In every work that feeds on data there is a threshold where only one of two things can happen: either the data remain a tool, or they become the spectacle. Hyperallergic, reviewing Dataland, the AI art museum just opened by Refik Anadol in Los Angeles, pinpointed that threshold with surgical elegance: there, it writes, “data itself is the star of the show, an end rather than a tool.” It is the most useful sentence uttered about generative art in recent months, because it names precisely the fault line separating two things we keep mistakenly calling by the same name. On one side there is the data-spectacle: the floating cloud of pixels, the pulsing surface, the aesthetics of computational power displayed for its own sake. On the other there is what I have long called the Work-as-System. The difference is structural, and it is rooted in the architecture of the work well before its appearance. Defending generative art today means defending this distinction before the market and the public erase it by confusing direction with decoration.
The misunderstanding about generative art that makes galleries pull the brake
The problem is first of all one of vocabulary. The Artsy AI Survey 2026 captures it mercilessly: out of more than three hundred galleries, only 9% consider AI-generated art a legitimate medium, while 25% read it as a destabilizing force for authorship and value. It would be convenient to file the figure away as conservatism. But that 25% has sensed something true: if the work coincides with the output, then authorship really is fragile, because the output of a generative model is in principle infinite, reproducible, anonymous. The distrust stems from a misunderstanding that we ourselves feed every time we present the image as the work.
The answer lies in the Work-as-System
The Work-as-System answers by overturning the plane. The work coincides with the generative grammar that produces the image: the sources, the thresholds, the constraints, the dataset, the editing, the display, the direction that connects image, sound and language. The identity of the work is established in the framework. My authorial gesture lies in building the system that decides what becomes image and what stays out; I leave the choice of the best frame to the system itself. Every single image, every soundscape, every video performance remains a provisional matrix of that system.
An artistic lineage that comes from far away
It is a position far older than AI. When Allan Kaprow stages his Happenings in the late 1950s, he moves the work from the object to the event: what matters is the score of actions, the protocol that makes them possible. It is already, in embryo, the idea that the work coincides with its rules of occurrence. Process Art, later codified by the Tate among others, takes a further step and makes the work coincide with the process itself. And Roy Ascott, with telematic art, definitively dissolves the idea of an author who fabricates an object, replacing it with that of an artist who designs a distributed system of relations, where meaning emerges from exchange more than from the hand. I have already written elsewhere, in these same pages, about how Eco’s Open Work and Burnham’s Systems Esthetics anticipate this trajectory, and I will not repeat myself. What interests me here is a different point: generative art carries with it a precise genealogy, and it is the legitimate heir of a line that, from Kaprow to Ascott, stopped identifying the work with the thing more than half a century ago. AI inherited this idea and gave it the means to be fully realized, because a generative system is literally a grammar that produces, a machine of possibilities more than an object that stands.
The two sublimes of generative art
Here lies the dividing line that the data-spectacle hides. In e-flux, Neda Atanasoski and Kalindi Vora recently described the “AURA” of AI as an aesthetics of the ubiquitous sublime, a regimented automation that captures human intimacy through emotional response. It is a sharp diagnosis, and it perfectly describes the sublime of Dataland: the sublime of computational capital, which exists to impress. There is, however, another possible sublime, one that arises from the weight of what the machine is called upon to process. The first seduces. The second implicates. The difference, once again, lies in the system that generates the output, and in what that system chooses to take seriously.
Defining the boundaries of generative art
This is where the Work-as-System stops being a statement of intent and becomes a verifiable practice. I work through protocols, prompts, constraints, datasets, thresholds, editing, display, as if they were a musical score. These are the real objects of my work, even before the images that result from them. It is in the choice of a threshold, in the decision to cite a source instead of blending it away, in the signed grammar that binds image, sound and word, that the whole difference is played out between a system that processes the world and a system that merely performs. So the question to ask in front of a generative work is just one: where is the framework? Who defined its rules, its sources, its limits? What prevents that work, structurally and beyond good intentions, from sliding into data-spectacle? If the work can answer these questions by pointing to a precise place in its own system, then we are facing a Work-as-System. When instead the answer comes down to the pulsing surface, the sublime has already won, and we have been sold decoration passed off as thought.
Taking a position in the field of generative art
The field is consolidating fast. At Art Basel the Zero 10 section curated by Trevor Paglen brought into the institutional spotlight artists working with AI and computational systems around the theme “The Condition”; in Rome a group show of eighteen hundred names turned itself into a work-as-system in social form, contained in a QR code. With this acceleration the risk grows that everything will be flattened under the undifferentiated label of “AI art.” Defending generative art as Work-as-System means defending a threshold more than a technology: the one where data remain a tool, the one where the work remains its grammar before its surface. An open question remains, and I keep it as working material rather than as closing rhetoric: does that point of resistance really exist inside the system, or is it the sublime itself, once switched on, that makes it impossible?
Dario Buratti