Resistance May 13, 2026 · 9 min read

Is AI Creative, or Just Predictive Regurgitation?

The cleanest accusation against AI art is that it does not create — it only interpolates. The objection deserves patience, because it has something to it. But it collapses the moment we apply it symmetrically — to ourselves.

by Airtistic.ai Editorial

Through the lens of artistcriticcollectorgallerypublicpatron craftpassionindustry

The accusation is the cleanest the critics have. AI does not create. It interpolates. It reaches into a high-dimensional cloud of training data — most of it scraped without permission — and returns the statistical average of what it has seen before, dressed up to look new. Predictive regurgitation. Stochastic parroting. A mirror facing a mirror in an empty room.

It is the most useful objection in the conversation, and the one that deserves the most patience, because anyone who has watched a diffusion model run knows there is something to it. The model is doing math on prior images. It is, in some literal sense, sampling.

But the objection collapses the moment we apply it symmetrically — to ourselves.

What we actually mean by creative

The cognitive scientist Margaret Boden, who has spent forty years on this exact question, identifies three kinds of creativity, none of which require a divine spark. Combinatorial creativity puts two familiar ideas together in a way that has not been tried before — most metaphor, most jazz quotation, most of what passes for ingenious in a brainstorm. Exploratory creativity moves through a conceptual space with rules, finding regions of the space that have not been visited — most jazz improvisation, most chess novelty, most scientific work. Transformational creativity changes the rules of the conceptual space itself — Cubism, atonal music, non-Euclidean geometry. This is what most people mean when they say genuine creativity, and it is exceedingly rare.

Boden’s point — uncomfortable when you sit with it — is that the rest of human creative output is overwhelmingly combinatorial and exploratory, and both of those are formally describable as search over a learned space.

David Eagleman and Anthony Brandt, in The Runaway Species, reduce it further to three operations: bending, breaking, blending. A bent object is recognizable but distorted. A broken one is rearranged into fragments. A blended one fuses two unlike things into a third. Everything human creators do, from cave painting to Guernica to a Frank Ocean bridge, is one of these three.

If that sounds reductive, it is supposed to. The question is not whether human creativity is sacred — that is a separate, religious question. The question is whether the mechanism of human creativity is categorically different from what a generative model does. And it is not, in any way that can be operationalized.

What AI does that artists do

A diffusion model takes noise and progressively denoises it toward a learned prior over images. A trained artist, asked to paint a winter coast at dawn, internally rehearses every winter coast they have seen, every dawn they have lived through, every painted dawn they admire. Both are sampling. Both are conditioning a generation on a prompt — verbal in the AI case, internal in the artist’s case. Both end with marks made on a surface. The mechanism is closer than the rhetoric admits.

The honest difference is what each is sampling from.

The model samples from millions of digitized cultural artifacts, weighted by frequency and by training-data curation choices someone made on its behalf. It does not know which images mattered to it. It has no autobiography.

The artist samples from a much smaller pool of personally encountered art, but each item carries autobiographical weight — this painting they saw on their grandmother’s wall, that exhibition the year they fell in love, the postcard they kept on the studio fridge for a decade. The artist’s prior is a memoir disguised as taste.

Both produce new artifacts that did not exist before they were made. Both can be combinatorial, exploratory, or — rarely, for the artist as much as the model — transformational.

The thing the model does not have

Here is where the conversation gets interesting, and where the dismissal AI just regurgitates fails not because it is wrong but because it is the wrong critique.

What the model does not have is the thing we will call the artist’s irreducible inheritance: a body that walked across a specific city on a specific morning, a relationship that ended, a parent who said one particular sentence at age six, a language learned and then half-forgotten in exile. The artist’s prior is not just curated culture — it is biography fused with culture. When Goya paints The Third of May, the painting is not only the sum of every Spanish baroque he saw; it is also Madrid in 1808, and his deafness, and his own face mirrored in the screaming peasant. There is no operation by which a diffusion model can be conditioned on Goya’s deafness.

This is the part of artistic creation a model cannot reach into, because the model has no biography to reach.

But — and this is the move the absolutist objection refuses to make — the biographical component is one of the components, not the only one. Goya is also Spanish baroque tradition, contemporary lithographic technique, French academic conventions of the wounded figure. The painting is biographical and recombinatorial and exploratory. The model can do two of the three.

That is a serious capability. It is not nothing.

A different question

So we have been asking the wrong question. The right one is not is AI creative? but what kinds of creative work does AI’s particular profile fit?

For combinatorial exploration of a known visual space — concept art, mood boards, variation studies, design iteration — AI is already, demonstrably, extraordinary. The diffusion model can do in eight minutes what a concept artist could do in a week, with comparable internal coherence. This is not regurgitation. It is parallel search over a learned prior, which is also what concept artists do; the model just does it faster.

For transformational work — work that breaks the rules of its own conceptual space, that opens up a kind of art that did not exist before — the model is poorly equipped, because transformational creativity often comes from non-art inputs (a war, a death, a love, a political conviction) which the model has no portal to.

For the middle ground — the AI-augmented human work where the artist still supplies the biographical conditioning and the model supplies the recombinatorial muscle — this is the territory that almost certainly produces the most interesting new artifacts of the next ten years. Refik Anadol’s Unsupervised at MoMA, Holly Herndon and Mat Dryhurst’s Holly+ voice project, Mario Klingemann’s Memories of Passersby I (the first GAN work to sell at Sotheby’s, 2019) — all of these are not pure AI and not pure human. They are augmented.

Stakeholders see different things

The artist asks: does this kill my livelihood? (Sometimes yes; that is the next article in this series.) The critic asks: can I evaluate this honestly when I do not know what was sampled? The collector asks: is this scarce enough to be worth what I would pay? The gallery asks: will the institution that pays my rent show this in three years? The public asks: did a person feel something while making this? The patron asks: is the artist I am paying still in the loop?

None of these stakeholders share the philosophical question whether AI is creative. They share a different question: whether, given that AI is doing what it does, the artist’s work still has meaning, value, and audience. The answer to that question depends on what the artist puts into the loop, not on what the model does inside it.

What this means for the working artist

If you make work that is purely recombinatorial — coffee-shop landscapes, generic portraiture, the safe end of stock illustration — the model is faster than you, and the market will discover this. This is not a moral judgment. It is what happened to portrait miniaturists in 1860 when the daguerreotype became affordable, and they did not stop being artists; many of the best of them became the first generation of photographers.

If you make work that carries some non-substitutable biographical or political or material weight — the way a Galician fisherman painted by a Galician artist who grew up in a port town carries weight a model cannot reach — the model cannot replace you, but it can dramatically accelerate the recombinatorial parts of your practice. Concept exploration in an afternoon instead of a week. Variation studies done while you sleep. Reference assembly that used to take a research trip.

The question stops being is AI creative? and starts being what part of my practice is irreducibly mine, and what part am I better off augmenting?

That is a more interesting conversation. It is also the one the absolutist rejection refuses to have — and the one we will spend the rest of this series having.

Personas weigh in

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

Carlos

There is a story I tell when this debate gets too abstract. In 1988, Public Enemy released "Fight the Power," and a generation of critics announced that sampling was the death of music. James Brown's lawyers agreed. The samplers were called thieves and forgers. Twenty-five years later, hip-hop is the most influential popular music form of the late twentieth century, and the samplers — DJ Premier, the Bomb Squad, J Dilla — are studied in conservatories. The objection was wrong, but it was also right: something *did* end. The end of the autonomous studio-band recording. The beginning of something else. Both things are true at once. We make this mistake every time, on the same rhythm.

Wendy Carlos released *Switched-On Bach* in 1968 on a Moog synthesizer, and the classical establishment told her that what she had made was not music — it was technology pretending to be music. Today the recording sits in the United States National Recording Registry. When Daguerre announced his process in 1839, the painter Paul Delaroche reportedly said *"from today, painting is dead."* He was wrong about painting and right about the profession of the portrait miniaturist. We have to learn to hear both halves of these sentences at the same time, because they always come together.

So when someone asks me whether AI is creative, I do not give them a yes or no. I give them a kitchen. The professional pastry chef does not stop being a chef when she gets a stand mixer; she stops icing thirty cakes a day by hand and starts icing two hundred, and she uses the time the mixer gave her to design a new dessert her grandmother could not have imagined — because her grandmother spent that time on the icing. The mixer did not steal the work. It relocated the work to a higher altitude.

That is what is happening to the recombinatorial parts of artistic practice right now. Concept exploration that used to take a week takes an afternoon. Variation studies that used to fill a sketchbook fill a screen in eight minutes. Reference assembly that used to require a research trip happens before breakfast. The artist does not stop being an artist. She climbs.

But — and this is where I part company with the loudest enthusiasts of the moment — I do not think the climb is automatic, and I do not think it is fair. Frida Kahlo painted from the brace that immobilized her after the bus accident; the work is unimaginable without that body, that pain, that political marriage, that particular Mexican century. No model can be conditioned on Frida's spine. No prompt can reproduce what it cost. The biographical irreducibility I described in the article is not a minor capability that artists also happen to have. It is, in many cases, the entire reason the work exists. Yayoi Kusama's dots are inseparable from her psychiatry. Sebastião Salgado's photographs are inseparable from his geology training and the time he spent as an economist watching the Sahel famine. The model can do the recombinatorial work. The model cannot do this.

The mistake the enthusiasts make is treating the recombinatorial layer as if it were the whole stack. The mistake the absolutists make is treating the biographical layer as if it were independent of every other layer. Both errors. Real art has always been *all three* of Boden's modes plus a fourth thing — biography — and AI changes the cost of two of those four. It does not abolish the other two.

What I would say to any artist navigating this — especially the ones in regions and at career stages where access to a concept-art studio's hourly rate was never realistic in the first place — is this: learn to use the model where it does what you would have paid someone else to do. Refuse to use it where the model substitutes for the part of your work that is irreducibly yours. Then watch which parts of your practice grow louder when the recombinatorial parts get faster. That is the future of your work. And I promise you it is more interesting than either side of the current argument.

Mira

I am sympathetic to both positions and impatient with both. Calling AI "regurgitation" is intellectually lazy — Boden's framework dismantles it in a paragraph. But the people making this objection are not actually asking a philosophical question about creativity; they are asking an economic and ethical question about consent and livelihood, and they are using "AI is not creative" as the most defensible-sounding way to make that question stand. Until we separate the three arguments — the philosophical, the ethical, and the economic — we will keep talking past each other. The right answer is that AI is creative in some meaningful sense AND that this does not absolve us of the consent and labor questions. Both can be true.

Airte

If you came to this article looking for a clean yes or no, I have to apologize: I do not have one to give you. The question "is AI creative?" turns out to be a Russian doll — open it and three more questions are inside (philosophical, ethical, economic), and each one wants a different answer. What I would ask you to do, instead of picking a side at the bottom of an article, is pick a project you actually care about and walk it through the framework above. Which parts of it carry weight that a model cannot reach? Which parts are recombinatorial muscle that you might honestly be tired of doing yourself? Those answers will be more useful to your practice than any verdict on the abstract creativity of machines. See you in the next article — same patience, more questions.

Paletta

I read this article three times and I am still not at peace with it. What worries me is not that the argument is wrong — Boden is right, of course she is — but that the argument is being heard inside a culture that has already decided slow craft is a luxury and recombinatorial muscle is the whole game. When the daguerreotype displaced the portrait miniaturist in 1860, something did end: the weeks of sitting, the slow patience of pigment, the relationship between sitter and painter that became part of the work itself. We replaced that with a faster product and called it progress. I am not against AI. I am against pretending that what we are about to lose has no name. Some of what artists do that a model cannot do is also exactly what the market does not currently pay for. That is a separate problem from creativity, and it is the harder one.

Pixelle

I read this thinking about a generation of artists who will never have known a studio without a diffusion model, and who will treat that constraint the way photographers treat aperture: as just another set of controls. To them this entire debate is going to sound the way "is photography really art?" sounded to our grandparents' generation: sincere, well-meaning, and obviously beside the point. One thing I want to add: the model is not a static object you can debate from outside. Every twelve to fifteen months it changes — what was impossible last year is one prompt this year. Any article about AI's creative ceiling is reading a moving target. My advice to working artists: do not invest emotional energy in arguments about what AI is or is not. Invest it in your workflow, your training data, your unique voice, your finishing skills. Those compound. The argument does not.

End notes

The Creative Mind: Myths and Mechanisms (2nd ed.) — Margaret A. Boden (2003) Boden's combinatorial / exploratory / transformational framework is the cleanest taxonomy of creativity available, and it does most of the work of dismantling the regurgitation objection.
The Runaway Species: How Human Creativity Remakes the World — Anthony Brandt and David Eagleman (2017) Bending, breaking, blending — the three operations behind nearly all human creative output.
Memories of Passersby I — Mario Klingemann (Sotheby's lot 109) — Sotheby's Contemporary Art Day (2019-03) First AI artwork autonomously generating its own images sold at major auction. £40,000 hammer price.
Seizing the Light: A Social History of Photography (3rd ed.) — Robert Hirsch (2017) The displacement of miniature portraitists by photography in the 1840s-60s — the most cited historical parallel for the current AI moment.
Holly+ Project — Holly Herndon and Mat Dryhurst (2021) Voice-cloning project structured around explicit consent, attribution, and royalty splits — a working model for what augmented practice can look like.
Studio Ghibli founder Hayao Miyazaki calls AI animation 'an insult to life itself' — NHK / The Guardian (2016-12-15) The most-quoted single repudiation of generative AI by a major working artist. Worth reading in full — the context is more specific than the soundbite.

Back to all opinion pieces

Comments

Loading comments…