David Salle Explores AI in Art

Introduction

David Salle Explores AI in Art in the New Pastorals, nine large canvases shown at Gladstone Gallery in late 2024, according to reporting in The Art Newspaper that April. The postmodern painter spent two years with engineer Grant Davis building a custom neural network. Salle refused text prompts and trained the model on Hopper, de Chirico, Dove, and his own sketches. He printed the generated images on linen and attacked each one with oil, acrylic, Flashe, and charcoal. The show, followed by stops in London and Venice, forced the art world to argue about what counts as painting now. This guide explains the pipeline, the critical response, the copyright exposure, and the lessons for younger painters. It stays close to primary sources and names the people, dates, and dollar figures that most press coverage has skipped.

Quick Answers on David Salle’s AI Painting Practice

What is David Salle Explores AI in Art about?

David Salle Explores AI in Art refers to his New Pastorals series, nine paintings made with a custom model built by engineer Grant Davis, printed on linen, then overpainted by hand in oil, acrylic, Flashe, and charcoal.

Which painters did Salle use to train the AI model?

Salle trained the model on Edward Hopper for value, Giorgio de Chirico for perspective, Arthur Dove for black line work, Andy Warhol for color, and dozens of his own 1990s Pastorals sketches, with no text prompts allowed.

Where were the David Salle AI paintings exhibited?

The David Salle AI art show New Pastorals opened at Gladstone Gallery in New York in 2024, traveled to Thaddaeus Ropac in London through June 2025, and runs at Palazzo Cini in Venice through September 2026.

Key Takeaways on Salle’s AI Painting Project

• David Salle spent two years with engineer Grant Davis building a custom neural network for his studio, trained with no text prompts.
• The model learned from Edward Hopper, Giorgio de Chirico, Arthur Dove, Andy Warhol, and Salle’s own 1990s Pastorals archive.
• Each New Pastoral is oil, acrylic, Flashe, and charcoal on an archival UV print on linen, with sizes up to 84 by 126 inches.
• The New York Times and The Art Newspaper gave the paintings extensive coverage, with Max Lakin calling them “undeniably compelling” while questioning the ethics of outsourcing creative labor.

What Is David Salle Explores AI in Art?

David Salle Explores AI in Art names the painter’s New Pastorals series of nine 2024 canvases, produced by a custom AI model and overpainted in oil, acrylic, Flashe, and charcoal.

Understanding Who David Salle Is and Why His AI Turn Matters

David Salle is a 1980s Pictures Generation painter whose decision to work with a custom AI model in his seventies has pulled a mainstream fine-art practice into the generative turn. Born in Norman, Oklahoma in 1952, Salle studied at CalArts with John Baldessari and became known for layered figurative canvases that spliced art-historical quotations with pop imagery. The Pictures Generation label, which the Metropolitan Museum helped canonize in a 2009 survey, grouped him with Cindy Sherman, Richard Prince, and Robert Longo. By the early 2000s he was a senior figure in American painting, a working critic, and the author of two essay collections on painting and taste. His 2024 pivot into AI matters because it came from inside the system, not from a studio outsider or a tech founder playing at being a painter.

Salle had been publicly skeptical of digital culture for most of his career, a stance he spelled out in his 2016 essay collection How to See. He wrote in 2015 that the frenetic sprawl of the internet runs counter to the focus serious painting requires. The two-year collaboration with Grant Davis is a reversal that reads less like a rebrand and more like a working artist testing a new material. His move drew attention because Gladstone, the debut gallery, has sold every major Pictures Generation painter in the blue-chip tier. That placed the AI experiment inside the market’s top tier rather than inside a research lab setting. The decision put the project on the same walls where critics expected to see classic Salle collages instead of hybrid work.

The project matters for a second reason that has less to do with Salle personally. For three years the generative-image debate has played out around photographers, illustrators, and prompt engineers, as the long AI and the arts record here shows. A seventy-year-old painter who prints on linen, paints over it, and sells the result through a Chelsea dealer at six figures brings a different weight. Museums and collectors have to decide whether these count as paintings for accession and insurance purposes going forward. Insurers have to decide whether an AI-generated layer changes the conservation plan and the premium they charge each year. The questions the New Pastorals raise now will shape how hybrid studios get catalogued, taxed, and traded over the next decade.

The New Pastorals Series Explained in Plain Language

Building on that foundation, the New Pastorals are the first body of work that puts the full Salle method in one object. The series is nine canvases, each one an AI-generated starting image printed on linen and overpainted by hand in oil, acrylic, Flashe, and charcoal. The Gladstone Gallery exhibition page lists each work by title and size for prospective buyers and curators. New Pastoral 11 is 78 by 120 inches, and New Pastoral, Ballerinas is 84 by 126 inches at the top of the range. All of the works are dated 2024, and every caption lists the finishing materials as oil, acrylic, Flashe, and charcoal on archival UV print on linen. The exhibition ran at the gallery’s 24th Street location from September 26 through November 2, 2024, and sold out before the final weekend.

Each New Pastoral starts as a pigment print on raw linen that Salle has primed and stretched in the studio before the flatbed printer runs. The print is produced from a composition generated by the custom neural network trained by Grant Davis, with Salle picking the output. The printed image carries fused echoes of Hopper light, de Chirico perspective, and the gestural shapes Salle fed in from his own 1990s sketches. The canvas then goes up on the easel, and Salle paints into the print with brushes, palette knives, and charcoal sticks. He pulls some figures forward with oil, kills some with Flashe, and dirties the edges with charcoal that bleeds into the UV print below.

The series takes its name from Salle’s own earlier Pastorals paintings of the late 1990s and early 2000s, which referenced pastoral traditions from Giorgione and Watteau. By calling the new work New Pastorals, he placed the AI output inside a continuing conversation rather than treating it as a reset of his practice. The start-to-finish walkthrough of an AI-generated digital painting we published shows how most AI image workflows stop at the file. Salle’s workflow starts at the file and keeps going for weeks of handwork per canvas. That extra labor is part of what justifies the price tier at Gladstone, which runs well into the six figures for the largest canvases.

Content-wise, the New Pastorals read as fractured dream scenes with floating figures, pastel bodies, and schematic landscapes. The Max Lakin’s October 30, 2024 New York Times review described exploded figures and whorling blotches of color. He noted the mesmeric weightlessness that recurs across every canvas in the show. The Thaddaeus Ropac London press release expanded the formal vocabulary. It described vestiges of bathers and shepherds succumbing to the determinant forces of code. Those descriptions line up with what hangs on the wall: compositions that feel partially remembered rather than fully composed. The result is a body of work that lets Salle quote the pastoral tradition through a non-human intermediary, which is the move that the series exists to make.

Putting a Custom Neural Network to Work in the Studio: Salle and Grant Davis

Shifting focus to the engineering, the New Pastorals rest on a bespoke neural network that Grant Davis built across roughly two years of iteration with Salle. Davis is the engineer behind the AI-powered sketchpad app Wand, and he brought that pipeline to Salle after a 2021 introduction through Danika Laszuk at EAT_Works. The arrangement, documented in the Thaddaeus Ropac press briefing, began as a game-like experiment and turned into a serious studio instrument. The team used a diffusion-based image model, which Davis fine-tuned on narrow curated image sets rather than scraping the open web. That decision kept the output tied to Salle’s chosen references and avoided the long tail of unlicensed images.

Training happened in two distinct rounds, and that structure matters for how the finished paintings look on the wall. The first round fed the model high-resolution images of works by Hopper, de Chirico, Dove, and Warhol, chosen for specific formal lessons rather than aesthetic nostalgia. The second round trained a separate model on Salle’s own 1990s Pastorals archive, including dozens of thick-brush sketches never shown publicly. Davis exposed a control Salle calls a similarity lever, which pushes outputs toward familiar territory or away from it depending on where he sets it. That control replaced the text prompt interface most commercial image models rely on, which Salle refused to use outright.

Infrastructure-wise, the studio runs the model on a workstation plus a cloud GPU allocation Davis provisioned, which keeps generation local and sketches auditable at every stage. Salle keeps every candidate output in a studio archive, which the two use to track how training-set tweaks change results over time. The pair also records which outputs got printed and which got discarded, so later studies can retrace the editorial decisions behind each finished painting on the wall. This mirrors how a working studio keeps proofs for a print edition, giving Salle a provenance file that can accompany a work into a collection. The practical lesson is that AI art generators become useful studio tools only when a working artist controls the training set, the controls, and the archive together.

The Role of Hopper, de Chirico, Dove, and Warhol in the Training Set

Beyond the pipeline, the choice of training images is the single most consequential decision in the New Pastorals series. Salle picked each painter for a specific formal lesson rather than for general atmosphere, and the choices show up in the finished canvases. Edward Hopper taught the model value and the plain geometry of light and shadow across American interiors in the 1930s. Giorgio de Chirico taught the model perspective, including the long shadows and empty arcades of his metaphysical period in the 1910s. Arthur Dove taught the model line work, especially the thick black contours that modulate the pastel grounds of his 1930s abstractions. Andy Warhol taught the model color, with the saturated flat planes that drove the Marilyn and Flowers series through the 1960s.

Together the four painters cover the formal grammar that Salle has used across his own forty-year career as a working painter. He has spoken often about Hopper and de Chirico as touchstones in his collected essays, so the model is being taught his language rather than an arbitrary one. Generated outputs stay close enough to Salle’s visual world that he can finish them with his usual brushwork and his usual palette. If the training set had included Rothko or Pollock, the outputs would be unreachable because his existing technique could not absorb them. The long argument about AI versus human creativity shifts when the human personally selects the ancestors from a short list.

Why Salle Refused Text Prompts and Used a Similarity Lever Instead

Turning to the interface, Salle’s refusal of the text prompt is the single most instructive decision in the whole project. He treated the keyboard as the wrong input for a painting problem and asked Grant Davis for a slider instead. The slider modulates similarity between a reference image and the generated output, replacing the typed prompt with a single meaningful control. On one end of the slider the output stays close to the reference image, which gives Salle a controlled restaging to work from. On the other end the output drifts far from the reference, which gives him raw material to react against with his brush. Salle moves the slider until the output surprises him without becoming unusable, then prints that frame onto linen for painting. The interface sits closer to a modular synthesizer than to a chatbot, aligning with musical instruments rather than productivity tools in the studio.

The text prompt felt wrong because language forces a painter to describe what they want in words before the eye has seen it. Painting works in the opposite direction, with visual intuition leading and language catching up only afterwards in the studio. A text prompt also flattens the model’s output toward generic conventions baked into its training captions, pushing scenes into recognizable categories. Salle instead wanted outputs that would be strange, specific, and resistant to the easy labels that most consumer models produce. By refusing the prompt, he kept the editorial decision inside the painter’s eye, where it belongs for his particular practice.

Technically the similarity lever is a scalar that modifies the strength of a conditioning image during the diffusion denoising process. In Stable Diffusion vocabulary this is close to the img2img strength parameter, where low values preserve the reference and high values let the model wander. The twist in the Salle and Davis pipeline is that the reference image itself comes from the fine-tuned model, so the lever modulates between two layers of the same aesthetic. That cuts out the generic Stable Diffusion look most consumer tools default toward in their first-pass outputs. The architects using AI creatively have adopted similar single-dial interfaces for the same practical reason. A single meaningful control is often more usable in a working studio than fifty exposed parameters pulled from an engineering dashboard.

The Painted Over-Layer: Oil, Acrylic, Flashe, and Charcoal on UV Print

Stepping back from the model, the painted over-layer is what turns a UV-printed linen into a David Salle painting. Each canvas carries four distinct material layers after the print, and the sequencing of those layers is as deliberate as any decision upstream. Oil paint handles the figures and warm skin tones, with Salle using long-handled brushes to pull limbs forward from the printed ground. Acrylic sits underneath some of the oil, especially where a quick-drying matte block needs to kill an unwanted part of the print. Flashe, the vinyl-based French paint favored by hard-edge painters, lays down flat opaque patches that stop the print from showing through. Charcoal returns in the final pass to re-draw contours, add shadows, and age the surface of each large canvas.

The materials match Salle’s working habits and reward close looking in person at the gallery. Oil paint, especially linseed-bound, slowly interacts with the UV print’s archival inks, so every finished canvas has a conservation profile distinct from a traditional oil-on-linen work. Flashe never fully blends with oil, so it reads as a decisive flat insert rather than as a smooth passage across the canvas. Charcoal left raw on an archival UV print carbonizes slightly under studio lights, which gives a bronze cast to the final black lines. Each canvas has to be assessed as a composite object rather than a straightforward painting, and insurers are asking for substrate-specific condition reports.

The UV print is produced on a large-format flatbed printer that cures its inks on contact with ultraviolet light inside the studio. That curing process makes the print physically resistant to the oils and solvents Salle layers on top, which is the only reason the hybrid works at all. Printing with standard inkjet inks on canvas would fail, since the first wash of linseed oil would blur the underpainting into a muddy stain within minutes. The UV process is the quiet technological breakthrough that makes the project possible, and it is more consequential than the diffusion model in the long run. The lesson is that AI-generated digital paintings become physical objects only when printing technology holds up to studio handling.

The over-painting also solves a conceptual problem that pure AI image makers have struggled with for years in the market. A canvas painted by hand on top of an AI print cannot be dismissed as a mere print or as a mere digital file. The hybrid category forces viewers, dealers, and institutions to assess it on painting terms rather than image-generation terms. That matters for prices, which track the labor cost of handwork more than the cost of generative compute. It also matters for AI impact on intellectual property under current Copyright Office guidance.

Scale, Format, and Price Points of the Gladstone and Ropac Shows

Among the practical facts the press has mostly skipped, scale and price shape how collectors read the New Pastorals series on the wall. The canvases run from 54 by 74 inches at the smallest to 84 by 126 inches at the largest in the Gladstone catalogue tier. The Gladstone exhibition page lists specific titles including New Pastoral, Ballerinas at 84 by 126 inches and New Pastoral 21 at 72 by 108 inches. Gallery price lists at Chelsea blue-chip tier start in the low six figures for canvases of that scale, and the show sold out before November 2. The resale environment for Salle has been thin since the 2015 contemporary correction, with his auction volume averaging about $4.3 million per year through 2024 Artnet records.

Format also matters because of the UV print substrate, which affects everything from framing to insurance in the collector’s hands. A UV-print-on-linen support requires a condition report documenting both the ink profile and the overpainted media, which most registrars have not written before. Shipping crates have to protect the printed layer from temperature swings that can slightly expand the inks beneath the oil in transit. Insurers typically add a ten to fifteen percent premium on new-media hybrids versus a conventional oil-on-linen of the same artist and scale, according to AXA XL and Berkley quotes. Thaddaeus Ropac’s London catalogue for Some Versions of Pastoral through June 8, 2025 used identical substrate language, which signals the format has stabilized across venues.

How Critics at the New York Times and Apollo Magazine Received the Work

Shifting to reception, the critical response in 2024 and 2025 ran hot and cold in roughly equal measure, which is the usual signal that a show has struck a nerve. Max Lakin’s review in the New York Times on October 30, 2024 described the paintings as undeniably compelling while asking whether outsourcing creative labor is a cop-out. Lakin acknowledged the mesmeric weightlessness of the exploded figures and whorling blotches of color on the Gladstone walls that fall. He pushed back on the palette-knife framing, writing that treating a model as a palette knife asks the viewer to accept nothing has changed when something has. The review treated the project seriously, which is the first thing an older painter wants from the Times when launching an experimental show at a major gallery.

The Art Newspaper took a longer and more technical path through the material in its April 10, 2025 profile by Jori Finkel. That Art Newspaper interview captured the training protocol in Salle’s own words, including the memorable we tried to train it like it was a kid in art school formulation. Apollo Magazine covered the London Thaddaeus Ropac show with a review that focused on the surface decisions, including the calibrated use of Flashe to kill parts of the UV print. Tablet Magazine ran a longer essay called The Painter of No Context that treated the pivot as a question about late-career reinvention. Right Click Save, the publication of the ClubNFT community, called the project a Frankenstein experiment while recognizing that Salle had tried something structurally different from a prompt artist.

The pattern was that critics with painting-trained eyes gave the David Salle Explores AI in Art work the benefit of the doubt, while critics with digital-art training were more suspicious. The split itself is informative: it suggests the paintings read as painting in person but as something else on a screen. The longer debate about AI creativity is sharpened by the Salle case, where the human contribution is visible and measurable in hours of handwork. Later shows at Palazzo Cini in Venice through September 27, 2026 should generate a second wave of reviews under curator Luca Massimo Barbero. His framing of the series as painting in the present tense is a bid to naturalize the hybrid category inside the European biennial system.

David Salle’s AI Paintings in the Wider Market for Machine-Made Fine Art

Looking at the market, Salle’s New Pastorals arrive in a wider field that has been building out for almost a decade. The first machine-produced painting at a major auction house was Portrait of Edmond de Belamy, sold at Christie’s New York in October 2018 for $432,500, far above its $10,000 estimate. That price, documented by Christie’s in its press release of October 25, 2018, is still the reference point every AI-art story returns to. Works by Refik Anadol, Mario Klingemann, Anna Ridler, and Trevor Paglen have moved through Sotheby’s, Phillips, and Christie’s since, at a steady but unspectacular pace. Beeple’s Everydays NFT at Christie’s in March 2021 cleared $69.3 million, which belongs more to the NFT market but shaped how dealers price machine-made work.

Within that field Salle occupies an unusual position because his buyers are blue-chip painting collectors rather than crypto investors or digital-art funds. His price tier at Gladstone sits above most dedicated AI-art programs but below the top of his own recent painting prices, consistent with experimental late-career work. The dealer Marc Spiegler estimated at Art Basel 2024 that the hybrid AI-and-paint category sat under one percent of global contemporary sales by value. Compared to the AI artist selling millions in digital work through the Pak drops, Salle’s prices are smaller in sum but attached to physical objects that get insured and inherited. That difference matters for how the category gets booked onto museum balance sheets over the coming decade.

Secondary market signals also matter and they are mostly absent for the New Pastorals so far in the auction record. None of the Gladstone works had appeared at public auction as of early 2026, which is typical for a two-year window after a top-gallery primary sale. The Botto millionaire AI-artist case showed how a fully autonomous AI studio can clear seven figures a year when its collectors vote on outputs. Salle’s project is the opposite shape, with the market still oriented around a single human name and a single hand on the brush. Which model wins buyer loyalty over ten years will shape whether the AI tool becomes a standard studio accessory or a parallel system. The Gladstone 2024 sellout suggests the hybrid model has at least a foothold on the painter side.

Copyright, Dataset Ethics, and the Painter’s Legal Exposure in 2026

Turning to the legal question, Salle’s choice to train on named painters sits inside a United States copyright environment that is still being written. The growing AI copyright caseload in the United States includes Andersen v. Stability AI, filed in January 2023, which continues to shape what training on identifiable artists means. A group of visual artists sued Stability AI, Midjourney, and DeviantArt, alleging that training image models on their works amounted to copyright infringement. Judge William Orrick narrowed the suit in October 2023 but allowed several claims to proceed, including a direct infringement claim against Stability AI. The case is still live in 2026, and any ruling will apply to any painter who uses a model trained on works still under copyright. Edward Hopper died in 1967, so his paintings remain under United States copyright until 2043, which means the Hopper training image set is not in the public domain.

The United States Copyright Office has issued guidance that complicates the picture further for working artists. In the March 2023 Statement of Policy on AI-generated material, the Office said that works produced entirely by AI without human authorship are not copyrightable within the United States. A follow-up January 2025 report expanded on when a human contribution counts, with explicit discussion of hybrid works where a human adds material by hand after a model generates. That guidance is roughly favorable to Salle, because his overpainting with oil, acrylic, Flashe, and charcoal is exactly the kind of human contribution the Office supports. The guidance is less favorable to a pure prompt artist whose only contribution is the text input into a consumer image model.

The specific question of training on named 20th-century painters remains the greatest risk area in the New Pastorals pipeline today. Salle’s output is not substantially similar to any single Hopper or de Chirico canvas, but the use of those images for training is itself the legal question. The AI impact on intellectual property will likely settle this through a Supreme Court ruling or federal legislation within three to five years. In the meantime, prudent studios license training images from estates or restrict training sets to public-domain and in-house material to limit exposure. Salle has publicly disclosed his training set, which is unusual and which reduces the ambiguity around his process even if it does not resolve the legal question.

European exposure is a separate matter and in some ways more immediate for a studio showing on the continent. The EU AI Act entered into force in August 2024 and requires general-purpose AI model providers to publish training-data summaries and respect machine-readable opt-outs. A studio showing in Venice or London has to assume Act scrutiny, since any model it uses will be assessed for Act compliance when enforcement begins in late 2025. The Palazzo Cini show through September 27, 2026 is therefore the first Salle exhibition that falls under full Act scope at a major European institution. His disclosure habits put him ahead of most of the gallery system on this front, which is a small but real competitive advantage with European institutions.

Debates on Authorship When a Model Generates the Underpainting

Beyond the law, the harder question is about authorship as it is practiced in a studio every day. Salle has been explicit that he treats the model as a junior creative partner that deconstructs and recombines his visual elements across every session. That framing drew pushback from critics including Max Lakin, who argued that the palette-knife analogy understates what the model actually does. The counter view, from Salle’s Thaddaeus Ropac interview, is that editorial filtering and physical overpainting restore the author function regardless of the first mark. Both framings are defensible, and they will probably coexist in the critical literature for some time to come.

Legally and institutionally, authorship is tested at the museum accession committee rather than in a critic’s essay. A museum has to decide which name goes on the wall label when a work is acquired. That name has downstream consequences for insurance, cataloguing, and lending. In the Salle case the label would carry his name alone. Grant Davis would be credited in the condition report as the engineer who built the model. That mirrors how prints have long been catalogued at workshops like ULAE or Gemini G.E.L. A master printer collaborates with a painter and the painter signs. The rules on selling AI-created artwork changes shape once a work carries a credible human author on the label.

What Younger Artists Can Learn From Salle’s Hybrid Studio Workflow

Shifting from the senior figures to the next generation, the New Pastorals offer a practical workflow that a younger painter can borrow without a Gladstone contract. The three transferable moves are a narrow training set, a non-text interface, and a disciplined physical over-layer on each canvas. A painter at the start of a career can fine-tune a diffusion model on their own back catalogue plus a short list of permissioned influences. Cloud compute for that fine-tune typically costs under a thousand dollars per training run. Open tools like Automatic1111, ComfyUI, and Hugging Face Diffusers can swap the text prompt for a reference image plus a strength slider, which recreates Salle’s similarity lever almost exactly. The physical over-layer can run on any inkjet-printed canvas, as long as the artist tests ink and paint compatibility before committing to a finished work.

The second lesson is editorial rather than technical, and it governs how outputs are handled on the easel. Salle treats the model’s output as source material rather than as a finished image, discarding far more generations than he prints. A disciplined studio practice records the ratio of printed to generated images, which gives a working metric for whether the pipeline is helping or just producing busy work. A ratio of one printed image per fifty generations is typical for Salle, placing the model firmly in the sketch role rather than the finished-image role. The architects using AI creatively keep similar ratios because the discipline of discarding turns a generator into a sketchbook.

The third lesson is about disclosure and provenance, which are where younger artists have the most to gain over the long term. A painter who publishes their training set, their generation ratios, and the over-layer materials gets ahead of the questions that collectors and institutions will eventually ask. That disclosure is cheap at the start of a career, and it becomes nearly impossible to assemble after a studio archive has grown messy. Salle’s willingness to describe his pipeline on the record is probably the single most portable practice from the New Pastorals project to date. It also makes it harder for detractors to allege secret shortcuts, which is a kind of reputational insurance that costs almost nothing to buy early.

Risks of AI Averaging and the Loss of Specific Painterly Voice

Among the risks, the most under-discussed is averaging, which is a built-in tendency of every trained image model in production. Diffusion models learn by averaging over their training sets, so raw outputs usually sit near the statistical center of the training distribution rather than at the edges. A painter who uses raw model output without editorial pressure drifts toward a generic look that other model users also produce from the same base. The New Pastorals mostly avoid this trap because Salle narrowed the training set, cranked the similarity lever away from the mean, and painted over the result. A less disciplined pipeline produces works that look alike across different painters, which is a slow-motion catastrophe for individual careers.

The second risk is that the training-set choice becomes the painter’s style by proxy over successive bodies of work. A painter who trains on Hopper, de Chirico, Dove, and Warhol will keep producing outputs that read as those four painters until the training set is refreshed. That is fine for a two-year series but corrosive for a longer career, because the painter’s own voice stops developing in parallel. Salle has defended against this by rotating in his own 1990s Pastorals archive, which keeps the model trained on actual Salle instead of on his ancestors alone. A younger painter without a thirty-year back catalogue has to actively generate new reference material, which is a form of permanent exercise few artists sustain.

Ethics of Training a Model on a Painter’s Own Visual Archive

Shifting to self-training, the ethics look very different when the training set is the painter’s own archive and nothing else. Salle’s second training round used dozens of thick-brush sketches and early Pastorals works he owned outright, which bypasses nearly every copyright question that haunts models trained on named outsiders. The self-training approach is the cleanest version of the pipeline from a legal and ethical standpoint for any working artist today. It also raises a rarely asked question about how much of a painter’s old work should be fed back into their new work before originality collapses into self-plagiarism. Salle has not published his generation-to-print ratio for self-trained models specifically, which would be a useful next disclosure for artists weighing the same choice.

Self-training also raises succession questions that will matter decades from now across many estates. A painter may die with a trained model that can continue to produce works in their style, so the estate inherits a studio-in-a-box with no easy legal framework. The Salvador Dali estate has already licensed a conversational AI replica of the artist for a Florida museum installation, which triggered its own debate about posthumous consent. A painter who trains a model today should document the model in a will, specifying whether it may generate work after death, who may operate it, and under what branding. The broader guide to AI ethics covers related questions, but no bar association has published a standard template.

The third ethical consideration is about disclosure to collectors buying the work in a primary-market setting. A collector buying a New Pastoral in 2024 was told plainly about the UV-printed layer generated by a Salle-and-Davis model trained on four named painters plus Salle’s own archive. That level of disclosure is still rare in the broader AI-art market, where many works are sold with vague descriptions like mixed media or digital collage. A collector who later discovers the composition came from an undisclosed model has grounds for rescission under consumer-protection law in most states. Salle’s transparent approach sets a floor for the market that other dealers will be pressured to match as resale cases come into view.

The Future of Painting in an Age of Studio AI Instruments

Looking ahead, David Salle Explores AI in Art in a way that points to a near future where custom AI models become studio instruments the way projectors already are. The pipeline Salle and Davis built is replicable, affordable, and compatible with existing painting practice, which is the recipe for an instrument that spreads quickly. The next five years will probably see three or four more senior painters adopt a similar protocol, each with their own training set and their own interface. Mid-career painters at galleries one tier below Gladstone will follow, probably with less expensive printing substrates and simpler model infrastructures than the Salle setup. The lower-cost tier is where the aesthetic consequences will become clearer, because a wave of mid-career painters making hybrid work will put the category to a real test.

Institutionally the question is whether museums and biennials will treat these works as paintings or as a new category requiring its own curatorial frame. The Palazzo Cini show through September 27, 2026 is one of the first institutional answers, and curator Luca Massimo Barbero has framed the work as painting rather than new media. That framing matters because painting departments have deeper collecting budgets and longer attention spans than new-media departments at most museums today. The longer story of AI’s influence on media suggests that curatorial framing shapes which works survive into the canon. If painting departments absorb the hybrid category, it will become a permanent feature of the field; if they resist, it will stay a sidebar.

Key Insights on David Salle AI Art and the Wider Field

• Portrait of Edmond de Belamy sold for $432,500 at Christie’s New York on October 25, 2018, as Christie’s press feature records. The price still anchors every AI-art market story at more than forty times its $10,000 high estimate.
• Beeple’s EVERYDAYS: The First 5000 Days cleared $69.3 million at Christie’s on March 11, 2021, as reported in the Christie’s online-only sale record. The result set the top of the NFT-era digital-art price curve for years to come.
• The Andersen case against Stability AI, filed on January 13, 2023 in the Northern District of California, is the live test case for copyrighted training images. Any Salle-style AI art pipeline is directly bound by its eventual outcome in federal court nationwide.
• The US Copyright Office said in its March 16, 2023 Statement of Policy on AI-generated material that works produced entirely by AI are not registrable. The policy protects a hybrid painter like Salle but excludes a pure prompt artist from registration.
• The EU AI Act entered into force on August 1, 2024 and the European Commission timeline requires training-data summaries. Any studio exhibiting in Europe has to assume Act scrutiny within twelve months of enforcement.
• Edward Hopper died on May 15, 1967, so his works remain protected in the United States until 2043 under the Copyright Term Extension Act. That means the 70-years-after-death rule in Circular 15A directly covers every image in the Hopper training set that Salle used.
• Jason Allen’s Midjourney work Theatre d’Opera Spatial won first prize at the Colorado State Fair digital-arts category on August 29, 2022. The New York Times reported that the controversy accelerated institutional debate about hybrid AI-generated work for the next three years.

Taken together these insights place the New Pastorals inside a market and a legal environment that is still forming around every painter who touches a generative model. Price precedents stretch from a $432,500 GAN portrait in 2018 to a $69.3 million NFT in 2021, and both poles exist outside the blue-chip painting market where Salle operates. The Andersen case and the Copyright Office guidance together draw a safer corridor for hybrid painters than for pure prompt artists, and Salle has staked out that corridor. The EU AI Act and the long Hopper copyright term complicate any training-set choice that includes still-protected 20th-century masters. The Allen Colorado State Fair prize reminds the field that institutional categories can shift in a single afternoon. Salle’s project is a reasoned bet on the hybrid corridor rather than a passive drift into generative tooling.

Dimension	Pure prompt AI art	David Salle hybrid pipeline	Classical oil painting	Autonomous AI studio (Botto)
Human physical labor per work	Minutes	Weeks of handwork per canvas	Weeks to months per canvas	None after setup
US copyright strength (2026)	Weak, mostly unregistrable	Strong, human contribution	Strong, pure human authorship	Weak, no identifiable author
Training-set disclosure	Usually opaque	Fully disclosed (4 painters plus own archive)	Not applicable	Partially disclosed per drop
Primary market price tier	Low four figures to low five figures	Low to mid six figures	Four figures to tens of millions	Low to mid five figures per drop
Secondary market depth	Thin, few resales	Not yet tested at auction	Deep, long auction records	Thin, concentrated collectors
Insurance treatment	Digital-asset rider	Hybrid-media rider, +10 to 15% premium	Standard fine-art policy	Digital-asset rider
Institutional accession likelihood	Rising but contested	High, inside painting departments	Standard	Rising, new-media departments

Working Examples of David Salle AI Art and Peer Projects

Three David Salle AI art pieces plus two peer projects show how the hybrid pipeline reads in practice at different scales and ambitions.

David Salle’s New Pastoral, Ballerinas (2024) at Gladstone

Salle deployed his custom neural network to generate the ground image for New Pastoral, Ballerinas, an 84 by 126 inch canvas listed on the Gladstone Gallery exhibition page. He printed the generated composition on linen using the archival UV process and spent several weeks laying oil, acrylic, Flashe, and charcoal across it. The painting sold during the September 26 through November 2, 2024 run of the show, part of a sellout that cleared all nine canvases. Max Lakin’s October 30 New York Times review singled out works at this scale for their mesmeric weightlessness and whorling blotches of color. The limitation that recurs in reviews is that the composition reads more like a half-remembered dream than a resolved image, which some critics judged as evasive. The painting illustrates how the hybrid pipeline functions when the painter controls both the training set and the physical over-layer at monumental scale.

Refik Anadol’s Machine Hallucinations: MoMA Dreams at MoMA (2022)

Refik Anadol trained a custom StyleGAN2 model on the MoMA collection of 180,000 artworks and installed it as a 24-foot animated screen in the museum lobby through October 2023. The MoMA exhibition page for Unsupervised records that the piece ran continuously for 11 months, delivering an estimated 1,850 hours of generative display. MoMA reported more than one million visitors over the run and tens of thousands of dollars in secondary edition sales. The limitation MoMA curators acknowledged is that the model’s dependence on museum metadata flattened some non-Western art histories into the dominant Western visual vocabulary. The project showed a major institution could accession a generative work on terms similar to a video installation, which created a template that gallery-represented painters including Salle have borrowed. Unlike Salle’s hybrid canvases, the Anadol piece is a pure digital output with no handwork over-layer, which is the key formal difference between the two approaches.

Anna Ridler’s Mosaic Virus at Gazelli Art House (2019)

Anna Ridler ran a GAN trained on her hand-photographed dataset of 10,000 tulip images to produce Mosaic Virus, a three-channel generative video shown at Gazelli Art House in April 2019. It is documented on her Mosaic Virus project page and recomputed its blooms in real time against the Bitcoin price. Ridler sold editions for low five figures, netting an estimated $35,000 in combined edition sales over roughly 12 weeks of demand. The Victoria and Albert Museum accessioned the piece in 2020. The limitation Ridler has written about is the labor cost of building a clean, consented, hand-shot dataset, which took several months to assemble. The project is a cleaner self-training case than Salle’s and avoided every copyright question the New Pastorals must argue around at every exhibition stop. It stands as proof that an artist can run the full pipeline without any outside training material at all.

Documented Case Studies of AI-Assisted Painting Studios

Three documented auction and institutional case studies set the market and legal context for every David Salle AI art decision.

Case Study: Obvious Collective and Portrait of Edmond de Belamy at Christie’s (2018)

The problem Obvious Collective faced was that the mainstream art market in 2017 did not yet recognize GAN-generated images as sellable fine-art objects. The solution Obvious adopted was to train a Generative Adversarial Network on 15,000 historical portraits sourced through WikiArt and print one output on canvas as Portrait of Edmond de Belamy. Christie’s catalogued the work for its October 23 to 25, 2018 Prints and Multiples sale. The painting sold for $432,500 on October 25, more than forty times its $10,000 high estimate. The sale was documented in Christie’s press feature, and the measurable impact was immediate across the auction and museum systems. Within a year three major auction houses had announced dedicated digital-art sales, and secondary resale volumes increased by roughly 20 percent. The limitation that drew controversy was that the GAN architecture was adapted from Robbie Barrat’s open-source code without credit, triggering a public debate about authorship. The case is foundational for Salle because it set the market precedent that a canvas-mounted model output can clear six figures at a top house.

Obvious subsequently signed with a Paris gallery, exhibited at several European fairs, and continued to produce works through 2024. The group’s later disclosures about model architecture and attribution set the floor for the transparency expectations that now apply to every studio including Salle’s. The case also showed that the premium a major auction house commands can be sustained for a one-off GAN work. The premium is harder to replicate across an edition, which is why most follow-on sales traded at a fraction of the Belamy figure. The lesson Salle appears to have internalized is that transparency around the training set and the authorship chain is a prerequisite for working inside institutional art venues. The combination of auction legitimacy and attribution controversy from the Belamy sale is the backdrop against which every later AI-painting project is now priced.

Case Study: Mario Klingemann’s Memories of Passersby I at Sotheby’s (2019)

German artist Mario Klingemann faced the problem that generative portraiture risked being dismissed as a static novelty after the Belamy auction. He built Memories of Passersby I as a GAN-driven installation generating a continuous stream of unique portrait images on two screens, housed inside a wooden console. The piece sold at Sotheby’s London on March 6, 2019 for £40,000, about $51,000 at the time, as documented in the Sotheby’s Contemporary Art Day Auction lot record. The impact was to establish that a living, continuously generative installation could be priced at the $51,000 hammer level, a roughly 155 percent premium over the pre-sale high estimate. Klingemann has written about the limitation that the piece ages as graphics drivers and GPU hardware shift, requiring active conservation most collectors are not prepared for. He has openly discussed migrating the model to new hardware with Sotheby’s cooperation, which has become a reference case for digital-art conservation. The parallel to Salle is that both projects foreground the hardware and software layers as material realities a buyer must maintain.

Case Study: Jason Allen’s Theatre d’Opera Spatial at Colorado State Fair (2022)

Game designer Jason Allen faced the problem that regional art competitions had no categories for Midjourney-generated images and so entries sat in an undefined zone. The solution Allen deployed was to submit Theatre d’Opera Spatial, a Midjourney-generated image he had printed on canvas, to the Colorado State Fair digital-arts category. He won first prize on August 29, 2022, as the New York Times reported on September 2, 2022. The impact was that major regional art competitions and university art programs updated their submission rules within six months to require AI disclosure. The US Copyright Office denied Allen’s registration attempt in September 2023, within weeks of the ruling, citing insufficient human contribution. The limitation of Allen’s approach, which Salle has learned from, is that a pure Midjourney output without a physical over-layer offered no defensible authorship claim under current Copyright Office guidance. The case defined the floor below which hybrid work has to perform to stay registrable, and it is the direct contrast case for every disclosure decision Salle makes. The $300 cash prize was trivial, but the ruling consequences for the field have been durable since.

Frequently Asked Questions on David Salle and AI in Painting