VII. Creativity is in the (A)I of the Beholder: Artificial Intelligence and the Visual Arts

Sold for over forty times its asking price, the Portrait of Edmond de Belamy stunned the art world when it reached Christie’s auction block. Signed with the algorithm that generated the portrait, Edmond de Belamy was the first major work of artificial intelligence (AI) art sold at a major auction house. ¹ Put forth by the artist collective Obvious, the 2018 entrée of Portrait of Edmond de Belamy captivated the cultural arts sector, sparking debates over how to conceptualize AI-created work as art and over the role of AI within the creative process. However, although many individuals are contemplating the use of AI for the first time, the practice has existed for decades stretching back to the late 1960s. ²

Within this paper, I will discuss the rise of artificial intelligence within the cultural arts sector beginning by defining artificial intelligence itself and by briefly contextualizing the history of AI. I will then delve into the recent popularity of artificial intelligence within the visual arts sector sparked by the record-shattering auction of Portrait of Edmond de Belamy. Once I have established an appropriate framework for the current position of AI within the cultural arts, I will then turn towards how the arts industry conceptualizes and understands work generated using AI as art, focusing on the perspectives of artists and of casual art viewers. After establishing how people view the output from AI as art, I will examine the role that AI plays within the creative process using prominent artists who work within AI as examples. While I argue that AI can inhabit a creative role outside of a simple tool for artists to use, it is the human artist that determines the role of AI within the creation of visual art.

The Rise of Artificial Intelligence

To begin, I would like to establish the definition of artificial intelligence that I use within this paper. Artificial intelligence according to cultural data scientist Vishal Kumar is when “computational tools start to possess cognitive abilities,” or to put it simply, AI is when technology can copy or perform tasks that only humans could previously do. ³ Those who conceive of artificial intelligence as the prescient Hal from Kubrick’s 2001: A Space Odyssey might be surprised to find everyday staples such as Apple’s Siri, Amazon’s Alexa, and even Spotify’s music recommendations listed among everyday AI technology.⁴ AI pervades our modern world in many ways, but the concept of anthropomorphic entities capable of performing human tasks has existed long before today’s era of digital assistants. Many credit the mythological Jewish golem (a clay creature brought to life by its creator) as the conceptual forerunner of artificial intelligence, while others also claim with Mary Shelley’s Monster from Frankenstein as an early inspiration.⁵ By creating the first algorithm, Ada Lovelace laid the groundwork for the future field of AI in the mid-1840s, while Alan Turing’s mid-20th century Turing Test (wherein he attempted to determine whether a computer could pass as a human being) forced society to consider the possibility of computers possessing indistinguishable human characteristics.⁶ While these developments formed the framework for the study of AI, it was not until 1956 that artificial intelligence became an established field, and not until 1968 that AI entered the cultural arts sector via Harold Cohen’s experiments with algorithmic art.⁷

While artists like Harold Cohen worked with artificial intelligence for decades in relative obscurity, the artist collective Obvious shot straight to international fame with Portrait of Edmond de Belamy. Although many established AI artists criticized Obvious for using an opensource algorithm rather than creating their own algorithm, the auction at Christie’s did have the noted benefit of bringing other AI artists into the spotlight.⁸ On the heels of Obvious’ success came Mario Klingemann’s Memories of the Passerby I – the second work of AI art to be sold at a major auction house in 2019.⁹ A prominent member of the AI art world, Klingemann’s success heralded the true arrival of AI art into the mainstream. Faceless Portraits Transcending Time – the first solo gallery exhibition devoted to an AI artist (and organized by the AI artist’s human counterpart) – became the next indicator of the success of AI art.¹⁰ With the significant rise in visibility of this sector of the art world, it becomes crucial to establish how the cultural arts sector can conceive of these works as art and to debate the role of AI within the creative process itself

How Do We Understand AI Art?

While it may seem simplistic, one question pursuing AI artwork is how to think about it as art? While there is an openness to accepting the work generated using AI as a legitimate artform, the technical complexity of the practice as well as the semi-nebulous role of the technology within the creative process can result in confusion for the unfamiliar. To understand how AI art fits into the cultural arts sector, one must first understand a little of how artists who work with AI often use it within their practice. While not employed by all AI artists, much of the artificial intelligence in this art uses neural networks such as GANs. Similar to the human brain, neural networks (or neural nets) are algorithms designed to process datasets and recognize patterns within that data.¹¹ A GAN, or generative adversarial network, is a two-part neural network. While the first part of the neural network (often known as a generator) creates output in an effort to replicate the dataset it received, the second part of the neural network (the discriminator) attempts to find the difference between its own output and the initial dataset.¹² The GAN will continue to produce output until the discriminator can no longer distinguish between the initial input and its own produced work. In the case of AI art, an artist may feed a large dataset composed of thousands of Renaissance portraits to a GAN. The GAN would then generate an infinite number of its own Renaissance-style portraits, editing itself to align as closely with the dataset portraits as possible - a process similar to how Obvious produced Edmond de Belamy. A recent derivative version of a GAN, a CAN (or creative adversarial network) is also a two-part neural network fed by a dataset. However, while one network of a CAN encourages the output to be similar to the aesthetics of the dataset, the other network actively discourages the CAN from replicating the style within the dataset.¹³ This difference aims to encourage greater creativity within the CAN’s output. Artists using a GAN or similar neural network can adjust their algorithms and curate both the datasets fed to the algorithm as well as the output to exert control over the final works.¹⁴ While not all AI art employs neural networks like GANs, these networks are one of the most common methods within the field and illustrate the complicated cognitive tasks that AI performs.

However, understanding the complexity of the algorithmic system involved in generating AI art is just the first step towards understanding this discipline within the broader arts sector. The next step is understanding how to measure the artistic merit of a work so that there is an objective working definition of what art is. Some follow the model of Marcel Duchamp wherein art is primarily preoccupied within itself and its artistry lies within whether it is innovative and expands the field of art.¹⁵ For those not willing to accept that art is what you make of it, there are several mechanisms to determine the artistic validity of a piece beyond skill and creativity: the intentions of the artist, whether the work merits and receives display within institutions of fine art, acceptance of the work as art by the casual viewer.¹⁶ It is safe to assume that the artists working with AI intend to produce artwork, and the discussion above provides several examples of exhibitions and auctions of AI art within the cultural arts sector. It is the final qualification regarding whether the casual viewer accepts the work as art that holds the next step towards conceptualizing the place of AI art within the larger cultural arts framework. However, beyond just acceptance, it is crucial to examine how AI artists and how casual viewers conceive of the work produced by AI as art.

When discussing how their work fits into the larger narrative, some AI artists, such as Leonel Moura, ascribe to Duchamp’s broad and inclusive definition of art while also relying upon validation from the arts sector to define AI art’s significance.¹⁷ Other artists, such as Trevor Paglen, describe the practice of AI art as similar to that conceptual art.¹⁸ Within conceptual art, artists believe that the idea is the true art and do not believe the idea must be made manifest for the art to exist. Conceptual artists locate the creation of their work within their mind rendering the physical unnecessary. This aligns with how many AI artists believe the true art of AI occurs within the neural network and algorithmic process rather than the output.¹⁹ While the art world did not accept the conceptual art movement until the 1970s, its ultimate inclusion allows the cultural arts sector to conceive of AI art using a similar framework.

While understanding how AI artists view their art undoubtedly informs the general understanding of AI art as a practice, it is also crucial to examine how casual viewers construct their own framework. Although it is undeniable that some people possess a bias against AI that renders them unwilling to accept that AI could ever produce art, there are objective methods to determine how casual viewers without this prejudice view AI art.²⁰ Many researchers have employed an updated interpretation of the Turing Test to test the casual viewers’ perspective on AI art. Initially conceived of by AI forefather Alan Turing, the Turing Test had a participant ask questions of two respondents and determine which of the respondents was human and which was a computer.²¹ If the computer could fool the participant into thinking it was human, Turing considered it a successful piece of AI technology since it was able to pass as a human. Adapted for contemporary AI art, the modern Turing Test asks viewers to determine whether an artwork was created by a human artist or by AI. The viewers’ inability to distinguish between a solely human-created piece of art vs. a work generated using AI illustrates that casual art viewers are capable of conceiving of AI art as fitting within the current artistic canon.

Multiple studies have employed a variation of this modern Turing Test to evaluate how casual viewers appreciate, assess, and process AI art. Within a study conducted by Ahmed Elgammal, director of The Art & Artificial Intelligence Lab at Rutgers University, 75% of visitors to an exhibition containing both AI and human-generated art believed the AI art to be the work of a solely human artist.²² Another study conducted by Rebecca Chamberlain, a psychologist who focuses on the psychological and neuroscientific basis of artistic perception, asked participants whether they thought certain works of art were attractive without the participants being aware that some of the art was AI art. Ultimately, the study concluded that viewers were unable to tell the difference between the two types of art and found them both attractive.²³ However, within a study conducted by scholars Joo-Wha Hong and Nathaniel Ming Curran, participants were asked to rate the artistic value of multiple artworks and were aware which artworks were AI artworks and which were solely human-generated. While the study concluded that casual viewers did not believe the AI art held the same artistic value as the non-AI art, it also concluded that some of the participants believed that AI could not create art regardless of the circumstances.²⁴ Therefore, those participants with this bias were unable to truly evaluate the artistic value of the AI art as their mental framework precluded the possibility of AI possessing artistry, thus casting doubt over the study’s overall conclusions. When taken together, these studies show that there is a definite subset of casual viewers who are unable to integrate the concept of AI art within the current cultural arts framework. However, outside of those who share this prejudice, even casual viewers who did not find the AI art to be of the same standard as solely human-generated art were able to evaluate and process the work as art using objective methodology around artistic and aesthetic value.

The Role of AI within AI Art

With a framework to understand AI art in place, we fall to what AI artist Leonel Moura considers the true controversy within the field – whether artificial intelligence can be creative, or if it is merely a tool in the hands of human artists.²⁵ The answer to this question lies within the juncture of the artist, the machine, the art, and the process. However, before addressing the role that AI plays within the artistic process, we must address the significance of creativity and establish how to objectively evaluate it. Creativity is the ability of an entity to produce new ideas or outputs based on its singular ability or imagination. It is “an uncensored, associative form of thought” that can both align with rules while also intentionally deviating from them.²⁶ Without creativity, any human or system is merely a tool – an algorithm becomes the same as an artist’s pastels or clay. Beyond the traditional measurements of the novelty and the quality of a work, the ‘creativity tripod’ of appreciation, imagination, and skill provides an objective scale to measure whether artificial intelligence can be creative.²⁷ Within this tripod, appreciation requires a system to critically evaluate its own work; imagination is the ability to produce unique outputs with intention and meaning; skill asks that the AI’s outputs possess quality and can be recognized as belonging to the intended group.²⁸ In the case of a GAN fed a dataset of Renaissance portraiture, the GAN would possess creativity if it intended to create and then created a portrait that viewers could recognize was of good quality and belonged within the category of Renaissance portraiture. Furthermore, the machine would then have to evaluate whether this portrait was better (or more aesthetically pleasing) than a previous portrait or future portrait it created. While not all AI is capable of meeting these benchmarks, these standards allow the cultural arts sector to establish that AI can possess creativity and to evaluate which AI systems have it.

Even though AI systems are capable of creativity, it is the human artist that determines the role AI plays within the artistic process, including whether that role is creative or not. Within the AI art sector, there is an entire spectrum regarding how artists use AI’s creative potential. At one extreme of this spectrum is AI-artist Anna Ridler who views the AI she works with as merely another tool to extend her practice.²⁹ While she acknowledges that the machine she uses is autonomous in certain ways and enables her to do things she wouldn’t normally be able to do, she asserts that the machine would not be able to do anything at all without her direction.³⁰ Thus, she rests firmly with those who do not endow AI with any creativity. On a similar end of the spectrum is Harold Cohen and his AI-machine AARON. Cohen compared his decades-long working relationship with AARON to a cyborg “in the sense of having computational implants in the brain, only [his] implant is sitting on [his] desk.”³¹ While Cohen worked with AARON for over fifty years, he was adamant that the machine possessed no creativity of its own.³² He believed that the artistic potential of AI was within its work with humans and that an AI machine’s inability to either question or break rules negated its creative potential.³³

At the other end of the spectrum, Leonel Moura believes his AI “artbots” possess a certain degree of creativity since they can produce pictures that he could not have predicted.³⁴ While he does not believe that AI systems can be artists in their own right, Moura likens his relationship to his artbots as that of a student and teacher. Just as an English teacher would not take credit for a successful novel created by one of their students, Moura believes the art produced by the artbots is unique and the result of the artbots’ own creative ability.³⁵ However, he maintains that “the will and skill” of these systems remains with the human artist.³⁶ Ahmed Elgammal, director of the Art and Artificial Intelligence Laboratory at Rutgers University, takes Moura’s attribution of creativity a step further and considers his system AICAN a creative collaborator.³⁷ Elgammal is the creator of the first CAN and believes that by feeding AICAN an uncurated dataset (at present it includes 80,000 images from five centuries of art), AICAN executes an inherently creative process since it uses its own creativity to draw inspiration from a random sampling of artwork.³⁸ In Elgammal’s view, it is the human artist who sets up the conceptual and algorithmic framework, but it is AICAN that is the creative lead in terms of the elements it includes and the principles of the art it generates.³⁹ Considering AICAN a full creative partner in his artistic process, Elgammal gives credit to AICAN for the artwork in all of their joint exhibitions.⁴⁰

What all of these artists in the AI field agree upon is that AI is not currently capable of existing as an artist without human intervention or collaboration. While this may change at some point in the future, there is still uncertainty regarding whether viewers would ever accept an AI system as an individual artist outside of a relationship with a human artist. Twentieth century philosopher Michel Foucault posited the concept of an ‘artist function’ as the idea that the artist of a work is not necessarily the person who created it, but rather the construct that people build about the artist in their own mind.⁴¹ In the realm of AI art, this means that if the viewer of AI art conceives of the AI system as a tool, then the viewer will consider the human who created or programmed the machine to be the artist, even if the AI system was the true artist.⁴² With this in mind, the potential for AI to exist as a true artist not only requires technology not currently in existence but also a mental and theoretical shift within the minds of viewers.

Conclusion

Within this paper, I have examined the history and recent rise of AI art within the cultural arts sector. By examining the perspectives of artists and casual viewers, I have established a framework to conceptualize how AI art fits within the arts sector, and I have evaluated the role of AI within the artistic process, from tool to creative collaborator. However, AI art presents a multitude of challenges for museums in particular. With issues pertaining to how to license and copyright these works of art, how to collect them in a sustainable fashion, and how to organize their attribution within databases, the debates surrounding AI art are hardly finished. While there remain a multitude of questions to answer regarding the future of AI and AI art, it is clear that AI art is now part of the cultural arts sector zeitgeist.

Notes