Creativity After Generative AI

Who Owns Creation in the Age of Machines?

In September 2022 a piece titled Théâtre D’opéra Spatial, produced with the Midjourney text-to-image system and submitted by Jason Allen, won the digital arts category at the Colorado State Fair. The controversy that followed concerned attribution and fairness to other entrants. It did not concern a question that the episode raises more directly than the dispute over the prize: what exactly had been produced, and by what process, such that the category “creativity” was even in play.

That question is philosophical before it is legal. A separate inquiry, addressed elsewhere in this project’s published work, asks who should hold rights over an output generated this way, what counts as infringement of the material a system was trained on, and how existing statutes should be read or amended to cover the case. Those are questions about entitlement, and they presuppose an answer to a prior question that they do not themselves settle: whether what has been produced is the kind of thing creativity has historically named. This essay addresses the prior question.

An invention with a short history

The idea that creativity is a distinct human faculty, and specifically that original artistic production expresses something located inside the individual who makes it, is often treated as timeless. It is not. Edward Young’s Conjectures on Original Composition, published in 1759, is usually credited with the modern separation of “original” from “imitative” genius, arguing that the true original writer draws from a wellspring within rather than from established models. The distinction had limited purchase before this point. Classical and medieval aesthetics generally treated artistic skill as the correct application of inherited technique, and imitation of admired predecessors was a virtue rather than a defect.

Immanuel Kant gave the emerging idea its most rigorous form in the Critique of Judgment, published in 1790. Genius, for Kant, is the innate mental disposition through which nature gives the rule to art: the genius does not follow rules that could be taught, and cannot explain how the resulting work was produced, because the work follows from a talent that operates outside the genius’s own conscious control. Kant’s account already contains an instability worth noting before any machine enters the picture. The genius, on this description, does not fully author the work in the sense of consciously designing it. Something operates through the genius that the genius cannot inspect. The Romantic figure of the sovereign creative individual was constructed, at its philosophical origin, around a process that the individual does not transparently control.

The twentieth century then dismantled the figure from the other direction. Roland Barthes’s 1967 essay “La mort de l’auteur” argued that a text is a tissue of citations drawn from countless centres of culture, and that treating the author as the text’s origin and final authority closes off the plurality of readings the text actually supports. Michel Foucault’s 1969 lecture “Qu’est-ce qu’un auteur?” reframed the author not as a person but as a function, a set of operations a culture performs on certain texts, historically variable and tied to specific legal and institutional arrangements, including the copyright regimes that emerged from the eighteenth century onward to assign liability and ownership. Neither Barthes nor Foucault was writing about machines. Both were arguing that even in the paradigm case, a human being writing alone, the notion of the author as sole originating source was doing less philosophical work than it appeared to.

The point of this history is not that creativity was always an illusion. It is that the concept under strain today was itself assembled at a specific moment, for specific purposes, and had already been shown to be composite before generative systems existed to complicate it further.

Separating what the concept had bundled

The most useful analytical tool available for this question is not philosophical in the traditional sense but cognitive-scientific. Margaret Boden’s The Creative Mind, first published in 1990, proposed that creativity is not one operation but three, distinguished by the kind of novelty each produces relative to a conceptual space.

Combinational creativity produces novelty by generating unfamiliar combinations of familiar ideas, as in a metaphor that yokes together two established concepts in a way not previously stated. Exploratory creativity produces novelty by traversing a conceptual space defined by a set of stylistic or structural rules, generating outputs that are new instances of an established style, as a jazz musician improvising within the harmonic conventions of a given idiom generates phrases never played before without leaving the idiom. Transformational creativity produces novelty by altering the rules that define the space itself, producing outputs that were not merely absent from the space but impossible within its prior definition, as when a new set of compositional constraints makes a previously unthinkable structure available for the first time.

The distinction matters here because it does not require a verdict on intention, consciousness or authorship to be applied. It asks only what relation an output bears to the space of possibilities that preceded it, which is a structural question a system’s outputs can be assessed against directly.

Generative systems are demonstrably capable of the first kind. A diffusion model, trained to reverse a process of progressive noising applied to images during training, produces at inference an image by iteratively removing noise from a random starting point under the guidance of a learned distribution, and the images it produces routinely combine visual elements, styles and motifs that did not co-occur in any single training example. This is combinational creativity by Boden’s definition, and denying that it occurs requires denying the definition rather than the evidence.

The second kind, exploratory creativity, is also demonstrably present, and it is where the bulk of current generative output falls. A system fine-tuned or prompted to work within a defined style explores that style’s space extensively, producing instances that satisfy its constraints in configurations a human working within the same style might never have reached, precisely because the system is not limited by the sequential, effortful search a human artist performs.

The third kind is where the evidence runs out rather than where it turns negative. Transformational creativity requires recognising the current constraints of a space as constraints, understanding what they exclude, and having a reason to relax or replace them. Nothing in the training objective of a generative system represents its own operating constraints as constraints in this sense, available for deliberate revision toward an end the system holds. When a generative system produces something that reads as a genuine departure from its training distribution rather than a further interpolation within it, the available evidence indicates this is better explained as a low-probability sample, a curation choice by the human selecting among outputs, or a deliberate change to training data or architecture made by engineers, than as the system transforming its own space. The claim is not that transformational creativity by generative systems is impossible in principle. It is that no clear demonstrated instance separates a system’s own transformation of its constraints from a human operator’s transformation of the system.

Where intention actually sat

W.K. Wimsatt and Monroe Beardsley’s 1946 essay “The Intentional Fallacy” argued, within literary criticism, that an author’s intended meaning is neither available nor authoritative as a standard for interpreting a text: what matters for criticism is the poem as a public artefact, which must be judged by what is actually achieved in it rather than by what the author privately meant to achieve. The argument was controversial within literary theory and remains contested. It is worth recalling here because it already established, well before any generative system existed, that a documented gap between an artefact’s meaning and its maker’s private mental state was not a novel disruption but an ordinary condition of interpretation, one criticism had already learned to work with.

What generative systems add is not a gap between intention and meaning. It is a redistribution of intention across positions that used to be occupied by one party. Consider what actually happens when a system produces an image from a text prompt. The person writing the prompt has an intention, typically underspecified relative to the output, expressed in language rather than in the visual choices the output will embody. The system executing the prompt has no intention in any sense continuous with the human case: it computes a conditional distribution and samples from it. The engineers who built the training objective and curated the training data had intentions, exercised months or years earlier, embedded now in weights that no longer represent them as intentions but as statistical regularities. The human raters whose feedback was used to fine-tune the system’s outputs toward preferred results, in the reinforcement learning from human feedback procedures common to systems of this kind, had intentions about what counts as a good output in general, applied to material they never anticipated in its particulars. And the artists whose work appeared, credited or not, in the training corpus had intentions embodied in the original images the system’s internal representations were derived from, at a distance the system’s own operation does not preserve or track.

None of these parties holds the intention that the Romantic model of authorship requires: a single mind, present at the moment of production, whose purpose the resulting artefact expresses. What exists instead is intention distributed across five positions, active at different times, none of which individually resembles the authorial intention the concept was built around. The question “who intended this work” does not have a false answer. It has a decomposed one, and the decomposition is the actual state of affairs rather than a problem to be solved before the state of affairs can be described.

Aesthetic judgment without a maker

There is a resource within Kant’s own aesthetics, largely separate from his account of genius, that bears on whether the absence of a unified intending author disqualifies an output from aesthetic value at all.

Kant’s analysis of the judgment of taste, earlier in the third Critique, holds that we judge something beautiful when it exhibits purposiveness without a purpose, meaning it appears as if designed for a function while no determinate function or concept actually governs it. Kant’s primary examples were natural: a wildflower, a bird’s plumage, a crystalline rock formation. None of these has a maker whose intention could be consulted, and Kant’s account of the judgment of taste does not require one. What it requires is a form that engages the free play of the viewer’s cognitive faculties, imagination and understanding, without being pinned to a concept the object was made to instantiate.

This account was developed to explain the aesthetic appreciation of nature, prior to and independent of any question about art. Its availability matters because it shows that a coherent, historically central aesthetic theory already permitted objects with no intending maker to be judged beautiful, on grounds entirely internal to the perceiving subject’s response to the object’s form. A generative system’s output, assessed under this framework, is closer in structural position to a Kantian natural object, something that exhibits apparent purposiveness without a determinate authorial purpose behind it, than it is to the paradigm case of intentional artistic production the Romantic account was built around. This does not settle whether such outputs should count as art in the institutional or legal senses the companion volume examines. It indicates that their aesthetic evaluability does not depend on resolving the authorship question first, because at least one major tradition of aesthetic theory never made authorial intention the ground of aesthetic judgment to begin with.

The institution decides what counts

George Dickie’s institutional theory of art, developed across the 1970s and stated in its mature form in Art and the Aesthetic: An Institutional Analysis, published in 1974, offers a different route to the same separation. Dickie argued that an object becomes a work of art not by possessing some intrinsic aesthetic property but by having a certain status conferred on it by the artworld, the loose network of artists, critics, curators, dealers and institutions whose collective practice determines what counts. Marcel Duchamp’s Fountain, a urinal submitted to an exhibition in 1917 under the pseudonym R. Mutt, is the theory’s paradigm case: nothing about the object’s physical properties changed when it was designated art, and what changed was its position within an institutional practice.

Sherrie Levine’s 1981 series After Walker Evans, rephotographing Evans’s Depression-era images without alteration and exhibiting the results as her own work, and Richard Prince’s rephotography of advertising imagery beginning in the same period, extended the same institutional logic to direct appropriation: the artworld conferred art status on the act of recontextualising an existing image, independent of any transformation applied to the image itself. The institutional theory therefore already had, decades before generative systems existed, a working account of how minimal or absent transformation of source material could still result in something the relevant institutions treated as a new work, provided the institutional context of presentation did the conferring.

Applied to generative output, the institutional theory predicts something the Romantic theory of genius cannot easily accommodate: that whether a given generated image counts as art will be settled by gallery placement, critical reception, market uptake and curatorial decision, not by resolving in advance whether the generating process contained anything answering to intention. This is a prediction about how the concept will in practice be applied, not an endorsement of the outcome. Institutional conferral has historically tracked economic and social interests as much as considered aesthetic judgment, and there is no reason to expect the pattern to change now that the objects being conferred upon are machine-generated.

What remains of the concept

Creativity, examined this way, turns out to have named at least three things that generative systems separate rather than one thing they either possess or lack. There is a generative operation, the production of candidates within or against a conceptual space, which Boden’s taxonomy shows current systems perform well for two of her three categories and unreliably, if at all, for the third. There is an intentional structure, historically assumed to be unified in a single author, which current systems distribute across designers, curators, raters and prompters in a configuration the Romantic model was never built to describe. And there is an evaluative and institutional apparatus, the artworld in Dickie’s sense, which confers status on outputs according to its own criteria and has done so before with objects, Duchamp’s and Levine’s among them, whose production involved minimal transformation of existing material.

None of the three requires the other two to be present in order to operate, and the fact that they used to arrive together, bundled inside a single human act of making, was a contingent feature of how art was produced before automated generation existed, not a conceptual necessity the word “creativity” enforces.

The question posed at the outset, what creativity is once production can be separated from human intention, has an answer that follows directly from this: it is whatever remains identifiable and assessable in each of the three components once they are no longer guaranteed to travel together, and the interesting empirical and critical work still to be done consists in tracking, case by case, which components a given generative practice actually exhibits, and which it merely appears to by borrowing the name.