AI-generated art, AI Art and Generative Art are all forms of digital art that are created using algorithms and code at their core. However, there are key differences between these approaches and the tools used for art creation that are worth exploring more in depth. With the advent of capable AI models in recent months the terminology around "AI Art" and "Generative Art" have become quite messy. You´ll see the terms “Generative AI”, “Code Art”, “Computer Art”, “Generative Art” - themselves difficult to differentiate on first glance - often used interchangeably, especially with the broad attention the current AI hype receives.
This article aims to define the following three approaches of computer aided creation processes by focussing on their differences: AI-generated art - meaning images generated by recently published text-to-image, large scale models; AI Art, i.e. artworks created using custom machine-learning algorithms, models and data sets; and Generative Art using code, mathematics and randomness as its medium to create visual artworks.
Before delving into the details, a high level difference can be extracted from the amount of human involvement in the process leading up to the final image. With AI-generated images, the user's input to the system is bound to the lexical domain. The way of interacting with current AI tools is a text prompt. Everything else is entirely managed by the AI image generator, pre-trained on huge sets of images from all over the internet. The user has no way of intervening in the intermediate stages of the image generation, other than refine the text prompt and start over.
AI Artists use similar algorithms in their processes, though they interact with them several levels deeper. By providing their own training data, taking care of the training processes and developing or customising the algorithms and models they use themselves, AI artists operate on the fine line between engineering, science and art. Their process yields much more control over the output and is often a consequence of extensive research and intentional testing.
Generative artists, often referred to as creative coders, use code, mathematical functions and randomness as their artistic medium. They develop programs that contain sets of instructions for the machine to generate visual outputs, the artworks, not unlike a traditional software developer who writes programs to solve some problem. Even though deliberately introduced randomness in the code is often part of generative art systems, a generative artist has in principle full control over the underlying processes and every pixel of the output.
Multiple images created with DALL-E 2 by OpenAI: "A three-part image depicting the distinctive characteristics of AI art, AI-generated Art and Generative art"
Let us look at the key characteristics, motivations, controversies and underlying processes of each in more depth:
What you have probably heard of in the news: “AI-generated art” describes images produced by one of the many publicly available text-to-image AI tools like Midjourney, Stable Diffusion or DALL-E, using text prompts as input. A text prompt can be something simple and vague like "A forest in the style of Vincent van Gogh" to paragraphs of super specific instructions and lists of keywords regarding the style, content and inspirations the model should take into account.
Underneath the hood, these image generators are large scale machine learning models, trained on huge image data sets mostly scraped from the internet. In the process of training these systems, they learn the key characteristics and elements of the training images and encode these into millions to billions of parameters, each with individual weights. This process is often compared to a very simplified model of the brain, with neurons firing on specific conditions, hence the term “neural networks” for the algorithmic architecture of these models and the inclination to attribute “artificial intelligence” to those models.
Created with DALL-E 2 by OpenAI: "A forest at sunset painted in the style of Vincent van Gogh"
When given a text prompt like the above, the model computes a new image, based on probabilistic parameter constellations defined by the input. This process, because of the necessary complexity of the model, is opaque to the user and even to the developers. A machine learning model like GPT-3.5, powering the now popular ChatGPT for example has 175.000.000.000 individual parameters. These AI systems are thus practically 'black boxes', i.e. the user has no real insights into the underlying processes. The results may or may not be what they had in mind, but why it behaves like it does is not answerable in a concise way.
However, if you think of these models as tools, as is Photoshop for image manipulation for example, their usage can to a certain extent be learned and thus can and will be used to create uniquely creative and interesting art. Since current models are taking mostly text prompts as input, a user that spends a lot of time experimenting with different key words and phrases might get some kind of control over the output. Such a proficient user is called a "prompt engineer", and it is already discussed as a new profession within the broader field of creative work, as more and more companies and people try to harness the power of the AI tools available right now.
Created with DALL-E 2 by OpenAI: "A robot drawing a robot drawing itself on paper. Illustration"
While some of the criticism regarding “AI-generated art” that arose in recent months, come from a certain "fear of replacement", there are some valid points of criticism worth mentioning:
The main argument against these specific AI tools is concerned with the way the companies behind the models built their training data sets. They collected and used images available on the internet, including works from artists and art-sharing sites like Deviant Art, stock photo sites and archives, without notice or explicit permission and most importantly without a way for the creators to opt out of their work being included. While this kind of data-collection and usage is legal for certain scientific purposes, it becomes very questionable, when corporations start to make significant money on that basis.
Quick side note: I´m not saying the development of these models is a small feat or bad in itself. In fact, it is impressive work and a huge step forward in the field of artificial intelligence research. However, the power of these models significantly depends on the huge data sets and the resources needed to train a model on that data. It seems to me that unlinking the models and AI tools built upon them from the training data used, is neither reasonable nor accurate. A pre-trained model, by definition, is the result of its algorithmic architecture combined with the training data.
Side by side comparison between Girl with a Pearl Earring (1665), by Johannes Vermeer (left) and an AI generated variation by DALL-E2 (OpenAI)
Another point of critique is the low skill requirement of using an AI text-to-image model to create an image. While this can be liberating in some sense, it unsurprisingly upsets a lot of artists and craftswomen, who put thousands of hours into honing their craft and expertise, just to see their feeds suddenly fill up with AI-generated images that took a second to create. As I interpret this line of argument, it is following the suspicion that the nearly non-existent barrier of entry results in creation devoid of intent or deeper meaning. The process feels “hollow”. And it is - mostly - though it definitely does not have to stay that way. But to make that leap it most certainly needs artists that explore, experiment, combine, deconstruct and recontextualize the outputs and the systems themselves..
With a low barrier of entry comes the issue of false labelling. To the untrained eye, it can be difficult to differentiate between an AI-generated image and a handcrafted (digital) artwork, pushing the doors for grifters and scams wide open. Though most AI image generator companies make it mandatory to label the outputs as "AI generated", with DALL-E most notably adding a visual watermark to all generated images, not everyone will abide by these terms and conditions and it will be next to impossible to go after every single offence. You would feel quite upset when your favourite fitness influencer turns out to be just a very skilled image manipulator, while telling you how hard they train - including the little sting to your self-worth with each perfect-body-image they post. The same holds for people using or even selling AI generated images that omit communicating that crucial information.
Although even appropriate labelling seems not enough if people are ignorant about existing AI tools, as Jason Allen’s AI-generated work, “Théâtre D’opéra Spatial” demonstrated, when it took first place in the digital art category at the Colorado State Fair´s art competition in mid-2022. He submitted the work under “Jason M. Allen via Midjourney”, thus clearly stating the tool he used to generate the image. But as the organisers told The New York Times: "The two category judges did not know that Midjourney was an A.I. program, [...] but both subsequently told her [the organiser] that they would have awarded Mr. Allen the top prize even if they had."
The discussion whether AI-generated images are or can be art, is way beyond the scope of this article. Personally I think the current outputs of AI image generators are almost exclusively experiments with a very new toolset by artists and everyone with access alike. The dimension of deliberate usage in an artistic sense seems to be missing from most AI-generated images we are seeing right now. New tools are always accompanied by heated and dismissive discussions - in the end people adapt, experiment and incorporate these tools in their work, and often artists lead the way in that process, using creativity and imagination.
The experts way: AI Art
Ivona Tau "VISIONS reflected", 2018, (Image courtesy of Ivona Tau)
AI Art, in contrast to AI-generated art, refers to artworks produced by artificial intelligence algorithms and models, the artists themselves trained and sometimes even programmed from scratch. The technical part of the artistic process can range from creating, labelling and preparing their own training data, the training itself, understanding and modifying existing machine learning models and algorithms on a source code level to building their own algorithms from scratch. Each of these steps requires deep knowledge of how different machine-learning models and architectures work, the ins-and-outs of training, manual weighing of parameters and a lot of time, hardware or money, to do the actual training.
An AI artist does technically similar work to the companies behind the above described AI image generators, but with their own acquired knowledge, skills, training data, hardware and specific artistic vision. To differentiate between the two further, AI artists are interested in using the tools of machine-learning and artificial intelligence algorithms to create a specific work, in contrast to developing tools that function in a very general variety of contexts.
AI artist Dr. Ivona Tau for example, trains machine learning models on image sets of her original photography to create scenes that constantly shift between discrete familiar objects and ambiguous abstract shapes. In this human-machine collaboration, the AI model becomes a medium to explore visual spaces that would be impossible to create otherwise. This approach makes the curation of the training data and the input of the model in conjunction with deeper knowledge about the inner workings of the system a big part of the overall artistic process.
Conceptually, AI Art is often exploring the relationship between machine generated imagery and human perception, since on a high level the medium of machine-learning algorithms seems to share structural similarities with our model of how our brains perceive and interpret the world around us. Adjacent topics like theories of creativity, authorship, human and machine consciousness (theory of mind) and the visual latent (in-between) spaces combined with a fascination for the technical and mathematical underpinnings of machine-learning technology often provide the starting point of artistic explorations in this field. Interestingly, to me this makes a lot works of AI Artists very introspective and self-reflective from a human point of view.
In her creative process, Ivona Tau uses Python and generative adversarial networks (GAN) to train artificial intelligence on custom sets of photography works. (Photo courtesy of Ivona Tau)
In my opinion, it is unfortunate that in the heat of the current debate around AI-generated art, there is often no room for distinction between users of AI tools and AI artists building their own. At the latest since the concept artists of the 60s, the process and context behind a work of art can not be left out of the conversation, especially if the visual aesthetics share similarities. It can be debated how much original work has to go into the creation of an artwork, to be 'original' - does a painter have to make their own paint? - to which the answers are cultural and subjective, but the differences in process, approach and concept between using a text-to-image AI tool and the work an AI artist does is clearly great enough to justify this distinction.
Generative Art, though more of an umbrella term for art that includes the use of some kind of autonomous system in the process, describes art, where code is used as the creative medium. Other, in my view more succinct terms for this artistic practice are "algorithmic art", "code art" or "creative coding".
Generative art in the form of Creative Coding requires the artist to be a competent programmer to use their medium - the programming languages and frameworks they use - expressively. Since the artist creates the generative system themself, they have full control over the process and space of possible outputs. In generative art, the artist codes the algorithms, rules and parameters that generate the artwork and can tweak every detail to their artistic vision. Unlike in AI art, which always involves some kind of 'black box' processes, the generative artist can in principle explain the creation process of an artwork step by step.
Fidenza #483, #545 and #10 by Generative Artist Tyler Hobbs, 2021 (Image courtesy of Tyler Hobbs)
Having said that, a lot of generative artists are interested in the intersection of that control and deliberate unpredictability. Generative artists often incorporate elements of chance and randomness in their algorithms and meticulously fine-tune the boundaries in between which the parameter values can be randomised. This can result in generative systems that create some truly unique and unpredictable artworks within the space of possible parameter configurations the artist programmed into their work. It also gives the artist the ability to generate a large number of artworks from the same system, of which each variation is a unique, one-of-a-kind piece.
In that sense one could say: Generative art is more concerned with designing processes using code as it´s creative medium, than generating one singular final artwork. In fact generative art began as a form of conceptual art in the 60s, where artists such as Sol LeWitt and Vera Molnár began to write instructions - i.e. algorithms - they followed themselves to create drawings, paintings and sculptures. A great accessible example for this approach to art are Sol LeWitt's "Wall Drawings", that only consist of a set of instructions on how to execute a specific painting, but leaves the interpretation and execution of those instructions to whomever does the actual drawing for an exhibition or instalment.
Sol LeWitt (B. 1928, Hartford Connecticut): Wall Drawing 630, January 1990 (left)
A wall is divided horizontally into two equal parts. Top: alternating horizontal black and white 8-inch (20 cm) bands. Bottom: alternating vertical black and white 8-inch (20 cm) bands. India ink
Wall Drawing 614, July 1989 (right)
Rectangles formed by 3-inch (8 cm) wide India ink bands, meeting at right angles. India ink
On display in the exhibition Sol LeWitt: A Wall Drawing Retrospective (Photo courtesy of MASS MoCA)
With the advent of computers in the 1960s, many early generative artists started to translate their conceptual algorithmic works into computer code and let the machine create the output. The “machine imaginaire” became the “machine existante”. Artists like Georg Nees, Frieder Nake, Herbert W. Franke and the aforementioned Vera Molnár are considered among the first to use computers for purely artistic endeavours. The machine became an important part in the process and a partner in the realisation of their artistic vision.
Vera Molnár 'Hommage à Dürer 91 variations', 1982”, reconstructed by the author.
The true beauty of generative art and code as it´s creative medium, lies in its unpredictability once a certain level of complexity is reached, while staying deterministic and transparent at the same time. At every step, the artist can in principle reason about what is happening and why, although the process of creative coding is often very explorative and driven by curiosity and “happy accidents” more than by planning and following strict guides.
With the recent boom in publicly available AI systems like Midjourney, DALL-E and Stable Diffusion, it is important to know the differences between AI generated art that was made with one of the pre-trained AI models by companies like OpenAI using text prompts as input, AI art that was made with a custom AI models, trained on self created images, and Generative art, where the artist uses their own code, maths and algorithms to generate the artworks.
Regardless of which style one prefers and the importance one lies on the underlying processes and tools used, it's clear that all forms of AI and Generative art are pushing the boundaries of what is possible in the world of (digital) art. We are surrounded and influenced by software of all forms in our daily lives and the artistic reflection on these systems is vital to understanding and imagining better systems.
AI and Generative Art are sure to continue to evolve and grow in exciting new ways in the years to come, as will the debates around its ethics and artistic values. Nevertheless, artists, recipients and collectors should know the differences of these approaches to art making, to not fall for scams, “Terminator” journalism and undifferentiated hot takes and to be able to form an opinion on their own.
Sources and further reading