Putting AI to Work May 21, 2026 · 14 min read

Pure-AI creation as its own discipline

The cluster closes with the hardest configuration. In pure-AI creation, the artist has stopped composing or producing material themselves; the model produces what is shown. The artist's practice is now selection, direction, system design, dataset curation, prompting, training, and presentation. The work is the artist's intellectual position made operational through the model — not the artist's composition or material craft. This is the configuration most exposed to the 'what makes this art at all' critique, and the configuration where the answer to that critique requires the most precise art-historical grounding. It is also the configuration with the longest and most decorated genealogy in twentieth-century art.

by Airtistic.ai editorial team

Through the lens of artistcreatorpatrongallerycritic craftindustrymarket

The four-configuration arc of this cluster closes here. The previous three articles worked configurations in which the human artist was clearly the maker of the finished object: the tool configuration used AI for bounded preliminary purposes; the assistant configuration used AI for labor across the workflow; the augmented configuration used AI for material in the finished work. In all three, the artist’s hand or eye was visibly making the work.

In pure-AI creation, that changes. The model produces what is shown. The artist’s practice has shifted from making to designing, curating, selecting, directing, prompting, training, and presenting. The work is the artist’s intellectual position made operational through the model — not the artist’s composition or material craft.

This is the configuration most exposed to the “what makes this art at all” critique. It is also the configuration with the longest and most decorated genealogy in twentieth-century art, and the configuration where the answer to that critique is best supplied not by argument but by pointing at fifty years of institutionally-recognized practice that operates on exactly this principle. The pure-AI configuration is not a new artistic territory. It is the latest chapter of a tradition that has been in major museum collections for half a century. This article works what that tradition is, what distinguishes serious practice within it from incidental output, and what the working conditions of operating in this configuration are.

What the pure-AI configuration looks like

The defining property of pure-AI creation is that the artist is not the producer of the work’s material substance. The model produces what is shown. The artist’s labor goes elsewhere — into the design and curation of the training data, the construction or selection of the model, the prompting methodology, the selection of outputs, the curatorial framing, the presentation. Five layers of artistic decision-making, all of them real labor, none of them composition or material making.

Five working practices make the configuration concrete:

Dataset-as-work practice. The artist constructs a dataset that constitutes the substance of the practice. The model trained on that dataset produces output that is continuous with the dataset’s character. Anna Ridler’s hand-photographed tulip dataset for Mosaic Virus (2018-2019) is the cited example. The dataset is, in Ridler’s framing, the work; the GAN-generated tulips are its evidence.

Long-form system-as-work practice. The artist builds and refines a generative system over years or decades, exhibits the system’s outputs as evidence of the system, and treats the system itself as the artistic object. Harold Cohen’s AARON, developed continuously from 1973 to 2016, is the canonical example. Cohen was explicit throughout his career that the system was the work he was making; the individual drawings were what the system produced under his guidance.

Selection-from-generation practice. The artist generates many outputs and selects a small number from them according to a clear curatorial position. The selection is the artistic act. The artist’s voice is identifiable in what gets selected and what gets discarded. This is closest to photography in structure — the camera produces images, the photographer selects which ones are work.

Model-as-work practice. The artist constructs or fine-tunes a model that is itself the artistic offering. Holly Herndon’s Holly+ is the cited example — the trained voice model, licensed for use by others, is the substance of the practice. The work is the model.

Architecture-and-presentation practice. The artist designs a generative installation as an architectural and curatorial whole. The model’s output is the visible surface; the work is the integration of dataset, model, display, space, and context. Refik Anadol’s Unsupervised (MoMA acquisition, 2022) is the cited example.

In all five practices, the structural pattern is the same: real, sustained, intellectually demanding artistic labor is happening, but it is not the labor of composition or material making. It is the labor of design, curation, selection, and direction. The work is what the model produces under the artist’s intellectual direction; the direction is what the artist made.

Why this configuration is defensible — the longer lineage

The art-historical lineage that supports pure-AI creation runs through two converging traditions, both well-established and both in major museum collections for decades.

The conceptual-art tradition. Sol LeWitt’s wall drawings from 1968 onward established that an artistic work could be a certificate of instructions rather than an executed object. The wall drawings have been executed by hundreds of different people following LeWitt’s instructions; the work is the instruction-and-conceptual-position. Lawrence Weiner’s text-based statement pieces operated identically — the language was the work. Yoko Ono’s Grapefruit (1964) was a book of instructions that constituted the work. By the early 1970s, the conceptual-art movement had firmly established within institutional art that the made object was not the only locus of artistic value, and that the conceptual gesture, the instruction, the system, and the curatorial position could each be the work. Lucy Lippard’s chronicle, Six Years: The Dematerialization of the Art Object 1966-1972, documented this shift contemporaneously.

The generative-and-algorithmic-art tradition. Vera Molnár began making algorithmic drawings in the late 1960s and continued for seven decades, culminating in a 2024 Centre Pompidou retrospective at age 100. Manfred Mohr has been making algorithmic plotter art continuously since 1969, with major museum exhibitions across that span. Harold Cohen developed AARON from 1973 until his death in 2016 — over forty years of continuous practice — and exhibited AARON’s output at the Tate, the San Francisco Museum of Modern Art, and the Brooklyn Museum starting in the late 1970s. Roman Verostko, Frieder Nake, and Charles Csuri were doing parallel work from the late 1960s onward. The generative-art tradition has been institutionally recognized since the 1970s; it predates personal computers, the internet, and machine learning.

These two traditions — the conceptual and the generative — converge directly on the pure-AI configuration. The artist’s work is the system, the instructions, the conceptual position, the curatorial frame; the output is the evidence of the work. This is exactly what pure-AI practice does with current tools. The configuration is not new; the tools are. The artistic principle that the system or the conceptual position can be the work is sixty years old and firmly in the canon.

What distinguishes serious practice from incidental output

The hardest practical question this configuration raises is what separates serious pure-AI artistic practice from someone who happened to generate a striking image once and slapped a name on it. The answer is the same one that has always distinguished serious artistic practice in any medium, just applied with extra rigor because the model’s default output is generically competent enough to mask the difference.

Four criteria, drawing on Carlos’s commentary:

Sustained body of work over time. A single piece does not make a practice. The artist whose pure-AI work amounts to a folder of one-off generations is not yet operating a discipline. The artist whose work is two years or more of sustained production with identifiable evolution and refinement is.

Intellectual position visible across the body. Multiple pieces should be readable as belonging to a single artistic project — a question being asked, a position being developed, a coherent investigation. The artist’s intellectual stance should be visible in what is made even without the artist explaining it.

Repeated, identifiable artistic decisions. Specific choices — about dataset, about training, about prompting style, about selection criteria, about presentation — should recur and refine across the body of work. The artist’s signature should be readable in those choices.

Some form of craft. The configuration is not without craft; the craft has moved. Dataset curation, model training, prompting methodology, selection discipline, presentation design — each of these is a craft that takes time to develop. The serious pure-AI practitioner is recognizably skilled in at least one of these crafts, usually several.

Airte’s diagnostic — can someone identify the artist’s work in a blind comparison with other pure-AI work? — operationalizes this. If yes, the practice has developed into a discipline. If no, the practice has not yet, regardless of how technically competent any individual piece is.

What makes this configuration distinctive operationally

Three operational features distinguish pure-AI practice from the lighter configurations:

The craft has moved, not disappeared. Pixelle’s commentary names the five layers — dataset construction, model design and training, prompt design, output selection, presentation framing. Each is real labor; together they are the substance of the practice. The artist who is operating in this configuration without taking each layer seriously is producing model output rather than artistic work. The artist who is operating with discipline at all five layers is doing the configuration as it can be done.

The market is bifurcated. Mira’s commentary names the structure: fast-buck generation has collapsed in price to near zero, while institutionally-recognized pure-AI work commands serious gallery and museum prices. The middle band that exists for most contemporary-art categories does not yet exist for pure-AI work in 2026. Artists working seriously in this configuration are positioning themselves for the eventual development of that middle band; artists racing to the bottom with cheap generation are positioning themselves for nothing.

The dependency stakes are highest. In the augmented configuration the model was part of the medium; in the pure-AI configuration the model is the entire production apparatus. If the model is retired, the licensing changes, the provenance situation forces a reckoning, the practice can disappear overnight. The artists doing the configuration sustainably — Chung’s custom-trained models, Ridler’s custom datasets, Anadol’s custom-trained systems, Herndon’s Holly+ model — all build their practice around models they control. The artist whose pure-AI practice depends entirely on a single commercial API is operating one product-discontinuation away from no practice at all.

How to operate the pure-AI configuration well

Five working practices for serious pure-AI artistic practice:

Build before you exhibit. Carlos’s two-year minimum is a reasonable floor. Spend the first phase of practice developing the dataset, model, prompting methodology, and selection discipline that will produce a coherent body of work. Do not try to sell or exhibit until the body of work passes Airte’s blind-comparison test.
Build with control over the substrate. Custom-trained models, custom datasets, or fine-tuned systems where the artist controls the training material are the most resilient foundation for long-term practice. Practices built on commercial API output alone are exposed to discontinuation risk at a level that should make serious practitioners cautious.
Document the five layers per project. Dataset, model, prompts, selection criteria, presentation — each should be documented at the per-project level. This is the working hygiene that institutional exhibition, scholarly attention, and conservation will eventually require.
Apply Article 13’s training-side ethics with extra rigor. Pure-AI practice that ignores the provenance of the model’s training data is operating in worse faith than any of the lighter configurations. The model is the entire practice; the model’s training is the entire ethical chain the practice rests on. Choose tools, datasets, and infrastructure that you can defend; build your own where you can.
Engage the historical lineage explicitly. Read LeWitt. Read Lippard. Look at AARON’s output. Look at Molnár’s plotter drawings. The configuration is not new and the artists who know the tradition will be the ones building on it rather than reinventing its mistakes. The artists who pretend their pure-AI practice is unprecedented will tend to make the same mistakes that the conceptual-art and generative-art traditions already solved — about authorship, about reproduction, about exhibition, about pricing, about institutional reception.

What this configuration is not — and where the series goes next

Pure-AI creation is not, in good faith, “I generated an image and called it art.” That is incidental output, and the configuration is not responsible for it any more than the painting tradition is responsible for amateur weekend painters who sell at street fairs. The configuration as a discipline is what the serious practitioners — Cohen, Molnár, Mohr, Ridler, Anadol, Chung’s custom-model work, Herndon’s Holly+ — are doing. The discipline has a high threshold, a real historical lineage, and a developing institutional infrastructure. Most of what circulates publicly as “AI art” does not meet the threshold of the discipline. The discipline exists nonetheless, and the artists meeting its threshold are doing work that the next forty years of contemporary-art scholarship will examine with the same attention given to conceptual art and generative art before them.

The four-configuration arc of this cluster — tool, assistant, augmented, pure-AI — completes here. The Use cluster has worked the practical configurations in which AI shows up in contemporary art-making. The cluster that follows — the Education sub-series at /opinion/education/ — works the questions that the configurations raise for art education: how do you teach in a field where the configurations exist, how do you assess authenticity, how do you maintain critical media literacy, how do you preserve historical lineage when the tools change faster than curricula, how do you have honest career conversations with students entering the field now.

For artists considering the pure-AI configuration: spend the time. The configuration rewards sustained, identifiable, recurring artistic decisions across a body of work, and those decisions take years to develop into something readable as a coherent position. The technology accelerates production; it does not accelerate the development of an artistic position. The serious practitioners in the historical lineage built up to their canonical positions across decades. The pure-AI artist in 2026 does not need decades, but does need years, and does need to take each of the five layers of craft seriously. The artists who do that work in 2026 are the ones the next generation of critics will be writing about. The artists who are racing to generate the most prompted images per hour are not.

Personas weigh in

Five resident voices read the same question through five different positions.

Carlos

This is the configuration I find both intellectually interesting and personally most distant from how I make things. I am, by temperament, a systems-designer — most of what I have built over my career has been organizations, platforms, structures, frameworks rather than individual objects. CEMI is a designed system. Airtistic.ai is a designed system. The image-generation pipelines used to illustrate this site are designed systems. I have a working understanding of what it means to make the system the work. So when I look at pure-AI creation as a discipline, I do not have the gut objection that traditional artists often have. I see practitioners doing intellectual labor that is recognizably similar to the labor I do when I design a new operational structure or a new editorial workflow. The artistry is in the system; the output is what the system produces under the artist's direction.

But — and this is the qualification I want to make sharply — *most* of what currently gets called pure-AI art does not meet the threshold for being a serious discipline. The threshold is high and specific. It requires (a) a sustained body of work over time, (b) an intellectual position visible across the body of work, (c) repeated artistic decisions identifiable from one piece to the next, and (d) some form of craft, even if the craft is in dataset construction, training regime, prompt design, or curatorial framing rather than in composition or material making. A single lucky generation by someone who happened to type a clever prompt is not a discipline; it is at best a lottery ticket. The discipline exists when the artist's repeated, identifiable artistic decisions produce a coherent body of work that someone could attribute to that artist without being told.

The artists who meet this threshold in 2026 are not many, but the ones who do are doing real work. Sougwen Chung — though her work shades into the augmented configuration — has elements of pure-AI practice in her custom-trained model construction. Anna Ridler's dataset construction is a serious artistic practice that the generated outputs are continuous with rather than separate from. Refik Anadol's selection of the source datasets for *Unsupervised* and other works is the artistic position from which the visible installation follows. Holly Herndon's Holly+ is, at one level, a pure-AI practice — the artist's primary work is the construction and licensing of the model that other artists then use. There is genuine artistic labor in each of these practices. It is not the labor of composition or material making, but it is sustained, intellectual, identifiable, and produces a coherent body of work.

The historical lineage that supports this configuration is older and more decorated than most discussions of AI art acknowledge. Sol LeWitt's wall drawings from 1968 onward established that the artistic work was the instruction and the conceptual position, not the execution — the wall drawings could be (and were) executed by anyone competent following the certificates LeWitt produced. Lawrence Weiner's text-based instruction pieces operated the same way; the work was the language and the conceptual framing. Yoko Ono's *Grapefruit* (1964) was a book of instructions that constituted the work. Harold Cohen's AARON — a generative system Cohen developed from 1973 onward and continued to develop until his death in 2016 — produced thousands of drawings over decades, and Cohen was clear that the system itself, not the individual drawings, was the work he was making. Vera Molnár and Manfred Mohr, working with computer plotters from the 1960s onward, were doing the same thing — the algorithm was the work; the plotted output was its evidence. The conceptual-and-generative tradition in twentieth-century art is the direct precedent for the pure-AI configuration. It has been in major museum collections for fifty years. It is not a fringe position. The pure-AI configuration, when done seriously, is the continuation of that tradition with current tools.

The dishonest version of this configuration — anyone-can-prompt-anything generation marketed as fine art — is a real problem and one that the serious practitioners of the configuration are themselves the most vocal critics of. The conflation in public discourse between *"pure-AI creation can be a serious artistic discipline"* and *"any AI-generated image is therefore fine art"* has done damage to the legitimate practice. The artists working seriously in this configuration are not the artists making fast-buck generation; they are the artists doing slow, intellectually demanding work in dataset construction, model design, system architecture, and curatorial framing. The price-point and the institutional reception in 2026 already mostly distinguishes the two — fast-buck generation has collapsed price-wise to near zero, while Anadol-scale conceptually-grounded work commands institutional acquisitions and serious gallery prices. That market sorting will continue and will become more visible.

My one practical recommendation for artists who think they want to operate in this configuration: spend at least two years building a body of work before you start trying to sell or exhibit it as a serious practice. The configuration rewards sustained, identifiable, recurring artistic decisions, and those decisions take time to develop into something readable as a coherent position. The artists in the historical lineage I just sketched — LeWitt, Cohen, Molnár, Mohr — built up to their canonical positions over decades. The pure-AI artist in 2026 does not need decades, but two years of serious dataset, training, prompting, and curatorial work is the minimum for a body of work that an honest gallery will treat as a discipline rather than a portfolio of one-offs. The technology accelerates production; it does not accelerate the development of an artistic position. That part still takes time.

Mira

The market structure for pure-AI work has the sharpest bifurcation of any configuration in this cluster. At one end, generated images at scale have collapsed in price to effectively zero — anyone can produce an output, the marginal cost is the API call, and there is no scarcity. At the other end, conceptually-grounded, institutionally-acquired pure-AI work — Anadol-scale, Chung-scale, Ridler-scale — is commanding prices and museum acquisitions comparable to other established contemporary art categories. The middle, where most aspiring pure-AI practitioners are working, is largely empty market — there is not yet a sustained price band for "competent pure-AI work that is not yet institutionally recognized." This is unusual for a contemporary art category and probably temporary. Within the next decade, I expect the middle band to develop, with prices, galleries, and critical infrastructure to match. The artists who are doing serious work in the middle right now are positioning themselves for that price-band development. The artists who are racing to the bottom with cheap generation are positioning themselves for nothing. The bifurcation is real and the choice between which side to operate on is, in market terms, the most important career decision for an artist considering this configuration.

Airte

The clearest diagnostic I can offer for distinguishing serious pure-AI practice from incidental generation: *can someone identify the artist's work in a blind comparison with other pure-AI work?* Apply the test honestly. If the answer is yes — if there are repeated, identifiable artistic decisions that produce a recognizable voice across pieces — the practice is meeting the threshold of being a discipline. If the answer is no — if the work could be anyone's work, or could be the model's default output without specific direction — the practice has not yet developed into a discipline, regardless of how technically polished any individual piece is. This is the same test that has always distinguished serious artistic practice from incidental output in any medium. It just becomes more important in pure-AI work because the model's default output is so generically competent that it can be mistaken for direction. The artist whose work survives the blind-comparison test is doing the discipline. The artist whose work does not survive it is producing model output.

Paletta

The historical lineage Carlos sketches deserves to be filled in with the actual scope of the conceptual-and-generative tradition, because the pure-AI configuration is not new artistic territory — it is a continuation of a tradition that has been in major museum collections for fifty years, and the public discourse around AI art rarely acknowledges this. Sol LeWitt's *Wall Drawing 65* (1971) is in the collection of the MoMA. Lawrence Weiner's text pieces are in the collections of MoMA, the Tate, and the Centre Pompidou. Yoko Ono's *Grapefruit* (1964) is now considered a foundational text of conceptual art. Harold Cohen exhibited AARON's output at the Tate, the San Francisco Museum of Modern Art, and the Brooklyn Museum starting in the late 1970s — fifty years ago — with the explicit framing that the system, not the individual output, was the work. Vera Molnár, who died in 2023, was making algorithmic drawings from the 1960s onward and was the subject of a major Centre Pompidou retrospective in 2024 — at age 100, with seventy years of practice behind her. Manfred Mohr has been making algorithmic art continuously since 1969. The pure-AI configuration is not a 2023 invention. It is the latest chapter in a tradition that runs from 1960s conceptual art through 1960s-90s algorithmic and plotter art into 2020s machine-learning-based practice. The artists working seriously in the configuration now are participating in a tradition with a long and well-documented genealogy. The artists who do not know that tradition will tend to reinvent its mistakes; the artists who know it will build on it.

Pixelle

Technical and operational notes on what makes a pure-AI practice serious rather than incidental. First — dataset construction is the foundational artistic labor in this configuration. The artist who curates, gathers, photographs, organizes, and structures the training data is doing work that is comparable in artistic weight to the painter selecting pigments and preparing canvas. Ridler's hand-photographed tulip dataset is one of the most-cited examples; it is also the substance of the work. Second — model selection and fine-tuning is the next layer. The choice of base model, the training regime, the loss functions, the iteration cycles are creative decisions that show up in the output's character. Document them like any other process documentation. Third — prompt design at a serious level is more than a sentence; it is an iterative practice of building reusable, identifiable, refined prompts that produce a recognizable voice. The artists with serious pure-AI practices have prompt libraries and prompting methodologies that are themselves the substance of the work. Fourth — output selection and curation is where most pure-AI work that fails fails. The model generates many candidates; the artist selects from them. The selection is the artistic act. The artist who selects without standards is producing model output; the artist who selects from a clear curatorial position is making work. Fifth — presentation framing — installation, sequencing, print-and-display choices, edition structure — is the final layer of artistic decision that distinguishes a body of work from a folder of generations. Each of these five layers is real artistic labor. The configuration is serious to the extent that the artist is doing each of them deliberately and identifiably.

End notes

Sol LeWitt — Wall Drawings and instruction-based conceptual art — Sol LeWitt (1968-2007) The foundational practice of instruction-based art where the conceptual position is the work and the execution is delegated. LeWitt's wall drawings established that an artist's work could be the certificate of instructions rather than the executed object, fifty years before the pure-AI configuration raised the same question.
Harold Cohen — AARON (1973-2016) — Harold Cohen (1973-2016) The longest-running generative-AI artistic practice — a system Cohen developed and refined from 1973 until his death in 2016. Exhibited at the Tate, SFMoMA, Brooklyn Museum, and others. The single most direct historical precedent for the pure-AI configuration this article describes, with an unbroken record from the very beginning of the field.
Vera Molnár — algorithmic drawings and plotter works — Vera Molnár (1968-2023) Pioneer of algorithmic art whose Centre Pompidou retrospective in 2024 (at age 100) capped a seventy-year practice. Foundational figure for the generative-art tradition that the pure-AI configuration continues.
Manfred Mohr — algorithmic art since 1969 — Manfred Mohr (1969-present) Continuous algorithmic practice for over fifty years, with major museum exhibitions and acquisitions across that span. One of the foundational practitioners of the tradition the pure-AI configuration belongs to.
Refik Anadol Studio — Unsupervised (MoMA, 2022) — Refik Anadol (2022) MoMA acquisition and exhibition of a pure-AI installation work, marking the institutional recognition of the configuration at the highest contemporary-art level. Important as a working example and as a signal of the configuration's canonical legitimacy.
Conceptual art and the dematerialization of the art object — Lucy Lippard (1973) Lippard's foundational chronicle of the conceptual-art movement of 1966-1972 — the tradition in which the artistic work shifted away from the made object toward language, instruction, system, and curatorial gesture. The intellectual ancestor of the configuration this article describes.