AI Art
AI art is visual work, mostly still images, made with machine-learning generative models, either by sampling from a trained model, by steering one with text prompts and reference images, or by training a model on a chosen body of work and treating the result as an artistic medium. It is distinct from earlier algorithmic art, in which the artist wrote explicit rules, because the models here learn visual structure from huge image collections rather than being programmed with a style. The term covers a fine-art lineage in which artists curated datasets and trained their own networks, and a much larger consumer phenomenon in which anyone can type a sentence and receive an image within seconds.
The first wave of recognized AI art used generative adversarial networks (GANs), introduced by Ian Goodfellow and colleagues in 2014. A GAN trains a generator to produce images and a discriminator to distinguish them from real ones, and the competition pushes the generator toward realism. Artists such as Mario Klingemann, Anna Ridler and Refik Anadol worked with GANs in the late 2010s. The moment that brought the field to mainstream attention was the October 2018 auction at Christie's of Portrait of Edmond de Belamy, made by the Paris-based collective Obvious using a GAN trained on historical portraits. It was estimated at roughly 7,000 to 10,000 US dollars and sold for 432,500 US dollars, and it prompted a long argument, including about how much of the underlying code came from Robbie Barrat's open-source work, over credit and novelty.
The 2021 and 2022 text-to-image systems changed who could make AI art, and the pace of release was compressed into about eighteen months. OpenAI announced DALL-E in January 2021 alongside CLIP, a model linking images and text that artists soon used to steer other generators, and followed with DALL-E 2 in April 2022, which used a diffusion approach and produced markedly sharper results. Midjourney, a small independent lab accessed through Discord, entered open beta in July 2022. Stable Diffusion, released by Stability AI in August 2022 and based on the latent diffusion research of Rombach and colleagues at LMU Munich and Heidelberg, was open-weight, so it could run on a consumer graphics card and be fine-tuned by the public, which produced a large ecosystem of derivative models and tools. Its training relied on the LAION image-text datasets scraped from the web, which later became the focus of consent and copyright complaints.
The Colorado State Fair episode of 2022 showed how unsettled the norms were. Jason M. Allen entered an image titled Théâtre D'opéra Spatial, produced with Midjourney and then refined and upscaled, in the digital arts category of the state fair's fine arts competition and won first prize in August 2022. When this became public in early September, many artists objected, while the judges said they had judged the work as presented and Allen said he had disclosed Midjourney in the entry. The episode was an early public argument over disclosure, categories and what a prompt contributes, and the US Copyright Office later declined to register the work as submitted because it contained Midjourney-generated material that was not disclaimed, a decision Allen contested.
Legal and economic disputes are ongoing and are not settled by any single case. In January 2023, the artists Sarah Andersen, Kelly McKernan and Karla Ortiz filed a class action against Stability AI, Midjourney and DeviantArt over training on their work, and Getty Images sued Stability AI the same year, with proceedings in the United States and the United Kingdom. Illustrators and concept artists have reported lost commissions, some platforms have restricted AI-generated submissions, and tools such as Glaze and Nightshade from the University of Chicago were released to let artists perturb images against style mimicry, with their effectiveness contested. Technical limits persist as well: text rendering, hand anatomy, counting and consistent characters across images were long-standing failure modes, improving but not eliminated.
A common misconception is that the model stores and collages existing pictures; it learns a statistical representation, although rare memorization of duplicated training images has been shown, and style imitation can still be close. The opposite misconception, that prompting involves no skill, overlooks composition, iteration, inpainting and editing. The more durable distinction is between work that treats the model as a medium with intent and curation, and output used as a cheap substitute for commissioned illustration.
Read AI art as a new image-making medium and a new production pipeline at once: GANs gave it a gallery debut, diffusion models gave it mass access in 2022, and the questions it raises about consent, credit and category are still being decided in courts, competitions and studios.