Art
Last reviewed
Sources
31 citations
Review status
Source-backed
Revision
v4 ยท 4,336 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
31 citations
Review status
Source-backed
Revision
v4 ยท 4,336 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Art ChatGPT Plugins
AI art is visual, audio, or other creative work produced with the help of artificial intelligence, most commonly images generated from text prompts by text-to-image models such as DALL-E, Midjourney, and Stable Diffusion. The relationship between artificial intelligence and art spans more than fifty years, beginning with rule-based drawing programs of the 1970s and accelerating into the consumer text-to-image era after 2021. AI art now refers to a broad set of practices in which artists use neural networks, generative adversarial networks, or diffusion models as collaborators, instruments, or autonomous agents. The field encompasses static images, video, music, poetry, sculpture, and live performance, and it sits at the center of unresolved fights over copyright, training-data consent, labor displacement, and authorship.
The single moment that pushed AI art into the mainstream came on 29 August 2022, when Jason M. Allen's Theatre D'opera Spatial, made with Midjourney, won first place in the digital-arts category of the Colorado State Fair, the first time a text-to-image image is widely reported to have won a fine-art competition.[19] The years that followed produced record auction prices, a string of U.S. Copyright Office refusals, and a wave of class-action lawsuits brought by illustrators against the companies behind the models.
What counts as AI art has shifted with each technical wave. Early practitioners wrote symbolic systems that produced drawings rule by rule. By the mid-2010s, generative adversarial networks and neural style transfer let artists hand higher-level decisions to the model. The 2021 arrival of text-to-image systems trained on web-scale image-text pairs collapsed the technical barrier to entry, drawing millions of non-artists into image generation through tools such as DALL-E, Midjourney, and Stable Diffusion.
The practice now spans three loose communities. The first is the long-standing computer-art and algorithmic-art tradition descended from Harold Cohen, which treats code as the medium. The second is the gallery and auction system, which began collecting AI-assisted works in 2018 and now hosts dedicated auctions and museum acquisitions. The third is the much larger commercial and amateur user base of consumer image tools, who often have no fine-art training and whose output dominates platforms like ArtStation, DeviantArt, and Reddit. Conflicts between these groups, particularly between commercial illustrators and users of scraped-data models, have driven the legal and cultural backlash of the 2020s.
The British-born American artist Harold Cohen began work on the drawing program AARON around 1968 at the University of California, San Diego, and named the system AARON in the early 1970s. After a 1971 presentation at the Fall Joint Computer Conference, Cohen was invited to Stanford's AI Lab, where AARON developed further from 1973 onward (Computer History Museum, 2016, https://computerhistory.org/blog/harold-cohen-and-aaron-a-40-year-collaboration/).[1] AARON began as a system for generating monochrome line drawings and gradually acquired the ability to depict figures, plants, and rooms, eventually producing color paintings on canvas.[1] Cohen continued to revise AARON until his death in 2016. The Whitney Museum of American Art mounted Harold Cohen: AARON in 2024 (Whitney Museum, 2024, https://whitney.org/exhibitions/harold-cohen-aaron).[2] AARON is widely cited as the longest-running AI art project in history.
The 1970s and 1980s saw a parallel scene in algorithmic and computer-generated art that included Vera Molnar, Manfred Mohr, and Frieder Nake. These practitioners did not use machine learning, but their work established that an artist could write code as a primary medium, an idea later extended by the AI generation.
The modern phase of AI art began with two technical events. In June 2014, Ian Goodfellow and seven co-authors published Generative Adversarial Networks at NIPS, introducing a training procedure in which a generator network competes against a discriminator that tries to distinguish real samples from fakes (Goodfellow et al., 2014, https://arxiv.org/abs/1406.2661).[3] In August 2015, Leon Gatys, Alexander Ecker, and Matthias Bethge posted A Neural Algorithm of Artistic Style, which used a deep convolutional network trained for image classification to transfer the brushwork of one image onto the content of another (Gatys et al., 2015, https://arxiv.org/abs/1508.06576).[4] Style transfer rapidly entered consumer apps such as Prisma in 2016.
GANs matured through a series of NVIDIA papers by Tero Karras and collaborators. A Style-Based Generator Architecture for Generative Adversarial Networks introduced StyleGAN at CVPR 2019, with StyleGAN2 following at CVPR 2020 (Karras et al., 2018, https://arxiv.org/abs/1812.04948; Karras et al., 2019, https://arxiv.org/abs/1912.04958).[5][6] Artists adopted these models quickly because pre-trained weights and fine-tuning code were open-sourced.
In June 2020, Jonathan Ho, Ajay Jain, and Pieter Abbeel published Denoising Diffusion Probabilistic Models, which framed image generation as the gradual reversal of a noise process and reported a state-of-the-art FID of 3.17 on CIFAR-10 (Ho et al., 2020, https://arxiv.org/abs/2006.11239).[7] In December 2021, Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer at CompVis (LMU Munich) preprinted High-Resolution Image Synthesis with Latent Diffusion Models, which moved the diffusion process into the latent space of a pre-trained autoencoder and made high-resolution synthesis tractable on consumer GPUs.[8] The paper appeared at CVPR 2022 and provided the architecture used by Stable Diffusion.[8]
On 5 January 2021, OpenAI introduced DALL-E and CLIP in companion blog posts, with DALL-E based on a 12-billion-parameter variant of GPT-3 trained on text-image pairs.[9] DALL-E 2, an unrelated diffusion-based successor, launched on 6 April 2022 with phased access; OpenAI removed the waitlist and opened it to the general public on 28 September 2022, by which point more than 1.5 million users were generating over 2 million images a day with the system (TechCrunch, 2022, https://techcrunch.com/2022/09/28/openai-removes-the-waitlist-for-dall-e-2-allowing-anyone-to-sign-up/).[10][31] Midjourney entered open beta on 12 July 2022, founded by David Holz, formerly of Leap Motion. Stability AI, CompVis, and Runway released Stable Diffusion publicly on 22 August 2022, shipping the v1.4 checkpoint on Hugging Face under the CreativeML Open RAIL-M license, the first major model of its class to ship with open downloadable weights and no access wall.[11]
| Tool | Year | Organization | Type |
|---|---|---|---|
| AARON | 1973 onward | Harold Cohen | Symbolic drawing program |
| Prisma | 2016 | Prisma Labs | Neural style transfer app |
| StyleGAN, StyleGAN2 | 2018, 2019 | NVIDIA | GAN model family |
| DALL-E | January 2021 | OpenAI | Text-to-image transformer |
| DALL-E 2 | April 2022 | OpenAI | Diffusion text-to-image |
| Imagen | May 2022 (paper) | Google Brain | Diffusion text-to-image |
| Midjourney | July 2022 (open beta) | Midjourney Inc. | Diffusion text-to-image |
| Stable Diffusion | August 2022 | Stability AI, CompVis, Runway | Open-weight latent diffusion |
| ControlNet | February 2023 | Lvmin Zhang and Maneesh Agrawala | Conditioning network for SD |
| Adobe Firefly | March 2023 | Adobe | Diffusion model trained on Adobe Stock |
| Stable Diffusion 3 Medium | June 2024 | Stability AI | Diffusion-transformer hybrid |
| FLUX.1 | August 2024 | Black Forest Labs | Flow-matching text-to-image |
DALL-E 2's training and architecture were described in Aditya Ramesh et al., Hierarchical Text-Conditional Image Generation with CLIP Latents (OpenAI, 2022). Stable Diffusion was trained on the 2-billion-image English subset of LAION-5B, a CLIP-filtered crawl of the open web assembled by the German nonprofit LAION.[11] Adobe Firefly, launched on 21 March 2023, was instead trained on licensed Adobe Stock images, openly licensed content, and public-domain works, a policy designed to limit commercial-use risk (Adobe, 2023, https://news.adobe.com/news/news-details/2023/adobe-unveils-firefly-a-family-of-new-creative-generative-ai).[12] ControlNet, introduced in February 2023, added spatial conditioning (edge maps, depth, pose) on top of Stable Diffusion and became a standard part of professional diffusion pipelines. Black Forest Labs, founded by several of the original Stable Diffusion researchers, released the FLUX.1 family on 1 August 2024 with a $31 million seed round led by Andreessen Horowitz (VentureBeat, 2024, https://venturebeat.com/ai/stable-diffusion-creators-launch-black-forest-labs-secure-31m-for-flux-1-ai-image-generator).[13]
Google's Imagen was described in a May 2022 paper by Saharia et al. and used three cascaded diffusion models conditioned on a frozen T5 text encoder. The model was not initially released to the public; a constrained version arrived later in 2022 through the AI Test Kitchen app under the names City Dreamer and Wobble. Subsequent Imagen versions were exposed through Vertex AI and the Gemini API.
| Artist | Work | Year | Technique |
|---|---|---|---|
| Harold Cohen | AARON drawings | 1973 onward | Rule-based symbolic system |
| Mario Klingemann | Memories of Passersby I | 2019 | Multi-GAN installation |
| Memo Akten | Learning to See | 2017 | Pix2Pix-style generation |
| Obvious Collective | Edmond de Belamy | 2018 | GAN (code based on Robbie Barrat) |
| Trevor Paglen | A Study of Invisible Images | 2017 onward | Computer-vision exposure |
| Sougwen Chung | DOUG drawing robots | 2015 onward | Robotic-AI collaboration |
| Holly Herndon | PROTO with Spawn | 2019 | Custom neural network on her voice |
| Refik Anadol | Unsupervised | 2022 | StyleGAN2 on MoMA collection |
| Jason M. Allen | Theatre D'opera Spatial | 2022 | Midjourney plus Photoshop and Gigapixel AI |
| Sasha Stiles | Technelegy | 2021 onward | Fine-tuned language model |
| Boris Eldagsen | Pseudomnesia: The Electrician | 2022 | DALL-E 2 |
| Botto (BottoDAO) | Weekly NFT drops | 2021 onward | VQGAN+CLIP, Stable Diffusion, community vote |
Mario Klingemann developed a technique he calls Neural Glitch, in which he randomly alters or deletes weights of trained GANs to push them into unfamiliar visual territory. Memories of Passersby I sold at Sotheby's in March 2019 for 40,000 pounds, the first autonomous AI installation to be auctioned by the house. Memo Akten's Learning to See (2017) used a Pix2Pix-style network trained on art history to reinterpret live webcam input as paintings.
Refik Anadol, born in Istanbul in 1985 and based in Los Angeles, has been working with machine learning since a 2016 residency at Google's Artists and Machine Intelligence program. Sougwen Chung's Drawing Operations Unit (DOUG), begun in 2015, pairs the artist with one or more robotic arms whose movements are guided by computer vision and, in later iterations, neural networks trained on Chung's own archive. Holly Herndon and Mat Dryhurst built the neural-vocal-ensemble system Spawn with developer Jules LaPlace in 2016 and integrated it into Herndon's 2019 album PROTO on 4AD.
Sasha Stiles, a Kalmyk-American poet, fine-tunes a language model she calls Technelegy on her own writing and research; her poem COMPLETION: When it's just you sold at Christie's in 2022, an early auction of AI-assisted literature. Botto, conceived by Mario Klingemann with the studio ElevenYellow and launched in October 2021, is governed by the BottoDAO. Each week, Botto generates roughly 70,000 images, presents around 350 to token holders, and auctions the most-voted work as an NFT on SuperRare.[29] Botto's sales had exceeded $5 million by late 2024 (CNBC, 2024, https://www.cnbc.com/2024/12/23/botto-the-ai-machine-artist-making-millions-of-dollars.html).[29]
Edmond de Belamy, from La Famille de Belamy, a GAN portrait produced by the Paris-based Obvious collective, sold at Christie's New York on 25 October 2018 for $432,500 including premium, against a presale high estimate of $10,000, roughly 43 times the estimate after about seven minutes of bidding among four bidders (Artnet News, 2018, https://news.artnet.com/market/first-ever-artificial-intelligence-portrait-painting-sells-at-christies-1379902).[14] Obvious trained the network on 15,000 portraits from the 14th to 20th centuries. The sale was the first AI work auctioned by a major house, but it generated immediate controversy when it became clear that Obvious had used code released by 19-year-old developer Robbie Barrat with minimal modification, an issue Mario Klingemann publicly raised and that Obvious's Hugo Caselles-Dupre conceded in the press.[15]
Mario Klingemann's Memories of Passersby I sold at Sotheby's London in March 2019 for 40,000 pounds. Refik Anadol: Unsupervised opened in the Gund Lobby of the Museum of Modern Art in New York on 19 November 2022 and ran until 29 October 2023. The piece used StyleGAN2 with adaptive discriminator augmentation, trained on metadata from MoMA's collection, and projected continuously evolving imagery onto a 24-by-24-foot LED wall while drawing live environmental data from the lobby (MoMA, 2022, https://www.moma.org/calendar/exhibitions/5535).[16] MoMA later acquired the work via a gift from Ryan Zurrer and the RFC Collection.
Christie's Augmented Intelligence sale, the first auction at a major house dedicated entirely to AI art, was announced in early February 2025 and closed in March 2025 with a total of $728,784, selling 28 of its 34 lots and outpacing its roughly $600,000 high estimate (The Art Newspaper, 2025, https://www.theartnewspaper.com/2025/03/05/christies-augmented-intelligence-sale).[17] Christie's reported that 48 percent of bidders were Millennials or Gen Z and 37 percent were new to the house. The sale included works by Refik Anadol, Harold Cohen, Holly Herndon and Mat Dryhurst, Alexander Reben, and Claire Silver, and ran into an open letter, ultimately signed by roughly 6,500 artists, arguing that many of the offered works used models trained on copyrighted images without licenses.[18]
Other notable shows include the Whitney's Harold Cohen: AARON (2024), which reactivated several of Cohen's working AARON systems on the museum floor, and the Serpentine's long-running engagement with Refik Anadol and Anna Ridler.[2] The Centre Pompidou, the Mori Art Museum, and ZKM Karlsruhe have all collected or commissioned work from AI artists.
The U.S. Copyright Office holds that copyright requires human authorship. Its position emerged from a string of decisions on AI-assisted works.
In 2018, computer scientist Stephen Thaler submitted A Recent Entrance to Paradise for registration, listing his system, the Creativity Machine, as the sole author. The Office refused. Thaler sued, and on 18 March 2025 the U.S. Court of Appeals for the D.C. Circuit affirmed the denial in Thaler v. Perlmutter, holding that the Copyright Act of 1976 "requires all eligible work to be authored in the first instance by a human being" (D.C. Cir., 2025, https://media.cadc.uscourts.gov/opinions/docs/2025/03/23-5233.pdf).[21] Judge Patricia Millett's opinion identified six provisions of the Copyright Act inconsistent with a non-human author, including the term provisions in 17 U.S.C. Section 302(a) and the joint-authorship intent test.[21]
In September 2022, artist Kris Kashtanova registered Zarya of the Dawn, a comic created with Midjourney, without initially disclosing AI involvement. After the Office learned of her social-media posts about the workflow, it reissued the registration on 21 February 2023 to cover Kashtanova's text and the selection and arrangement of the images, but not the images themselves, which it deemed not the product of human authorship (USCO, 2023, https://www.copyright.gov/docs/zarya-of-the-dawn.pdf).[22] The Office wrote that a person who provides text prompts to Midjourney "does not 'actually form' the generated images and is not the 'master mind' behind them."[22]
On 16 March 2023, the USCO issued a statement of policy on registration of works containing AI-generated material, requiring applicants to disclose AI use and describe the human author's contribution, and explaining that text prompts alone are generally insufficient to confer human authorship (USCO, 2023, https://www.copyright.gov/ai/).[23] The Office subsequently published a multi-part report on AI and copyright, with Part 2 on copyrightability released in January 2025.
Jason M. Allen's Theatre D'opera Spatial, the Midjourney-generated work that won the Digital Arts category at the Colorado State Fair on 29 August 2022 and a $300 cash prize, was denied registration in September 2023 by the USCO Review Board, which held the human contribution was de minimis. Allen sued the Office in the U.S. District Court for the District of Colorado on 26 September 2024 (Case No. 1:24-cv-2665), arguing that his 624 prompt iterations and subsequent edits in Photoshop and Gigapixel AI amounted to copyrightable human authorship (CPR News, 2024, https://www.cpr.org/2024/09/26/colorado-springs-artist-appeals-copyright-rejection-ai-art/).[24]
A wave of class actions has tested whether scraping copyrighted images to train generative models constitutes infringement.
Andersen v. Stability AI was filed on 13 January 2023 in the Northern District of California by illustrators Sarah Andersen, Kelly McKernan, and Karla Ortiz, who later expanded the class to include Hawke Southworth, Grzegorz Rutkowski, Gregory Manchess, Adam Ellis, Julia Kaye, Jingna Zhang, and others. The suit names Stability AI, Midjourney, DeviantArt, and Runway and centers on the use of the LAION-5B dataset of roughly 5.85 billion image-text pairs to train Stable Diffusion. On 12 August 2024, Judge William H. Orrick refused to dismiss the artists' core direct and induced copyright claims and denied Midjourney's motion to dismiss the Lanham Act false-endorsement claim, writing that the plaintiffs had plausibly alleged that Stable Diffusion was "created to facilitate that infringement by design" (Loeb and Loeb, 2024, https://www.loeb.com/en/insights/publications/2024/08/andersen-v-stability).[25] The case moved into discovery, with trial scheduled to begin on 8 September 2026.[25]
Getty Images v. Stability AI was filed in the High Court of England and Wales on 16 January 2023 and separately in the U.S. District Court for the District of Delaware. Getty alleged that Stability AI had scraped roughly 12 million images from Getty's catalog to train Stable Diffusion. In the UK case, Getty narrowed its claims significantly during trial after conceding it had no evidence the training itself took place in the United Kingdom; on 4 November 2025 the High Court rejected the remaining copyright-infringement and database-right claims and found only limited trademark infringement linked to Getty watermarks reproduced in some outputs (Latham and Watkins, 2025, https://www.lw.com/en/insights/getty-images-v-stability-ai-english-high-court-rejects-secondary-copyright-claim).[26] The court reasoned that AI model weights are "not a copy" of the training images in the sense the UK statute requires. The Delaware case continued separately.
Additional suits were filed by the Authors Guild and writers against language-model developers, and by image agencies, photographers, and music labels against various AI companies through 2023 and 2024. The legal questions of fair use in the U.S. and the text-and-data-mining exceptions in the EU and UK remain unresolved at the appellate level for image models.
A research group at the University of Chicago led by Ben Zhao and Heather Zheng released two adversarial tools aimed at giving artists a technical defense.
Glaze, released in March 2023, applies imperceptible perturbations to an image so that a fine-tuning run on the modified image learns a different style than the one the artwork actually depicts. The project reported more than 8.5 million downloads since launch (MIT Technology Review, 2024, https://www.technologyreview.com/2024/11/13/1106837/ai-data-posioning-nightshade-glaze-art-university-of-chicago-exploitation/).[27] Glaze received the 2023 USENIX Internet Defense Prize and a special mention in TIME's Best Inventions of 2023.
Nightshade, released in January 2024, is a more aggressive data-poisoning tool that injects perturbations designed to corrupt models that train on a poisoned image. Where Glaze is defensive (protecting a single artist's style), Nightshade is offensive (intended to damage scraped models). Nightshade reported about 2.5 million downloads in its first year.[27]
A separate effort, Spawning, was founded in September 2022 by Mat Dryhurst, Holly Herndon, Jordan Meyer, and Patrick Hoepner. Spawning runs the site HaveIBeenTrained.com, which lets artists search the LAION datasets to see whether their images were included, and an opt-out registry that Stability AI agreed in December 2022 to honor for Stable Diffusion 3 training.[30] By the March 2023 opt-out deadline, artists had removed roughly 80 million images from the Stable Diffusion 3 training set through the registry.[30]
The most visible artist backlash unfolded on the portfolio site ArtStation in December 2022. After AI-generated images began appearing on the platform's trending page, illustrators uploaded a "NO TO AI GENERATED IMAGES" sign designed by Alexander Nanitchkov to their portfolios. The protest image dominated the front page within a week (Kotaku, 2022; PetaPixel, 2022).[28] ArtStation, owned by Epic Games, later added tags and filters for AI-generated content rather than banning it outright.
In the United States, the Concept Art Association ran a GoFundMe in late 2022 to lobby policymakers on AI and copyright. The fundraiser, which raised roughly $270,000, helped finance trips to Washington and amicus participation in subsequent litigation. The Authors Guild, SAG-AFTRA, and the Writers Guild of America have made AI training and synthetic performance use central issues in industry bargaining.
The core artist argument is that consumer image generators were trained on billions of scraped images, many of them copyrighted illustrations on which a class of working professionals depends, without licensing or compensation. Defenders of the tools argue that the resulting models do not retain copies of the training images, that the synthesis process is transformative, and that the technology lowers barriers for non-traditional creators. Both positions are reflected in the pleadings of Andersen v. Stability AI.
At a workflow level, AI tools have spread into concept art, advertising, storyboarding, book covers, and game asset pipelines. Studios increasingly require disclosure when AI is used, and some hold positions limiting it for credited work. Wizards of the Coast pulled and reissued a Magic: The Gathering card art piece in 2024 after artist complaints about apparent AI use, and Marvel briefly faced backlash for an AI-style title sequence on the Secret Invasion series in 2023. Adobe Firefly's training set was positioned as a commercial-safe alternative, though it too drew artist complaints when Adobe Stock contributors found their work in the training set.
The Colorado State Fair incident on 29 August 2022 made Theatre D'opera Spatial one of the first AI-generated images to win a fine-art prize.[19] Jason M. Allen entered the work in the Digital Arts and Digitally Manipulated Photography category and took first place in the emerging-artist division. The two judges later told The Washington Post they did not know Midjourney was a generative AI system but said they would have awarded the prize regardless (Washington Post, 2022, https://www.washingtonpost.com/technology/2022/09/02/midjourney-artificial-intelligence-state-fair-colorado/).[19] Allen, defending the win, told the paper, "I'm not going to apologize for it," and the state fair began requiring disclosure of AI use in 2023.[19]
The Sony World Photography Awards 2023 incident centered on Berlin-based photographer Boris Eldagsen, who won the Creative category of the Open competition in March 2023 with an image titled Pseudomnesia: The Electrician that he had generated using DALL-E 2. Eldagsen publicly refused the prize in April 2023, saying AI images and photographs should not compete in the same award.[20] World Photography Organisation officials said Eldagsen had disclosed the co-creation in advance; Eldagsen disputed that they understood what disclosure meant. The episode prompted several photography contests to clarify their rules (Hyperallergic, 2023, https://hyperallergic.com/german-photographer-refuses-award-for-his-ai-photo/).[20]
The Christie's Augmented Intelligence sale in February to March 2025 drew an open letter signed by more than 6,400 artists asking the house to cancel the auction, organized in part by concept artist Reid Southen, who argued that a substantial share of the works on offer used models trained on copyrighted material (CNN, 2025, https://www.cnn.com/2025/02/11/style/christies-ai-auction-open-letter-tan/index.html).[18] Christie's proceeded with the sale, which exceeded its high estimate.
Other flashpoints include the 2023 fight over Adobe Stock contributors discovering their work in Firefly's training set, the 2024 controversies around Magic: The Gathering and Wizards of the Coast art submissions, and recurring complaints when scraped models reproduce identifiable styles of named living artists (Greg Rutkowski's name was one of the most-used prompts on early Stable Diffusion, a fact he objected to publicly).