AI art (artificial intelligence art)

What is AI art?

AI art (artificial intelligence art) is any form of digital art created or enhanced with AI tools. Though commonly associated with visual art -- images and video, for example -- the term AI art also applies to audio compositions, including music.

Since the earliest pictures found on cave walls, human creativity alone has driven art's history. Inspired humans using hand-held tools -- musical instruments or paintbrushes -- generated all manner of art throughout recorded history. AI art shatters that paradigm.

Using machine learning algorithms, computer technology -- trained on a body of art to learn what art is and how to describe it -- applies various techniques, such as a generative adversarial network (GAN), to change or enhance existing human creations or generate entirely new works of art.

AI art challenges the millennia-old requirement of humans as the sole creators of art. Its introduction raises questions about the genesis of creativity and carries ethical and legal concerns. It’s also an opportunity to extend the boundaries of art -- and creativity -- in many ways.

AI art allows anyone to create works or even entire collections of art, but in a small fraction of the time non-AI methods afford. In addition, AI art can create visual or audio compositions that would be difficult to create otherwise. With text-to-image generative AI tools, such as Dall-E or Stable Diffusion, humans no longer need to attempt to draw the image they want; they simply type a text prompt into the tool, which generates the desired imagery.

History of AI-generated art

The earliest iterations of AI art appeared in the late 1960s, with the first notable system appearing in 1973 with the debut of Aaron, developed by Harold Cohen. The Aaron system was an AI assistant that used a symbolic AI approach to help Cohen create black-and-white art drawings.

AI-generated art began its recent ascent in 2014, when GANs -- a foundation of generative AI technologies -- were first discussed. In 2015, Google released DeepDream, which uses a convolutional neural network (CNN) as an experimental approach to AI art, further advancing the field.

Ganbreeder was launched in 2018 and rebranded itself as Artbreeder, using GAN models to allow humans to use AI to modify existing images and create new ones. That same year, an artist collective operating under the name of Obvious made headlines by selling a painting called Edmond de Belamy, created using GAN models, at Christie's auction house for the princely sum of $432,500. Those GAN models were trained on a corpus of 15,000 portraits from the 14th to the 19th century that were publicly available on the WikiArt website.

The public debut of text-to-image GAN-based online services for image generation sparked the imagination and interest of users around the world in January 2021. That month, OpenAI launched Dall-E, providing a publicly accessible and usable system that enabled anyone with internet access to create AI art with text prompts, giving the world a look at AI art’s possibilities.

In May 2022, Google announced its Imagen text-to-image technology as another option for AI art. This was followed in August 2022 by Stability AI, which launched Stable Diffusion's services, another GAN-based, publicly accessible option to create AI art with text prompts.

The growth of AI art tools continued in 2023, with large software vendors joining the market. Notably, the Adobe Firefly service was announced in March 2023. This GAN-based approach integrates with Adobe's popular image and video editing tools, including Photoshop and Premier.

What types of AI are used to generate art?

Though AI art uses a variety of models and techniques, the fundamental process remains the same. The first step is machine learning, during which an AI model is trained on a data set to begin to form a knowledge base. Once an understanding of a data set is established, models can begin the next step: creating and generating images. As part of an interface to the models, modern AI art tools often employ some form of natural language processing, or NLP, to understand and interpret the text users input in their request to generate an image.

Different types of AI models used to generate art include the following:

• Generative adversarial network. In a GAN, multiple neural networks are used together for deep learning operations to help predict, or generate, the end result the user seeks based on the prompt.
• Convolutional neural network. With the CNN approach, the deep learning model identifies objects, which can then be useful for generating new images.
• Neural style transfer. An NST is used in conjunction with a CNN as a deep learning technique that enables the transfer of the style of one image to another. For example, a user could use an NST to generate AI art in the style of Van Gogh.
• Recurrent neural network. RNNs are utilized for generating sequences of data like music. They use a feedback loop to produce a sequence of outputs based on prior inputs, which enables them to generate new outputs that resemble the inputs on which they were trained.

How are artists using AI?

Again, the primary tools of artists were physical items like brushes, paints, chisels or musical instruments. But the introduction of AI expands the palette of capabilities available to all artists in the following ways:

• Art therapy. Art is used for personal enjoyment and relaxation, and therapists have occupational AI art to help patients on a case-by-case basis.
• Democratization. AI enables more people than ever before to create and generate their own art, supporting a new generation of would-be artists.
• Education. Educators and teachers are using AI art tools to help teach a new generation of artists.
• Enhancing existing creations. AI tools and features help to enhance, augment and improve existing creations. For example, AI can be used to reimagine an existing piece of art in a particular artistic style.
• Fully AI-generated art. AI tools help artists create entirely new pieces of visual art, videos and music.
• Inspiration for new art. AI tools inspire artists with starting points that can lead to new pieces of art.

How difficult is it to make AI art?

Making AI art is an increasingly simple task for artists of nearly any skill level.

At the most advanced, complex level, an artist can choose to train an AI model to create art. In this approach, the artist first needs to collect or have access to a data set of art. Once the target data set has been assembled, the next step is to train the model to learn from the assembled data. With the trained data set on an appropriate GAN model, the next step is to generate the art.

It is significantly easier for an artist to use an AI tool that has already been trained on a data set of existing art. It is possible, depending on the tool, to focus additional training on an artist's own set of images to further refine the model. With the pre-trained model and any customization, the artist can then generate images. Images can be generated with text prompts, then refined after they have been generated. Some tools will allow for further generation with supplemental text prompts, while others can provide artists with additional visual design tools to fine-tune a creation. Many of the tools offer new users with free credits to explore the process of AI art.

Among the many AI image generator tools available to generate AI art today are the following:

• Adobe Firefly.
• Artbreeder.
• Dall-E.
• Deep Dream Generator.
• DreamStudio.
• Midjourney.
• Playform.
• Stable Diffusion.

What ethical concerns are associated with AI-generated art?

As explained above, AI-generated content has many positive aspects, but potential pitfalls to AI art include the following:

• Authorship. Artists have long enjoyed the practice of signing their names to works of art. But how does authorship work with AI art? An emerging ethical dilemma concerns who actually created the work: the AI, or the human who instructed the AI with the prompt to create the work?
• Bias. Any AI model is only as diverse as the data on which it was trained. There is the probability for bias if the data on which the model was trained lacks diversity and sensitivity toward equity and issues of discrimination.
• Copyright. A major concern is intellectual property theft. For example, large GAN-based AI tools for art have sometimes been trained on data sets without obtaining full legal copyright access. Because of these situations, multiple lawsuits have been filed that extend the ethical concern of AI-generated art into the legal domain. Getty Images, for example, filed a lawsuit in January 2023 against Stable Diffusion for alleged infringement on copyrights held by Getty.
• Originality. While defining what constitutes "art" has long been the subject of debate, one common attribute is that it is -- in some way, shape or form -- original. With AI art there is an ethical question about whether generated works are genuinely original or just derivative.