AI art
Last reviewed
Sources
35 citations
Review status
Source-backed
Revision
v7 ยท 7,022 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
35 citations
Review status
Source-backed
Revision
v7 ยท 7,022 words
Add missing citations, update stale details, or suggest a clearer explanation.
AI art is artwork created with the assistance of artificial intelligence systems, spanning visual images, music, video, and other creative outputs generated or co-created by algorithms, neural networks, and machine learning models. The most common form is AI image generation: text-to-image tools such as Midjourney, DALL-E, and Stable Diffusion that turn written prompts into pictures. The field spans a wide spectrum, from fully autonomous systems that produce work without human intervention to collaborative tools where artists use AI as a medium alongside traditional techniques. AI art moved from an obscure academic pursuit in the 1970s to a mainstream cultural force after 2022, when consumer text-to-image tools went public and a Midjourney image won a state-fair art prize, igniting fierce debates over authorship, copyright, training data, and the future of human creative labor. In September 2023 the U.S. Copyright Office concluded that an image generated solely from a text prompt cannot be copyrighted because its "traditional elements of authorship" are "determined and executed by the technology, not the human user."[10]
AI art refers to any creative work in which generative machine learning models play a substantive role in producing the output. In practice the term most often describes images created by diffusion models and generative adversarial networks, but it also covers AI-generated music, video, text, and 3D content. The defining technical feature is that the model, rather than a human hand, renders the final pixels, audio samples, or frames, while the human contributes prompts, parameters, curation, and post-processing. Whether that human contribution is enough to constitute authorship is the central legal and philosophical question the field raises, and courts and the U.S. Copyright Office have so far answered that fully autonomous AI output is not copyrightable while sufficiently human-directed work may be.[12][11]
The roots of AI art stretch back to the earliest days of computing. In the 1960s, artists and researchers began experimenting with algorithmic and computer-generated imagery. Pioneers like Vera Molnar used early mainframe computers to produce geometric compositions, and A. Michael Noll at Bell Labs created computer-generated visual patterns that echoed the work of abstract expressionists.
The most significant early project was AARON, a software program created by British artist Harold Cohen.[1] Cohen, who had represented Britain at the Venice Biennale and exhibited at Documenta, relocated to the University of California, San Diego in 1968, where he was introduced to computer programming.[1] In 1971, he presented an initial prototypal painting system at the Fall Joint Computer Conference, which led to an invitation to Stanford's Artificial Intelligence Lab. There, in 1973, AARON was born.[1]
Cohen's goal was to build an autonomous program capable of generating original artworks. AARON combined formal rules (such as starting in the foreground of a drawing and moving to the background) with random events to generate curved lines, straight lines, and closed figures.[1] Over the following four decades, Cohen continually expanded AARON's knowledge base, adding rules for rendering everyday objects, plants, and human figures. The program eventually gained the ability to make color decisions and apply paint using a custom-built robotic system.[1] AARON stands as one of the longest-running, continually maintained AI systems in history, and Cohen continued developing it until his death in 2016.[1] In 2024, the Whitney Museum of American Art mounted a major retrospective exhibition of Cohen's work with AARON.[2]
The modern era of AI art began in 2015 with two breakthroughs that demonstrated the creative potential of deep learning.
In June 2015, Google engineers Alexander Mordvintsev, Mike Tyka, and Christopher Olah released DeepDream, a computer vision program that uses a convolutional neural network to find and amplify patterns in images.[3] By feeding an image through a trained neural network and iteratively enhancing the features the network recognized, DeepDream produced hallucinatory, psychedelic visuals filled with swirling patterns and animal-like forms.[3] Google open-sourced the code on July 1, 2015, and the resulting images went viral, introducing millions of people to the idea that neural networks could produce visually striking, surreal artwork.[3]
That same year, Leon Gatys, Alexander Ecker, and Matthias Bethge published "A Neural Algorithm of Artistic Style," a paper that introduced neural style transfer.[4] Their technique used the internal representations of a convolutional neural network to separate and recombine the content of one image with the style of another.[4] A photograph could be rendered in the style of Van Gogh's Starry Night or Picasso's cubist paintings. The paper, first released on arXiv in August 2015 and later accepted at CVPR 2016, demonstrated that neural networks had learned rich, hierarchical representations of visual style that could be manipulated independently from content.[4]
Together, DeepDream and neural style transfer sparked a wave of experimentation. Artists like Mike Tyka used style transfer techniques to adjust the colors and textures of images, showing that a human operator could guide and shape the output of neural networks in meaningful ways.
Generative adversarial networks (GANs), introduced by Ian Goodfellow and colleagues in 2014, became the dominant paradigm for AI image generation through the late 2010s. A GAN consists of two neural networks: a generator that creates images and a discriminator that evaluates whether they look real. The two networks are trained in opposition, with the generator gradually improving its output until the discriminator can no longer distinguish generated images from real ones.
Several landmark GAN architectures pushed the boundaries of what AI-generated imagery could look like:
| Model | Year | Developer | Key Contribution |
|---|---|---|---|
| DCGAN | 2015 | Radford et al. | First stable GAN architecture using deep convolutional layers |
| Progressive GAN | 2017 | NVIDIA | Progressive training from low to high resolution |
| BigGAN | 2018 | DeepMind (Brock et al.) | Massive scale training with batch sizes of 2,048, high-fidelity ImageNet generation |
| StyleGAN | 2018 | NVIDIA (Karras et al.) | Style-based generator allowing control over image attributes at different levels [7] |
| StyleGAN2 | 2020 | NVIDIA | Eliminated common GAN artifacts, improved image quality |
| StyleGAN3 | 2021 | NVIDIA | Alias-free generation, smooth latent space interpolation |
BigGAN, published by Andrew Brock, Jeff Donahue, and Karen Simonyan at DeepMind in 2018, demonstrated that scaling up GAN training (using batch sizes eight times larger than previous models and 50% more channels per layer) produced dramatically better results. StyleGAN, released by NVIDIA researchers Tero Karras, Samuli Laine, and Timo Aila in December 2018, introduced a style-based generator architecture that gave users unprecedented control over generated images.[7] StyleGAN could produce photorealistic human faces that did not belong to any real person, leading to the viral website "This Person Does Not Exist."[7]
The GAN era reached a cultural milestone on October 25, 2018, when the Portrait of Edmond de Belamy became the first artwork created using artificial intelligence to be sold at a major auction house.[5] The portrait was created by Obvious, a Paris-based collective composed of Hugo Caselles-Dupre, Pierre Fautrel, and Gauthier Vernier, using a GAN algorithm trained on a dataset of 15,000 portraits painted between the 14th and 20th centuries.[6]
Christie's in New York offered the blurry, gilt-framed portrait with a presale estimate of $7,000 to $10,000.[5] After a bidding war between three phone bidders, an online participant in France, and a bidder in the room, the work hammered at $350,000 and sold for $432,500 including buyer's premium, a 4,320 percent increase over the high estimate.[5] It was the second most expensive lot in the sale, surpassed only by an Andy Warhol suite of screenprints.[5]
The sale generated significant controversy. Artist Robbie Barrat, whose open-source GAN code Obvious had used, questioned the collective's originality.[6] Critics debated whether the artistic merit belonged to the algorithm, the collective that selected and framed the output, or the researchers who built the underlying technology. Regardless, the sale signaled that the mainstream art market was paying attention to AI-generated work.
The GAN era produced several artists who gained recognition for their creative use of these systems:
Starting in 2021, diffusion models rapidly overtook GANs as the dominant architecture for AI image generation. Unlike GANs, which learn through adversarial competition, diffusion models learn by gradually adding noise to images during training, then learning to reverse this process to generate new images from random noise.
The shift began with DALL-E, announced by OpenAI on January 5, 2021.[18] DALL-E used a modified GPT-3 architecture to generate images from text descriptions, demonstrating the remarkable ability to combine concepts in novel ways (such as "an armchair shaped like an avocado").[18] On April 6, 2022, OpenAI announced DALL-E 2, which produced images at four times greater resolution than its predecessor (up to 1024x1024 pixels) with dramatically improved realism.[18] DALL-E 2 entered beta on July 20, 2022, and became publicly available on September 28, 2022.[18] DALL-E 3, integrated into ChatGPT in October 2023, further improved prompt accuracy and coherence.[18]
Midjourney was founded by David Holz, the former co-founder of Leap Motion, in August 2021 in San Francisco.[19] Holz and a team of 10 engineers built the Midjourney platform, which launched its Discord server in February 2022 and entered open beta on July 12, 2022.[19] Midjourney quickly became known for its distinctive artistic quality and painterly aesthetic. The company released Version 2 in April 2022, Version 3 in July 2022, Version 4 in November 2022, Version 5 in March 2023, and Version 6 in December 2023.[19] Version 7, released in April 2025, introduced a draft mode for rapid prototyping at ten times the speed and half the cost of standard generation. Holz told journalists in August 2022 that the company was already profitable.[19]
Stable Diffusion, developed through a collaboration between Stability AI, RunwayML, researchers at LMU Munich, EleutherAI, and LAION, was publicly released on August 22, 2022.[20] Its defining characteristic was openness: the full model weights were made freely downloadable under a Creative ML OpenRAIL-M license, allowing commercial and non-commercial use.[20] The model was trained on a 512x512 subset of LAION-5B, an open dataset of roughly 5.85 billion image-text pairs scraped from the web.[29] Within days, the open-source community had the model running on consumer hardware, including Windows laptops and Apple M1 Macs.[20] Downloads surpassed 10 million within two months, and a massive ecosystem of derivative projects, fine-tuned models, and community tools emerged on platforms like GitHub, Hugging Face, and Civitai.[20]
Flux, a 12-billion-parameter model from Black Forest Labs (founded by former Stability AI researchers), launched in mid-2024 and gained significant traction in 2025, rivaling Midjourney in quality while embracing open weights for broader access.
| Platform | Developer | Release | Type | Key Strength |
|---|---|---|---|---|
| DALL-E 3 | OpenAI | 2023 | Closed/API | Prompt accuracy and text rendering |
| Midjourney v7 | Midjourney, Inc. | 2025 | Closed (Discord/Web) | Artistic quality and aesthetic appeal |
| Stable Diffusion 3.5 | Stability AI | 2024 | Open weights | Customization, fine-tuning, local deployment |
| Flux | Black Forest Labs | 2024 | Open weights | Compositional accuracy, text rendering |
| Imagen 3 | Google DeepMind | 2024 | Closed/API | Photorealism, detail quality |
| Firefly | Adobe | 2023 | Closed/Integrated | Commercial safety, Creative Cloud integration |
The most publicized controversy in AI art occurred on August 29, 2022, when Jason M. Allen won first place and a $300 prize in the emerging-artist "Digital Arts/Digitally-Manipulated Photography" category at the Colorado State Fair's fine arts competition.[8] His entry, "Theatre D'opera Spatial," depicted an elaborate theatrical scene rendered with rich detail and dramatic lighting, and beat a field of 18 images submitted by 11 participants in the category.[8][9]
Allen had created the image using Midjourney, running at least 624 text prompts and revisions and producing roughly 900 iterations before selecting his three favorites.[8] He then refined the chosen images in Adobe Photoshop and increased their resolution using Gigapixel AI.[8] Allen had disclosed his use of Midjourney when submitting the work, listing the author as "Jason M. Allen via Midjourney."[8]
When news of the win spread on social media, it ignited a fierce debate. Thousands of commenters called the win unfair, arguing that using AI was a form of cheating.[8] Traditional artists expressed frustration that AI-generated work was competing in categories designed for human creativity. Allen defended his work, noting that he had spent approximately 80 hours refining and iterating on the images and stating that he "did not break any rules."[8] The two competition judges later said they did not know that Midjourney used AI to generate images, but stated they would have awarded Allen the top prize regardless; one judge, art historian Dagny McKinley, said the work reminded her of Renaissance art.[9]
In September 2023, the U.S. Copyright Office denied Allen's application to register a copyright for the image, concluding that his "sole contribution to the Midjourney Image was inputting the text prompt that produced it" and that the subsequent modifications were insufficient for copyright protection.[10] Allen appealed the decision in September 2024.
Working with text-to-image AI systems has given rise to a new creative practice often called prompt engineering. Rather than placing brush to canvas or stylus to tablet, the AI artist crafts textual descriptions that guide the model's output. This process involves several distinct stages.
Prompt crafting is the foundational skill. Artists develop detailed text descriptions specifying subject matter, composition, lighting, color palette, artistic style, and mood. Experienced practitioners learn the vocabulary and syntax that particular models respond to most effectively. A prompt for Midjourney might include references to specific artists, photographic techniques ("shot on Hasselblad"), and stylistic modifiers ("volumetric lighting," "octane render") to shape the result.
Iteration and variation follow the initial generation. Artists typically produce dozens or hundreds of variations, adjusting parameters and prompt language to explore the creative space. In Midjourney, this involves using the "vary" and "remix" features. In Stable Diffusion, it might mean adjusting the "seed" value, changing the sampling method, or altering the guidance scale that controls how closely the output follows the prompt.
Curation and selection represent a critical artistic judgment. From hundreds of generated images, the artist selects those that best express their vision. This curatorial act parallels the choices photographers make when selecting frames from a contact sheet.
Post-processing and refinement often involve traditional digital tools. Many AI artists use Photoshop, Lightroom, or specialized upscaling tools to refine their selected images, correcting artifacts, adjusting colors, compositing multiple generations, or adding hand-painted elements.
Proponents argue that this workflow constitutes a legitimate form of art direction, comparable to how a film director guides a production without personally operating the camera, designing the sets, or performing the roles. Critics counter that the gap between intention and execution is too wide, and that the unpredictable nature of AI generation means the artist cannot claim full authorship of the result.
AI art extends well beyond visual imagery into music composition and production.
AIVA (Artificial Intelligence Virtual Artist) was created in February 2016 by Pierre Barreau, Denis Shtefan, and Vincent Barreau in Luxembourg.[17] AIVA became the first virtual composer to be recognized by SACEM (Societe des auteurs, compositeurs et editeurs de musique), the French and Luxembourgish professional association of authors and composers.[17] This recognition granted AIVA the status of a composer, allowing its music to be registered with SACEM.[17] AIVA's first studio album, "Genesis," was released in November 2016. The system specializes in orchestral, cinematic, and classical composition, and is frequently used for scoring films, video games, and documentaries.
Jukebox, published by OpenAI in April 2020, was an autoregressive model that generated music (including rudimentary singing) as raw audio.[16] Built on a hierarchical VQ-VAE architecture combined with Sparse Transformers, Jukebox was trained on a dataset of 1.2 million songs (600,000 in English) paired with lyrics and metadata.[16] While it could produce coherent short passages conditioned on genre, artist, and lyrics, it had significant limitations: rendering one minute of audio took approximately nine hours, and the generated songs lacked larger musical structures like repeating choruses.[16]
The landscape of AI music generation transformed dramatically in 2023 and 2024 with the emergence of consumer-facing platforms capable of producing full songs with vocals.
| Platform | Founded | Specialty | Key Feature |
|---|---|---|---|
| AIVA | 2016 | Orchestral, cinematic composition | SACEM-registered; no vocal generation [17] |
| Suno | 2023 | Full songs with vocals | Studio workspace with stem editing, MIDI export |
| Udio | 2023 | Full songs with vocals | Inpainting tool for section-level regeneration |
| Soundraw | 2020 | Background and royalty-free music | Customizable loops and structures |
| ElevenLabs Music | 2024 | Voice synthesis and music | High-quality vocal cloning |
Suno emerged as the leading consumer AI music platform, with its v4.5 model delivering professional-quality output. Users can generate complete songs with vocals, instrumentals, and arrangements from text descriptions of genre, mood, and lyrics. Suno Studio provides a DAW-like workspace with stem editing and MIDI export capabilities.
Udio, founded by a team of former Google DeepMind researchers, differentiated itself through precision controls. Its standout feature is an inpainting tool that allows users to select and regenerate specific sections of a song without affecting the rest.
In June 2024, the Recording Industry Association of America (RIAA) filed landmark copyright infringement lawsuits on behalf of Sony Music Entertainment, UMG Recordings, and Warner Records against both Suno and Udio.[15] The lawsuits alleged that the AI music platforms trained their models on copyrighted recordings without authorization, with damages claimed at up to $150,000 per work infringed.[15] Both companies acknowledged training on copyrighted material but argued their use constituted fair use.
AI-powered video generation advanced rapidly beginning in 2023, with several competing platforms pushing the boundaries of what automated systems could produce.
Runway was founded in 2018 by Cristobal Valenzuela, Alejandro Matamala, and Anastasis Germanidis after they met at New York University's Tisch School of the Arts ITP program.[22] The company played a key role in the development of Stable Diffusion before pivoting to focus on video generation.[22] In February 2023, Runway released Gen-1 (video-to-video) and Gen-2 (text-to-video), among the first commercially available generative video models.[22] Runway's tools have been used in productions including the Oscar-winning film "Everything Everywhere All at Once" and "The Late Show with Stephen Colbert."[22] Runway Gen-4, released in 2025, introduced director-style parameters including motion brushes and camera path tools.
Pika Labs was founded in April 2023 by Demi Guo and Chenlin Meng, who left Stanford University's AI PhD program after becoming frustrated with the quality of existing AI video tools while participating in an AI Film Festival organized by Runway. Pika 1.0 launched in November 2023, and the company secured $55 million in funding within six months of founding, followed by an $80 million Series B in June 2024.
Sora, OpenAI's text-to-video model, was first announced on February 15, 2024, when OpenAI previewed examples and published a technical report.[21] CEO Sam Altman posted a series of tweets responding to user prompts with Sora-generated videos.[21] The model generated significant excitement due to its ability to produce realistic, coherent video clips from text descriptions. OpenAI provided limited access to a small group of creative professionals and safety researchers before publicly releasing Sora for ChatGPT Plus and Pro users in the U.S. and Canada in December 2024.[21] Sora 2, released in early 2025, represented a significant leap in quality, and a September 2025 update added synchronized dialogue and sound effect generation.
| Platform | Developer | First Release | Key Capability |
|---|---|---|---|
| Runway Gen-4 | Runway | 2025 | Director-style controls, motion brushes, camera paths |
| Sora 2 | OpenAI | 2025 | Photorealism, cinematic quality, synchronized audio |
| Pika 2.5 | Pika Labs | 2025 | Accessible pricing, fast generation, social media focus |
| Veo 3 | Google DeepMind | 2025 | Integrated audio generation |
| Kling AI 1.6 | Kuaishou | 2025 | Motion quality and consistency |
| Luma Dream Machine | Luma AI | 2024 | 3D-aware video generation |
The rise of AI art has triggered a wave of legal disputes and regulatory questions that remain largely unresolved. The recurring theme across U.S. rulings is that copyright requires human authorship: output a machine generates autonomously from a prompt is not protectable, while work in which a human exercises sufficient creative control may be.
The foundational U.S. copyright case for AI-generated art is Thaler v. Perlmutter. Stephen Thaler, an AI scientist, sought to register a copyright for "A Recent Entrance to Paradise," a visual artwork created by his AI system called the "Creativity Machine."[12] He listed the AI system as the author and argued that the copyright should transfer to him as the machine's owner.[12]
On August 18, 2023, the U.S. District Court for the District of Columbia ruled against Thaler, holding that "human authorship is a bedrock requirement of copyright."[12] On March 18, 2025, the D.C. Circuit Court of Appeals unanimously affirmed the lower court's decision.[13] Thaler petitioned the U.S. Supreme Court for review on October 9, 2025, but the Court denied certiorari on March 2, 2026, leaving the human authorship requirement firmly in place.[13]
Importantly, the courts distinguished between fully autonomous AI creation (which cannot be copyrighted) and AI-assisted creation (which may be copyrightable depending on the degree of human involvement).
The U.S. Copyright Office's February 2023 decision regarding "Zarya of the Dawn" established a more nuanced framework.[11] Artist Kristina Kashtanova created a graphic novel using text she wrote and images generated by Midjourney.[11] The Copyright Office granted copyright protection for Kashtanova's text and her selection and arrangement of the text and images together, but denied protection for the individual AI-generated images themselves.[11] The Office reasoned that Midjourney users do not exercise sufficient control over the AI's output to qualify as authors of the resulting images, unlike photographers who control framing, lighting, exposure, and other variables.[28]
In January 2023, Getty Images filed suit against Stability AI in the United Kingdom, alleging that Stability AI unlawfully copied and processed millions of copyrighted images to train Stable Diffusion without a license.[14] Getty's claims included primary copyright infringement, secondary copyright infringement, database right infringement, trademark infringement, and passing off.[14]
Stability AI acknowledged that "many" copyrighted works had been used to train the model and admitted that "at least some" of those images came from Getty's websites.[14] However, Getty accepted during proceedings that the training and development of Stable Diffusion did not occur in the United Kingdom, which undermined its copyright claims, and shortly before closing submissions it abandoned its primary copyright and database right claims.[30] On November 4, 2025, the High Court of England and Wales issued its judgment (Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch)), rejecting the central copyright allegation and finding only limited liability for trademark infringement.[14][30] The Court held that an AI model's trained weights are not a "copy" of the images it learned from, so importing or distributing Stable Diffusion in the UK did not constitute secondary copyright infringement; the only successful claims concerned a small number of generated images reproducing Getty watermarks.[30]
A parallel case filed by Getty Images in the United States remains ongoing.
Andersen v. Stability AI is the leading U.S. class-action brought by visual artists. Filed in early 2023 in the Northern District of California by artists Sarah Andersen, Kelly McKernan, and Karla Ortiz against Stability AI, Midjourney, DeviantArt, and (added later) Runway AI, the suit alleges that the defendants copied the artists' works into training datasets, principally the LAION dataset, without consent.[31] On August 12, 2024, Judge William Orrick issued an order that dismissed the plaintiffs' Digital Millennium Copyright Act Section 1202 claims with prejudice but allowed the direct copyright infringement, induced infringement, and trademark claims to proceed to discovery.[31] Legal trackers report the case is scheduled for trial in the Northern District of California beginning September 8, 2026.[32]
The legal landscape includes several additional high-profile cases:
In March 2023, the U.S. Copyright Office issued formal guidance titled "Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence." The guidance established that:
AI art raises deep questions about the nature of creativity and authorship. When a person types a text prompt and an AI system produces an image, who is the author? Possible candidates include the person who wrote the prompt, the developers who built the model, the artists whose work was used in training data, or no one at all.
Some philosophers and legal scholars argue that creativity requires intentionality, consciousness, and the ability to make meaningful choices. Under this view, AI systems are merely sophisticated tools, and the creative act lies in how humans choose to use them. Others contend that if the output is novel and aesthetically valuable, the process by which it was created is irrelevant to its artistic merit.
A persistent critique of AI art is that generative models can only recombine patterns found in their training data, making their outputs inherently derivative rather than truly original. Defenders of AI art note that human artists also learn by studying existing work, and that all art exists within a tradition of influence and reference. The question is whether there is a meaningful difference between a human artist who has studied thousands of paintings and a neural network that has processed millions of images.
Research on public perception indicates that while audiences acknowledge the technical sophistication of AI-generated art, they consistently rate it lower on emotional resonance and authenticity compared to human-created work.
Proponents of AI art tools argue that they democratize creative expression, allowing people who lack traditional artistic training to realize their creative visions. A person who cannot draw can now create detailed illustrations, concept art, or visual narratives. Critics respond that this democratization comes at the cost of devaluing the skills and labor of trained artists, and that the ease of generation leads to a flood of low-effort content that drowns out more thoughtful work.
The rapid adoption of AI art tools has had measurable effects on creative professionals and the industries that employ them.
A January 2024 report by CVL Economics, co-commissioned by the Animation Guild and other entertainment unions, estimated that nearly 204,000 positions across the U.S. film, television, and animation sectors would be adversely affected by generative AI over the following three years; the projection was based on a survey of 300 industry leaders conducted in late 2023.[33] Several game studios have reduced their concept art teams, using AI to generate initial designs that senior artists then refine. Freelance illustrators have reported significant drops in commission work as clients turn to AI-generated alternatives.
According to research from the Stanford Graduate School of Business, when AI-generated art enters the market, consumers benefit from lower prices and greater variety, but artists face reduced demand and downward pressure on compensation.[23] The study found that AI art tends to substitute for rather than complement human-created work in commercial contexts.[23]
A United Nations Conference on Trade and Development (UNCTAD) report examined the broader implications for creative industries globally, noting that the replacement of human artists by AI systems poses particular risks for developing countries where creative industries represent a growing share of economic output.[24]
However, the impact is not uniformly negative. Some artists have embraced AI as a tool that enhances their workflow, using it for brainstorming, generating reference images, or handling repetitive tasks. The role of the artist may be evolving rather than disappearing, with less time spent on technical execution and more on conceptual development and creative direction.
Artist anxiety about legal protection runs high: in a 2023 survey of 218 artists conducted by the platform Book An Artist, roughly 89% said they believed current copyright laws do not protect them from generative AI.[34]
Refik Anadol is a Turkish-American media artist recognized as a pioneer in data visualization and AI art.[25] His work merges art, technology, science, and architecture through media embedded into existing spaces, live audiovisual performances, and immersive installations.[25] Notable projects include "WDCH Dreams," which transformed the exterior of the Walt Disney Concert Hall with projections derived from the Los Angeles Philharmonic's 100-year archive.[25] In 2022, the Museum of Modern Art (MoMA) in New York acquired his piece "Unsupervised," making it one of the first major AI artworks to enter a major museum's permanent collection.[27] "Unsupervised" used a StyleGAN2 model trained on 138,151 images from MoMA's collection and, during its run from November 2022 to April 2023, drew more than three million visitors.[35] Anadol co-founded Dataland, a 20,000-square-foot museum dedicated to AI art, scheduled to open in spring 2026.
Holly Herndon, an American musician and sound artist, created a high-fidelity AI model of her own voice in collaboration with Mat Dryhurst. This "vocal twin," called Holly+, can take any incoming audio and re-render it in Herndon's voice. She made the tool available online, allowing anyone to create music using her AI-cloned voice. To address the ethical implications of vocal deepfakes, the project is governed by a decentralized autonomous organization (DAO). When someone creates a song using the AI voice, the community votes on whether it qualifies as an official Holly+ work. Revenue from approved works is split: 50% to the creator, 40% to the DAO, and 10% to Herndon.
| Artist/Collective | Medium | Notable Work or Contribution |
|---|---|---|
| Harold Cohen | Software/painting | AARON (1973-2016), one of the longest-running AI art systems [1] |
| Obvious | GAN imagery | "Portrait of Edmond de Belamy" (2018), first AI art sold at Christie's [5] |
| Mario Klingemann | Neural networks | "Memories of Passersby I" (2019), sold at Sotheby's |
| Robbie Barrat | GAN imagery | Trained GANs on art historical datasets; collaborated with NVIDIA |
| Sofia Crespo | Neural networks | AI-generated speculative natural history and biological forms |
| Anna Ridler | GAN/dataset art | Hand-crafted datasets as artistic medium (e.g., tulip photographs) |
| Tom White | Neural networks | "Perception Engines" series exploring machine vision |
| Gene Kogan | Creative coding/AI | Educational initiatives and generative art |
AI art has found commercial outlets through both traditional and emerging platforms:
AI art has achieved growing institutional validation:
The following table provides an overview of the major tools and platforms used for AI art creation across different media:
| Tool | Type | Developer | Access | Primary Use |
|---|---|---|---|---|
| Midjourney | Image generation | Midjourney, Inc. | Subscription (Discord/Web) | Artistic image creation |
| DALL-E 3 | Image generation | OpenAI | API/ChatGPT subscription | Prompt-accurate image creation |
| Stable Diffusion | Image generation | Stability AI / Community | Open source (free) | Customizable local generation |
| Flux | Image generation | Black Forest Labs | Open weights | High-quality open image generation |
| Adobe Firefly | Image generation | Adobe | Creative Cloud subscription | Commercially safe image generation |
| Sora | Video generation | OpenAI | ChatGPT subscription | Cinematic video from text |
| Runway | Video generation | Runway AI | Subscription | Professional video editing and generation |
| Pika | Video generation | Pika Labs | Freemium | Quick video generation |
| Suno | Music generation | Suno, Inc. | Freemium | Full songs with vocals |
| Udio | Music generation | Udio | Freemium | Precision music generation |
| AIVA | Music generation | AIVA Technologies | Subscription | Orchestral and cinematic scoring |
| Artbreeder | Collaborative generation | Artbreeder | Free/Premium | Image blending and exploration |
| NightCafe | Multi-model generation | NightCafe | Credits-based | Community-driven AI art creation |
| ComfyUI | Workflow builder | Community | Open source (free) | Node-based Stable Diffusion workflows |
Supporters of AI art highlight several contributions to culture and creativity:
Critics raise several persistent objections:
The question of whether AI-generated outputs constitute art continues to divide opinion. Traditionalists argue that art requires human intentionality, emotional expression, and mastery of a craft. Proponents counter that the definition of art has always expanded to embrace new technologies and methods, from photography in the 19th century to video art in the 20th. Marcel Duchamp's readymades demonstrated nearly a century ago that the act of selection and presentation can itself be an artistic gesture, an argument that resonates with how AI artists curate outputs from generative models.
AI art continues to evolve rapidly. Several trends are shaping its trajectory: