sec.7: both describe a system that resamples itself, reducing its own variance; the former on a cultural level, the latter on a statistical one. This also shifts the political issue raised by Attali onto a technical plane. Attali’s thesis also resonates with that of Wal- ter Benjamin [90]: both diagnose what occurs when a work of art enters the era of its technical reproducibility. For Attali, it is a matter of power; for Benjamin, it is the loss of aura (that dimen- sion of uniqueness that mass reproduction would dissolve). In a generative context, however, the picture becomes interestingly complicated: by def- inition,every generative output is unique(at least within the assumed operating regime). Accord- ing to Benjamin, therefore, the aura should not dissolve; rather, it condenses upon the generative system that produces the artifact [2]. Aura as a localized excitation emerging directly from the cul- tural fieldthat gave rise to the model. This shift in the aura requires a third term to describe the output of a generative system, one that transcends 13 the traditional opposition betweenoriginaland copy, which is proposed here asmanifestation. 11 For example: a unique object produced by a generative system (whether it be a drawing made by M ́eta-matics, a drawing created by Nees’s software, or a piece produced by an AI) is not anoriginalbecause infinite other equiva- lent instances exist, nor is it acopybecause it is not the reproduction of a pre-existing object: it is amanifestationof a process, analiasfor an aura residing elsewhere. Themanifestationis thus the new status of the artwork in the era of gen- erative production: a unique, unrepeatable object that is simultaneously non-individual. What can be endowed with the adjectiveoriginal, imbued with aura, is the process that generated it. Within this framework, it is useful to revisit the distinction proposed by Nelson Goodman [91] betweenautographicandallographicarts. An autographic work is one in which even a faith- ful replica remains a forgery: authenticity depends on the object having been produced by a specific hand at a particular moment (such as a painting). In contrast, an allographic work is identified by its adherence to a notation: any correct performance of a scoreisthe work itself, not a copy of it. Music, for example, is allographic: there are no forgeries of Bach; there are only performances that either conform to or deviate from what Bach wrote. Within this topology, a generative process can be defined as allographic: for instance, code is an executable notation, and its outputs are execu- tions. However, the analogy with the musical score is short-lived: the correct performances of a sym- phony are notationally equivalent, whereas the executions of a generative system produce objects that are unique each time 12 . Yet, generative out- put is neither autographic (lacking a production history that confers uniqueness) nor strictly allo- graphic (as it is not interchangeable with other outputs from the same system). It escapes both categories; for this reason, too, the termmanifes- tationis proposed as a third ontological status; the 11 Used here in its ontological sense (cf. Gabo, Sec.3), not in the sense popularized by contemporary “manifesting” culture. 12 Far more variable than any performance of written music could be. One might object that Terry Riley’sIn C[ 92] is a work composed of ever-changing performances; however, it can readily be classified as a generative work, and thus such performances are best described asmanifestationsof the work. manifestationof a process, endowed with singu- larity without autography, and with processuality lacking notational identity. It should be noted that, in the same essay, Attali speaks primarily ofnoise: for him, noise possesses no distinct technical or social signifi- cance; rather, it is inherently political, as estab- lished by the very definition provided by infor- mation theory. Attali demonstrates that what a society identifies as disturbance systematically coincides with those voices that threaten the cod- ified order: minorities, unaligned aesthetics, and aspects of reality intended for marginalization. Contemporary models, trained on mainstream corpora and optimized for statistical robustness, replicate this exact gesture: they amplify the cen- ter of the distribution and marginalize anything that deviates from it, acting, without explicit declaration, as apparatuses of power. It is therefore argued that older models (or, more broadly, those systems that manifest and exploit, more or less deliberately, the noise of a culture) are indeed bearers of an aura: they gen- erate authentic difference in the sense described by Attali, rendering the process recognizable, irreplaceable, andoriginal. In contrast, modern frontier models (with their forced robustness and pronounced convergence toward the center of dominant cultures, even if they do not generate copies) are themselves, at a higher level,copies of a cultural process (the technical reproduction of something that already exists). If culture itself is instead considered a generative process, then these models are itsmanifestation, just as the output is that of the model. Both interpreta- tions converge on the same conclusion: in its naive use, the frontier model introduces no inherent dif- ference relative to the cultural field from which it originates; it thus remains devoid of an aura distinguishable from that of the culture that pro- duced it. Thus,Algorithmic Repetitionoperates in a cascade across two levels: that of the model rel- ative to culture, and that of the output relative to the model. These models can only succeed in claiming an aura of their own if artists manage to do with them what the Barrons did with circuits, and what Munari theorized (perhaps provocatively) in theManifesto del Macchinismo: to use them as an object to be tortured, forced, and bent to one’s will, so that their own voiceandthat of the 14 machine, along with its hidden imperfections, may emerge. Therefore, if one wishes to leverage these new models in this manner, the artist must first act as anentropic agent, dismantling the system to recover entropy where homogenization has arti- ficially suppressed it, and subsequently act as a negentropic curatorto re-establish a new order. In this way, rather than being a device of power that amplifies the center of the distribution and per- petuatesAlgorithmic Repetition, the generative system becomes a device of liberation that ampli- fies marginalized voices, and the artist becomes an agent of social change operating through the manipulation of noise. 11 Conclusion In summary, drawing upon a historical genealogy and technological contextualization, this paper introduces: a taxonomy of functional categories for generative systems (medium,artwork,instru- ment), where attribution is aneditorial actrather than an ontological property; a taxonomy of four artist roles (entropic agent,systems designer, explorer, andnegentropic curator) based on the premise that the creative partner in the genera- tive chain is not the algorithmic system (which, for now, remains either medium, instrument, or artwork), but the collective of humans distributed throughout the pipeline, thereby establishing a framework of distributed authorship;environmen- tal enrichmentas a response to Model Collapse and cognitive atrophy;Algorithmic Repetitionas an aesthetic degeneration of aligned generative systems; andmanifestationas a third status tran- scending the original/copy dichotomy. Before concluding, from a perspective of ret- rospective positioning and comparison with the literature, it is worth verifying the robustness of the proposed contributions against Galanter’s nine problems for Generative Art, regardless of technology [ 93]. While these problems are rad- icalized rather than resolved by Generative AI, the work addresses at least seven of them. The problem of authorshipand theproblem of inten- tionare addressed through the taxonomy of roles and the concept of asynchronous co-authorship (Sec.8). Theproblem of uniqueness(the paradox of “unique objects produced in mass”) finds an ontological resolution in the concept ofmanifesta- tion, which names what Galanter left as an aporia (Sec.4.1and10). Theproblem of authenticityis addressed through the distinction between aes- thetic value and generative intention, in continuity with Calvino (Sec.6). Theproblem of dynamics (whether art resides in the process or the artifact) is shown to be an editorial rather than ontologi- cal distinction within the taxonomy of generative systems (Sec.4.1). Theproblem of locality, code, and malleability(where the work resides—in the code, the system, or the output—and who holds the power to define it) is addressed in Secs.3and 4.1and spans the entire contribution up to the concept ofmanifestationin Sec.10. Theproblem of creativityis developed through the distinction between different types of creativity applied to generative systems, and through the critique of Algorithmic Repetitionas its degeneration (Sec.6 and10). Finally, theproblem of postmodernity finds its answer not in the dissolution of the refer- ent (thesimulacrum[94]), but in its displacement from object singularity to system processuality (Sec.10). Finally, several social aspects must be reaf- firmed: First, risks like cognitive atrophy and Model Collapse can be mitigated throughenvi- ronmental enrichment,namely policies safeguard- ing and incentivizing human creative activities. Another aspect involves a formalization of the artist’s roles, which can be identified through- out the generative pipeline:entropic agentwithin the data,systems designerin the architec- ture,explorervia prompting or interaction, and negentropic curatorin post-generation. Third, the legitimacy of distributed authorship (an asynchronous, largely unconscious co-authorship) requires explicit recognition for all contributors, including those providing training data in the case of Generative AI. The regulatory dimen- sion, though moving in the directions discussed in Section8.1(mandatory licensing, labeling, protec- tion of voice and style, copyright reform), remains structurally lagging behind technological evolu- tion and constitutes the primary area of ongoing development. Finally, the last social aspect concerns the rela- tionship between generative models and power: these become instruments of liberation only when ceasing anAlgorithmic Repetition, that is, when 15 they stop gravitating around the center of a sta- tistical distribution that marginalizes noise and the unexpected (the native behavior of imperfect models). In this sense, the artist’s role is that of an entropic agentandexplorerwho dismantles the homogenization, and anegentropic curatorwho restores voice to marginalized noise. Methodological Note and AI Statement This article is a manifestation of the “Artifi- cial Intelligence for Music” course taught by the author at the University of Milan. Based on a talk transcript and a series of notes, the structure was organized into a coherent sequence through a clus- tering operation [95,96] based on embeddings [97] and subsequent optimal linearization [98]. Some of the resulting passages were merged and reworked using Claude Sonnet 4.6 (Anthropic) and Gemma 4 (Google DeepMind); the result was curated and rewritten for stylistic uniformity, semantic refinement, and bibliographic integration. 13 As for the models: behind ’Claude’ and ’Gemma’ lie vast, often uncredited teams, but above all, the most anonymous contributors: the authors of the corpus of millions of texts on which these models were trained. References [1] Snow, C.P.: The Two Cultures and the Scien- tific Revolution. Cambridge University Press, Cambridge (1959). Rede Lecture, 1959 [2] Fern ́andez-Castrillo,C.:TheAI work of art in the age of its co- creation. magaz ́en4(2), 357–384 (2023) https://doi.org/10.30687/mag/2724-3923/2023/02/008 [3] Park, S.: The work of art in the age of generative AI: aura, liberation, and democra- tization. AI & Society40, 1807–1816 (2024) https://doi.org/10.1007/s00146-024-01948-6 13 The process exemplifies the proposed taxonomy: the author asentropic agentin material selection,systems designerin clustering architecture,explorerin prompt formu- lation for language models, andnegentropic curatorin final rewriting. [4] Ambrosini, L.: The aura in the algo- rithm:ReimaginingWalterBenjamin’s aesthetic theory in the era of artificial intelligence. In: Nayak, B.S. (ed.) Dialec- ticofDigitalEnlightenment.Radical Theologies and Philosophies, p. 133– 160. Palgrave Macmillan, Cham (2025). https://doi.org/10.1007/978-3-031-95469-6 8 [5] Goodfellow, P.: The distributed authorship of art in the age of AI. Arts13(5), 149 (2024) https://doi.org/10.3390/arts13050149 [6] Norouzi, M.,Prinz, J.:From canvas tocode:lessonsfromgenera- tiveart.Synthese207,72(2026) https://doi.org/10.1007/s11229-026-05464-6 [7] Rohrmeier, M.:On creativity, music’s aicompleteness,andfourchallenges for artificial musical creativity. Trans- actionsoftheInternationalSociety for Music Information Retrieval (2022) https://doi.org/10.5334/tismir.104 [8] Damasio, A.: The Feeling of What Happens: Body and Emotion in the Making of Con- sciousness. Harcourt Brace, New York (1999) [9] Varela, F.J., Thompson, E., Rosch, E.: The Embodied Mind: Cognitive Science and Human Experience. MIT Press, Cambridge, MA (1991) [10] Anderson,P.W.:Moreisdifferent. Science177(4047),393–396(1972) https://doi.org/10.1126/science.177.4047.393 [11] Diderot, D., d’Alembert, J.l.R.: Ency- clop ́edie, Ou Dictionnaire Raisonn ́e des Sciences, des Arts et Des m ́etiers. Bri- asson, David, Le Breton, Durand, Paris (1751–1772) [12] Butcher, M.: How the Ancient Greeks InventedProgramming.Talkpresented at Strange Loop, St. Louis, MO; pub- lished on InfoQ, 29 December 2012 (2012). https://w.infoq.com/presentations/Philosophy-Programming/ 16 [13] Angius, N., Primiero, G., Turner, R.: The Philosophy of Computer Science. In: Zalta, E.N., Nodelman, U. (eds.) The Stanford Encyclopedia of Philosophy, Spring 2025 edn. Metaphysics Research Lab, Stanford Univer- sity, ??? (2025) [14] Knuth, D.E.: Computer programming as an art. Commun. ACM17(12), 667–673 (1974) https://doi.org/10.1145/361604.361612 [15] Galanter, P.: What is generative art? com- plexity theory as a context for art theory. In: In GA2003–6th Generative Art Conference (2003) [16] Gabo, N., Pevsner, A.: Realistic Manifesto, Mosca (1920) [17] Bach, J.S.: Die Kunst der Fuge (L’arte della fuga), BWV 1080 (1740–1750) [18] Kirnberger,J.P.:DerAllezeitFertige Menuetten-undPolonaisencomponist. Winter, Berlin (1757) [19] Moholy-Nagy, L.: Telephone Pictures (Kon- struktionen in Emaille / EM 1, EM 2, EM 3) (1922–1923) [20] Munari, B.: Verbale Scritto. Nugae, 25. Il Melangolo, Genova (1992) [21] Holtzman, A., Buys, J., Du, L., Forbes, M., Choi, Y.: The curious case of neural text degeneration. In: International Conference on Learning Representations (ICLR) (2020) [22] Munari, B.: Manifesto del Macchinismo vol. 10. M.A.C. (Movimento Arte Concreta), Milano (1952) [23] Cage, J.: Silence: Lectures and Writings. Wesleyan University Press, Middletown, CT (1961) [24] Kroeber, A.L.: The superorganic. American Anthropologist19(2), 163–213 (1917) [25] Tinguely, J.: M ́eta-matics (1959) [26] Barron, L., Barron, B.: Forbidden Planet (colonna sonora – “electronic tonalities”) (1956) [27] Hiller, L., Isaacson, L.: Illiac Suite (poi String Quartet No. 4) (1957) [28] Hiller, L., Isaacson, L.: Experimental Music: Composition with an Electronic Computer. McGraw-Hill, New York (1959) [29] Noll, A.M.: The digital computer as a cre- ative medium. IEEE Spectrum4(10), 89–95 (1967) [30] Balestrini, N.: Tristano. Feltrinelli, Milano (1966) [31] Balestrini, N.: Tristano. Romanzo Multiplo. DeriveApprodi, Roma (2007) [32] Eno, B.: Music for Airports / Ambient 1. EG Records (1978) [33] Munroe, R.: Machine Learning. xkcd, n. 1838 (2017). https://xkcd.com/1838/ [34] Vaswani, A., Shazeer, N., Parmar, N., Uszko- reit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems (NeurIPS), vol. 30 (2017) [35] Harris,Z.S.:Distributionalstruc- ture. WORD10(2–3), 146–162 (1954) https://doi.org/10.1080/00437956.1954.11659520 [36] Firth, J.R.: A synopsis of linguistic theory, 1930–1955. In: Studies in Linguistic Analysis, p. 1–32. Blackwell, Oxford (1957) [37] Barthes, R.: The Death of the Author (1967) [38] Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y., Madotto, A., Fung, P.: Survey of hallucina- tion in natural language generation. ACM Computing Surveys55(12), 1–38 (2023) https://doi.org/10.1145/3571730 [39] Green, D.: The Serendipity Machine: A Voy- age of Discovery Through the Unexpected 17 World of Computers. Allen & Unwin, Sydney (2000) [40] Eno, B., Schmidt, P.: Oblique Strategies (1975) [41] Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023) [42] Manovich, L.: AI Aesthetics. Strelka press Moscow, ??? (2018) [43] Turing, A.M.: Computing machinery and intelligence. Mind59(236), 433–460 (1950) [44] Audry, S.: Art in the Age of Machine Learn- ing. Mit Press, ??? (2021) [45] Boden, M.A.: The Creative Mind: Myths and Mechanisms. Weidenfeld and Nicolson, London (1990) [46] Wiggins, G.A.: A preliminary framework for description, analysis and comparison of creative systems. Knowledge-Based Systems 19(7), 449–458 (2006) [47] Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., Chi, E.H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., Fedus, W.: Emergent abilities of large language models. Transactions on Machine Learning Research (2022) [48] Schaeffer, R., Miranda, B., Koyejo, S.: Are emergent abilities of large language models a mirage? In: Advances in Neural Information Processing Systems (NeurIPS) (2023) [49] Lovelace, A.: Notes by the Translator. In: L.F. Menabrea, “Sketch of the Analytical Engine Invented by Charles Babbage”, Scien- tific Memoirs, vol. 3 (1843) [50] Wimsatt, W.K., Beardsley, M.C.: The inten- tional fallacy. The Sewanee Review54(3), 468–488 (1946) [51] Hume, D.: Of the standard of taste. In: Four Dissertations. A. Millar, London (1757) [52] Jauss, H.R.: Toward an Aesthetic of Recep- tion. University of Minnesota Press, Min- neapolis (1982) [53] Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforce- ment learning from human preferences. In: Advances in Neural Information Processing Systems (NeurIPS), vol. 30, p. 4299–4307 (2017) [54] Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. In: Advances in Neural Information Processing Systems (NeurIPS), vol. 36, p. 15903–15935 (2023) [55] Calvino, I.: Cibernetica e fantasmi (appunti sulla narrativa come processo combinatorio). In: Una Pietra Sopra. Discorsi di Letteratura e Politica. Einaudi, Torino (1980) [56] Maxwell, J.C.: Theory of Heat. Longmans, Green, and Co., London (1871) [57] Kosmyna, N., Hauptmann, E., Yuan, Y.T., Situ, J., Liao, X.-H., Beresnitzky, A.V., Braunstein, I., Maes, P.: Your Brain on Chat- GPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872, MIT Media Lab (2025) [58] Parasuraman, R., Manzey, D.H.: Com- placency and bias in human use of automation: An attentional integration. Human Factors52(3), 381–410 (2010) https://doi.org/10.1177/0018720810376055 [59] Risko,E.F.,Gilbert,S.J.:Cogni- tiveoffloading.TrendsinCognitive Sciences20(9),676–688(2016) https://doi.org/10.1016/j.tics.2016.07.002 [60] Anderson, B.R., Shah, J.H., Kreminski, M.: Homogenization effects of large language models on human creative ideation. In: Pro- ceedings of the 16th Conference on Creativ- ity and Cognition (C&C ’24), p. 413–425 (2024) [61] Padmakumar, V., He, H.: Does writing with 18 language models reduce content diversity? In: 12th International Conference on Learning Representations (ICLR) (2024) [62] Perry, N., Srivastava, M., Kumar, D., Boneh, D.: Do users write more insecure code with AI assistants? In: Proceedings of the 2023 ACM SIGSAC Conference on Computer and Com- munications Security (CCS ’23), p. 2785– 2799. ACM, Copenhagen, Denmark (2023). https://doi.org/10.1145/3576915.3623157 [63] Glickman, M., Sharot, T.: How human– AI feedback loops alter human perceptual, emotional and social judgements. Nature Human Behaviour9(2), 345–359 (2024) https://doi.org/10.1038/s41562-024-02077-2 [64] Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., Gal, Y.: Ai models collapse when trained on recursively generated data. Nature631, 755–759 (2024) https://doi.org/10.1038/s41586-024-07566-y [65] Liang, W., Izzo, Z., Zhang, Y., Lepp, H., Cao, H., Zhao, X., Chen, L., Ye, H., Liu, S., Huang, Z., McFarland, D.A., Zou, J.Y.: Monitoring AI-modified content at scale: A case study on the impact of ChatGPT on AI conference peer reviews. In: Proceed- ings of the 41st International Conference on Machine Learning (ICML). PMLR, vol. 235, p. 29575–29620 (2024) [66] Hebb, D.O.: The effects of early experience on problem solving at maturity. American Psychologist2, 306–307 (1947) [67] Bengio,Y.,Courville,A.,Vincent, P.:Representationlearning:Areview andnewperspectives.IEEETransac- tions on Pattern Analysis and Machine Intelligence35(8),1798–1828(2013) https://doi.org/10.1109/TPAMI.2013.50 [68] Wang-Chen, S., Stimpfling, V.A., Lam, T.K.C., ̈ Ozdil, P.G., Genoud, L., Hur- tak,F.,Ramdya,P.:Neuromechfly v2:Simulatingembodiedsensorimo- tor control in adult drosophila. Nature Methods21(12),2353–2362(2024) https://doi.org/10.1038/s41592-024-02497-y [69] Lubart, T.: How can computers be partners in the creative process: classification and com- mentary on the special issue. International journal of human-computer studies63(4-5), 365–369 (2005) [70] Kantosalo, A., Toivonen, H.: Modes for cre- ative human-computer collaboration: Alter- nating and task-divided co-creativity. In: Pro- ceedings of the Seventh International Confer- ence on Computational Creativity, p. 77–84 (2016) [71] Guzdial, M., Riedl, M.: An Interaction Framework for Studying Co-Creative AI (2019). https://arxiv.org/abs/1903.09709 [72] Horvitz, E.: Principles of mixed-initiative user interfaces. In: Proceedings of the SIGCHIConferenceonHumanFac- tors in Computing Systems. CHI ’99, p. 159–166. Association for Computing Machinery, New York, NY, USA (1999). https://doi.org/10.1145/302979.303030. https://doi.org/10.1145/302979.303030 [73] Colton, S., Charnley, J.W., Pease, A.: Com- putational creativity theory: The face and idea descriptive models. In: ICCC, p. 90–95 (2011). Mexico City [74] Cornock, S., Edmonds, E.: The creative pro- cess where the artist is amplified or super- seded by the computer. Leonardo6(1), 11–16 (1973)https://doi.org/10.2307/1572505 [75] Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personal- izing text-to-image generation using textual inversion. In: 11th International Conference on Learning Representations (ICLR) (2023) [76] Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer 19 Vision and Pattern Recognition (CVPR), p. 22500–22510 (2023) [77] Eco, U.: Opera Aperta: Forma e Indetermi- nazione Nelle Poetiche Contemporanee. Bom- piani, Milano (1962) [78] Ascott, R.: Is there love in the telematic embrace? Art Journal49(3), 241–247 (1990) https://doi.org/10.2307/777114 [79] Colton, S., Wiggins, G.A.: Computational creativity: The final frontier? In: ECAI 2012 – 20th European Conference on Artificial Intelligence, p. 21–26. IOS Press, ??? (2012) [80] Foucault, M.: Qu’est-ce qu’un auteur? Bul- letin de la Soci ́et ́e fran ̧caise de philosophie 63(3), 73–104 (1969) [81] Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J.W., Wallach, H., Daum ́e I, H., Crawford, K.: Datasheets for datasets. Communications of the ACM64(12), 86–92 (2021) https://doi.org/10.1145/3458723 [82] Samuelson, P.: Generative ai meets copyright. Science381(6654), 158–161 (2023) [83] European Parliament and Council of the European Union: Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artifi- cial Intelligence (Artificial Intelligence Act) (2024) [84] Menkman, R.: The Glitch Moment (um) vol. 4. Institute of Network Cultures, ??? (2011) [85] Sivertsen, C., Salimbeni, G., Løvlie, A.S., Benford, S., Zhu, J.: Machine learning pro- cesses as sources of ambiguity: Insights from AI art. In: Proceedings of the CHI Conference on Human Factors in Computing Systems. CHI ’24, p. 1– 14. ACM, New York, NY, USA (2024). https://doi.org/10.1145/3613904.3642855 [86] Desmedt, E.: The syntactic synthesizer. Interactions33(4), 14 (2026) [87] Mordvintsev, A., Olah, C., Tyka, M.: Inceptionism: Going Deeper into Neural Networks. Google Research Blog (2015). https://blog.research.google/2015/06/inceptionism-going-deep [88] Borges, J.L.: La biblioteca de babel. In: Fic- ciones. Sur, Buenos Aires (1944) [89] Attali, J.: Bruits: Essai sur L’ ́economie Poli- tique de la Musique. Presses Universitaires de France, Paris (1977) [90] Benjamin, W.: Das Kunstwerk im Zeitalter seiner technischen Reproduzierbarkeit (1935) [91] Goodman, N.: Languages of Art: An Approach to a Theory of Symbols. Bobbs- Merrill, Indianapolis (1968) [92] Riley, T.: In C (1964) [93] Galanter, P.: Artificial intelligence and prob- lems in generative art theory. In: Proceedings of EVA London 2019 (2019). BCS Learning & Development [94] Baudrillard, J.: Simulacres et Simulation. Galil ́e, Paris (1981) [95] Ward,J.H.:Hierarchicalgrouping tooptimizeanobjectivefunction. JournaloftheAmericanStatistical Association58(301),236–244(1963) https://doi.org/10.1080/01621459.1963.10500845 [96] Murtagh,F.,Legendre,P.:Ward’s hierarchicalagglomerativeclustering method:Whichalgorithmsimple- mentWard’scriterion?Journalof Classification31(3),274–295(2014) https://doi.org/10.1007/s00357-014-9161-z [97] Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D., Liu, Z.: M3-Embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distilla- tion. In: Findings of the Association for Computational Linguistics: ACL 2024, p. 2318–2335. Association for Computational Linguistics, Bangkok, Thailand (2024) [98] Bar-Joseph, Z., Gifford, D.K., Jaakkola, 20 T.S.:Fastoptimalleafordering forhierarchicalclustering.Bioinfor- matics17(suppl1),22–29(2001) https://doi.org/10.1093/bioinformatics/17.suppl 1.S22 21

Instant research discovery

The Aura in the Machine: Genealogy and the Status of the Work of Art in the Generative Era

Intelligence

Summary

Entities (13)

Relation Signals (12)

Abstract

Tags

Links

PDF

Full Text

Paper	Published	Area	Tags	Intel	Citations