When Machines Learned to See: A Century of Art and Technology in Conflict

The "painting is dead" myth and what actually happened

On January 9, 1839, the astronomer François Arago presented Louis Daguerre's invention to the Académie des Sciences in Paris. By August, the French government had declared photography a gift "Free to the World." The painter Paul Delaroche allegedly intoned, "From today, painting is dead!"

Except he almost certainly never said it. Historian Stephen Bann traced the quote to an 1873 survey, 34 years after Delaroche supposedly spoke. The phrase, Bann notes, "has served more than any other to fix the relationship between painting and photography in a condition of mythic stasis." Delaroche kept painting until his death in 1856, and he championed photography, calling it "an immense service to the arts."

What photography did kill was a business. By the early 1850s, Parisian daguerreotype studios produced over 100,000 portraits a year, gutting the painters who had made their living doing portraits for the middle class. Cut loose from pure representation, painting went somewhere nobody predicted: the Impressionists.

Edgar Degas had what art historian Aaron Scharf called "a photographic eye." He "loved and appreciated photography at a time when artists despised it or did not dare admit they made use of it," and by autumn 1895 a friend wrote that "Degas abandons himself entirely to his new passion for photography." The twist is that the cameras of his early career couldn't take snapshots at all; exposures ran several minutes. So his frozen motion, his arbitrary cropping, his visual disjuncture weren't copied from photographs. They showed what the camera couldn't. He had "set about inventing an idiom that showed us the world in the way the camera could not."

The women who taught machines to dream

Before AI could generate art, someone had to teach it to see. That story starts with a Hungarian émigré making marks on paper by hand, imagining a machine that didn't exist yet.

Vera Molnár arrived in Paris in 1947, the year she met Brancusi, Kandinsky, and Vasarely at Café Le Select in Montparnasse. By 1960 she'd developed what she called her "machine imaginaire," an imaginary computer she ran by hand, executing simple algorithmic programs and feeding her drawings through systematic variations. "You could say I am a computer gone wrong, that turned into a woman," she later joked.

When she finally got access to a real computer in 1968, at a Sorbonne research lab, she taught herself FORTRAN. She was 44. She worked without stopping until her death in December 2023 at age 99, and in 1989 she stated the paradox at the center of the whole field: "It may seem paradoxical that this so-called inhuman machine helps to realize what is most subjective, what is most profound in man."

Harold Cohen had represented Britain at the 1966 Venice Biennale before he found programming two years later, as a visiting professor at UC San Diego. By 1973 he'd named his creation AARON, after the biblical figure who spoke for Moses. The program ran and evolved for over four decades, "one of the longest-running, continually maintained AI systems in history." Cohen made two bargains with himself: "I would never accept the position of having to apologize because this was done by a computer," and "the work the program did would have to stand on equal terms with art made by hand." In 1982 he wrote what now reads as prophecy: "We are in the process of coming to terms with the fact that 'intelligence' no longer means, uniquely, 'human intelligence.'"

The establishment took its time catching up. In summer 1968, London's Institute of Contemporary Arts hosted "Cybernetic Serendipity," the first major computer art exhibition, drawing 44,000-60,000 visitors. The Evening Standard caught the mood: "Where in London could you take a hippy, a computer programmer, a ten-year-old schoolboy and guarantee that each would be perfectly happy for an hour without you having to lift a finger to entertain them?" From that show to the Whitney Museum seating a digital art committee took nearly 50 years.

August 2022: the moment everything accelerated

The releases came one on top of another: DALL-E 2 in April 2022, Midjourney in July, Stable Diffusion in August. Midjourney turned a profit within two months on zero marketing spend, and by November 2023 it had 16.4 million users.

Then, on August 29, 2022, Jason Allen won first place in the "Digital Arts/Digitally-Manipulated Photography" category at the Colorado State Fair with "Théâtre D'opéra Spatial," a work made with Midjourney. He'd run at least 624 text prompts and spent roughly 80 hours refining the image through Photoshop and Gigapixel AI.

The backlash was instant. The artist OmniMorpho tweeted, "We're watching the death of artistry unfold right before our eyes." Allen wouldn't back down: "I'm not going to apologize for it." Later he put it more bluntly: "Art is dead, dude. It's over. A.I. won. Humans lost."

The harder moment belonged to Greg Rutkowski, a Polish digital artist who'd spent years building a reputation on fantasy illustrations for Dungeons & Dragons and Magic: The Gathering. His name was turning up as a Stable Diffusion prompt more than 93,000 times, more than Picasso, Michelangelo, or Da Vinci combined. "I became like a meme," he said. "All that we've been working on for so many years has been taken from us so easily with AI."

The artists strike back: a legal reckoning

On January 13, 2023, three artists — cartoonist Sarah Andersen ("Sarah's Scribbles"), watercolor artist Kelly McKernan, and concept artist Karla Ortiz — sued Stability AI, Midjourney, and DeviantArt. Their attorney, Matthew Butterick, noted that his own visual work appeared in the training datasets.

The core allegation was that the LAION dataset, 5.85 billion images scraped from the internet without consent, had been used to train these systems. Andersen had already watched her style turned against her once, by far-right groups who edited her comics; now AI could reproduce it on command. "I see a monster forming," she wrote in the New York Times.

The case runs toward a September 8, 2026 trial date. In August 2024, Judge William Orrick let it proceed to discovery, finding the "model theory" — that a model contains compressed copies of its training data — plausible.

Getty Images filed separately in January 2023 with its own evidence: AI-generated images sometimes reproduced Getty's watermarks, mangled and distorted, which is hard to explain unless training included their copyrighted photographs. The UK judgment in November 2025 split the difference. The court ruled that model weights are not "infringing copies," but Getty won limited trademark claims.

The ArtStation protests that December showed how much fury had built up. On December 14, over 2,000 artists flooded the platform's trending page with "No AI Art" images until they drowned out everything else. What made it sting was that "Trending on ArtStation" had become one of Stable Diffusion's most popular prompts, precisely because the platform's images had been scraped to train it.

The $69 million JPG and the hangover that followed

On March 11, 2021, Christie's sold a JPG for $69,346,250. The work was "Everydays: The First 5000 Days" by Beeple (Mike Winkelmann), a graphic designer who had been posting one digital artwork a day since May 1, 2007. Before NFTs he'd never sold a print for more than $100.

The buyer, Singapore-based cryptocurrency investor Vignesh Sundaresan (pseudonym MetaKovan), paid 42,329 ETH. Beeple's reaction on Twitter was "holy fuck." Critics pointed out that buyer and seller both had a stake in the boom: the sale drove up B20 tokens, a speculative asset tied to other Beeple works, which later collapsed.

The peak came fast. By September 2021, NFT art was running about $881 million in 30-day volume, and the celebrities arrived. Stephen Curry bought a Bored Ape for $178,000. Justin Bieber paid $1.31 million for his, later worth around $59,090. Pak's "The Merge" sold for $91.8 million in December 2021 through a mass-participation mechanism, 28,983 collectors buying 312,686 units.

Then the floor gave out. Trading volume fell 97%, from $17 billion in January 2022 to $466 million that September, with the Terra/LUNA collapse in May and FTX's implosion in November speeding it along. Average NFT prices dropped 92%, from $3,894 in May 2022 to $293 by February 2023. Art NFT volume crashed 93%, from $2.9 billion in 2021 to $197 million in 2024.

The human stories cut both ways. Brittany Pierre, a 36-year-old photographer living paycheck-to-paycheck on jobs paying under $15 an hour, made $109,000 in nine months in 2021. "Last year has been the first year that I could live completely abundantly," she said. Victor Langlois (FEWOCiOUS), a transgender teen artist, sold five lots at Christie's for $2.16 million, thanking buyers for "believing in me and my story."

And then there's Douglas Newton, an 85-year-old artist approached on LinkedIn with an "opportunity" to convert his work to NFTs. Over months he paid escalating fees for "gas," "taxes," and "verification," and never received a cent. "It was like a nightmare," he told investigators. At least two other artists were taken by the same people.

By Q1 2025, 95% of NFT collections had a market cap of zero. Active traders numbered 19,575, down 96% from the peak.

David Hockney at 72, drawing on glass

When the iPad launched in 2010, David Hockney was 72. His sister Margaret had recommended the Brushes app by Steve Sprang, and Hockney, who had announced in 1980 that "I love new mediums," dove in.

He drew at first with his fingers, working en plein air with a device light enough to replace the heavy sketchbooks and supplies. The backlit screen did one thing traditional media couldn't: "Once, when we were just sitting outside the house, we put all the lights off to see the moonlight more clearly. The moon could then be seen to cast shadows of the trees on the grass, so with my backlit iPad I could draw it. This would have been virtually impossible without it."

His Yorkshire series, "The Arrival of Spring in Woldgate, East Yorkshire in 2011," ran to 51 iPad drawings, each titled by date like a visual diary. Then the tool moved under him. In 2016 Brushes was updated: "They altered it, to improve it they said. Well, I thought they had ruined it." By 2018 his assistant Jonathan Wilkinson and a mathematician in Leeds had built him a custom app with brushes Hockney designed himself.

When COVID-19 shut down the world in 2020, Hockney was in Normandy making 116 prints of spring arriving, emailing the daily work to friends. His message to a frightened world was four words: "They can't cancel spring."

Inside the immersive empire of teamLab

In 1998, five engineering students at the University of Tokyo started sketching ideas for something that didn't have a name yet. Toshiyuki Inoko, who would graduate in Mathematical Engineering and Information Physics, founded teamLab officially in 2001 as "an experimental place of collective creation."

The early years were brutal. "We were still ignored by the art world," Inoko recalls. The team sold technology solutions by day and made art late into the night. The break came in 2011, when Takashi Murakami invited them to the Singapore Biennale.

Their philosophy draws on pre-modern Japanese scroll paintings that "contained several scenes derived from multiple perspectives combined into a single painting," what they call "Ultrasubjective Space." "Throwing one's whole body into the art," Inoko explains, "causes the border established between oneself and an artwork to become ambiguous."

When teamLab Borderless opened in Tokyo's Odaiba district on June 21, 2018, running on over 400 projectors, it did better than anyone expected. The "museum without a map," where the artworks drift between rooms, set a Guinness World Record for the most visited museum dedicated to a single group or artist: 2,198,284 visitors in 2019, past the Van Gogh Museum. Half of all overseas visitors said teamLab was their main reason for coming to Tokyo.

The collective now runs around 1,000 members — artists, programmers, engineers, CG animators, mathematicians, architects — all of them calling themselves "ultra-technologists." Inoko's hope for it is plain: "I hope that after people come and see the artworks, it expands how they perceive the world. Everything in this world exists in a continuity."

The man who makes buildings dream

Refik Anadol was eight years old in Istanbul when he watched Blade Runner. "Since that moment," he says, "one of my inspirations has been that question: 'What can a machine do with someone else's memories?'" At the same age he was teaching himself basic programming on a Commodore 64.

After an MFA at UCLA, where he was mentored by the digital art pioneers Casey Reas and Jennifer Steinkamp, Anadol founded his studio in 2014 and became the first artist-in-residence at Google's Artists and Machine Intelligence Program in 2016. There he ran into the DeepDream algorithm, and "Machine Hallucinations" followed.

For the Los Angeles Philharmonic's centennial in September 2018, Anadol turned Frank Gehry's Walt Disney Concert Hall into a canvas. Using 42 projectors at 50K resolution, he processed 587,763 images, 1,880 videos, and 17,773 audio files from the orchestra's archives, 44.5 terabytes of memory in all. The three-part performance climaxed in "Dream," where the visual and audio data merged into hallucinations the New York Times called "a sort of combinatorial Fantasia."

In November 2022, MoMA installed his "Unsupervised" in the main lobby, a 24-foot data sculpture pushing 138,151 records from the museum's collection through custom machine learning to draw maps in 1024 dimensions. Visitors stayed an average of 38 minutes, against under 20 seconds for a typical museum artwork. More than 3 million people saw it before it closed in October 2023, "one of the museum's most popular exhibitions of all time," and MoMA acquired it as the first generative artwork in its permanent collection.

The critics split. Jerry Saltz of New York Magazine compared it to "a lava lamp." Ben Davis of Artnet called it "An Extremely Intelligent Lava Lamp," high-tech formalism treating art history as "just a bunch of random visual tics to be permuted." Anadol answers by building more. His DATALAND, billed as the world's first AI art museum, opens in downtown Los Angeles in 2025.

Trevor Paglen: photographing what shouldn't exist

Trevor Paglen holds a PhD in Geography from UC Berkeley, and it shows in the work, which treats physical space as political terrain.

His "Limit Telephotography" series documents classified military sites through telescopes with focal lengths between 1300mm and 7000mm, shooting from as far as 40-65 miles away. Dense atmosphere, dust, and debris blur the images, and the blur is the point; it makes the apparatus of invisibility visible. "Photographing a secret military base means insisting on the right to do it, and enacting that right," he explains.

With WikiLeaks contributor Jacob Appelbaum, Paglen built "Autonomy Cube" (2014), a Plexiglas sculpture housing circuit boards that run an open Wi-Fi hotspot routing all traffic through the Tor anonymizing network. The sculpture itself becomes a Tor relay used by people worldwide. "We've come to learn that the network is hostile," Paglen says. "The Internet was supposed to be the greatest tool of global communications... And it is. But it has also become the most effective instrument of mass surveillance and potentially one of the greatest instruments of totalitarianism in the history of the world."

His most viral project arrived in September 2019. "ImageNet Roulette," made with AI researcher Kate Crawford, let people upload selfies to be classified by a model trained on the ImageNet database. The results exposed what was sitting in the training data: Black users got labels like "wrongdoer" and racial epithets, Asian users got slurs like "gook, slant-eye."

The response was fast. Within five days, ImageNet announced the removal of 600,040 images and 438 "unsafe" categories. Paglen and Crawford shut the project down on September 27: "ImageNet Roulette has made its point — it has inspired a long-overdue public conversation about the politics of training data."

The aura problem: can digital art be authentic?

In 1935, Walter Benjamin wrote that "even the most perfect reproduction of a work of art is lacking in one element: Its presence in time and space, its unique existence at the place where it happens to be." He called this the "aura," the authenticity that comes from uniqueness, tradition, and distance.

Digital art dissolves the category. Every copy is identical, there's no "original" in the old sense, and the work exists across distributed networks at once. Some pieces pull in real-time data, so each viewing differs in a way Benjamin never had to consider.

Recent scholarship proposes "semi-aura" for AI-generated art, a partial authenticity that comes from human intent and algorithmic generation working together. The idea, from a 2025 AI & Society paper, places AI art "somewhere between the fully unique aura of a historically situated original and the purely mechanical or digital copy."

Not everyone buys it. The artist Molly Crabapple: "AI is a way for witless dullards to suck every bit of anima, pathos, and humanity out of art." The designer Matt Corrall: "Through the sweat and the hours they put in, the artist communicates those feelings to us... Take away that substance, and what do you have? A lifeless matrix of pixels."

The preservationists have a different problem. Vint Cerf, co-inventor of the internet protocols, puts it this way: "What faces all of us with digital preservation is the amount of infrastructure that needs to be available in order to retain the experience... In the older artistic productions, our eyes and ears and sense of touch are sufficient for us to experience those preserved works of art." When MoMA examined early 1990s work by Teiji Furuhashi in 2011, they found "much of the playback and display equipment needed for experiencing this artwork had become obsolete."

Rhizome, affiliated with the New Museum, has built tools to fight this: emulation-as-a-service, web archiving, semantic linked data. Format obsolescence is still on track to swallow a whole generation of digital work.

The democratization paradox: everyone can create, no one can rest

The numbers look like good news. In 2024-2025, 51% of high-net-worth collectors bought digital artwork, and the average share of digital art in collections jumped from 3% to 13% in a single year. Saatchi Art hosts 1.4 million original works from more than 94,000 artists. Procreate has been the best-selling iPad app since 2018, and 32% of artists worldwide now work with digital tools.

The success stories are real. Dina Brodsky credits Instagram with carrying her from career uncertainty to "a financially secure existence where a majority of my time would be spent in the studio, painting." FEWOCiOUS went from transgender teen artist to Christie's auction star, and Brittany Pierre climbed out of poverty-wage jobs on NFT sales.

The cost shows up in what the platform asks in return. Professional illustrator Steven Zapata called DeviantArt's AI launch "completely tone-deaf." Musician Jim Kroft describes "a relentless loop of performance, visibility, and noise," and adds that "every hour you give to the digital world, you're borrowing it from somewhere else in your life. And most artists don't have hours to spare."

The artist @ArtbyKellyanne put the bargain plainly: "I dedicated hours of my day to this social media quest, like a drug dealer, but instead of drugs, I was pushing hashtags. My artistic journey had morphed from a creative adventure into a numbers game. And so, burnout set in." Dan Howard names the slower version of the same damage: "At some point, I stopped creating to grow as an artist and started chasing the algorithm. Comparison kills art. It doesn't happen overnight, but it's like rust — it spreads if you don't catch it."

Instagram content stays visible for an average of 21 hours before the feed buries it. Then you make more.

Where we stand in early 2026

The legal picture is still unresolved, but it's taking shape. In February 2025, Thomson Reuters v. Ross Intelligence became the first major ruling to reject fair use for AI training on copyrighted content. In June 2025, Bartz v. Anthropic found that training itself was "exceedingly transformative" fair use, while downloading pirated copies to do it was not. Anthropic faces billions in potential damages for the piracy, even as the principle of training on legally obtained work survives.

The Copyright Office is unambiguous on authorship: "If content is entirely generated by AI, it cannot be protected by copyright." But in February 2025 it granted copyright to "A Single Piece of American Cheese" by Kent Keirsey after he documented 35 iterative human edits, which points toward a path for AI-assisted work.

The tools keep moving. Midjourney v6.1 promises "best-in-class output quality," Runway Gen-4 holds identity and style steady across persistent characters, Pika 2.0 does physics-defying effects. Video generation is the new frontier.

Adobe Firefly sells itself as "ethical AI" trained on licensed images, but Bloomberg reported in April 2024 that roughly 5% of the training data was AI-generated images from competitors, Midjourney among them. Adobe Stock now holds 57 million AI-generated images, 14% of its library.

Christie's ran its first AI-only auction in March 2025. "Augmented Intelligence" sold $728,784 of work, with Refik Anadol's "Machine Hallucinations - ISS Dreams - A" taking $277,200, and 48% of bidders were Millennials or Gen Z. More than 6,500 artists signed an open letter protesting the sale.

The previous October, thousands of creators — Radiohead members, Julianne Moore, Kevin Bacon, Kazuo Ishiguro among them — signed a statement: "The unlicensed use of creative works for training generative AI is a major, unjust threat to the livelihoods of the people behind those works, and must not be permitted."

The pattern holds — but faster

History offers a hopeful read. Photography didn't kill painting; it freed painters from representation and opened the door to Impressionism. The early computer art pioneers insisted their tools served human work, and each disruption since has, eventually, widened what art could be.

The catch is the clock. From Daguerre's announcement to Impressionism took roughly 30 years. From Stable Diffusion's public release to multi-million-dollar AI art auctions took less than three. The artists suing won't see trial until 2026, by which point the technology will have changed underneath the case. The old pattern may hold, but it's never had to hold this fast, and whether the people displaced along the way get reabsorbed or just left behind is exactly the part history doesn't promise.

Vera Molnár worked until she was 99, and near the end she said this: "Listen, I can't say it any other way: what makes me happy is getting up in the morning, making myself some tea, picking up a pencil... and I'm completely bowled over by this joy that a pencil leaves on paper when you move it around. And that's not even art yet, it's nothing, but it's leaving a mark, making something that hadn't existed until then."

A pencil, some tea, a mark on paper that wasn't there before. That part has survived every machine so far.

---

Key figures reference

Figure	Contribution
Vera Molnár (1924-2023)	Pioneer of computer art; "machine imaginaire" concept; worked until age 99
Harold Cohen	AARON system creator; 40+ years of continuous AI art development
David Hockney	Embraced iPad at 72; created pandemic spring series at 83
teamLab	World record museum attendance; 1,000-member collective
Refik Anadol	MoMA's first generative artwork; DATALAND museum founder
Trevor Paglen	ImageNet Roulette; surveillance documentation
Greg Rutkowski	Most-prompted artist in Stable Diffusion (93,000+ times)

---

Timeline

Year	Event
1839	Daguerre presents photography to French Academy
1968	"Cybernetic Serendipity" exhibition; Molnár accesses first computer
1973	Harold Cohen begins AARON
2010	Hockney starts iPad drawings
2018	teamLab Borderless opens; Anadol's Disney Hall projection
2021	Beeple's $69M NFT sale
2022	Stable Diffusion/Midjourney launch; Colorado State Fair controversy
2023	Artist lawsuits filed; MoMA acquires "Unsupervised"
2025	First AI-only Christie's auction; legal precedents set