Defining an AI-Generated Artwork: A Transdisciplinary Concept for Cognitive Science, Computer Science, and Art Theory

Defining an AI-Generated Artwork: A Transdisciplinary Concept for Cognitive Science, Computer Science, and Art Theory

Research article ^[1]

1. Introduction

In recent years, the relationship between artistic creation and artificial intelligence (AI) has garnered significant attention from artists (Google Arts & Culture, 2019; Hsu, 2019; Ridler, 2021), philosophers (Moruzzi, 2020b, 2020c, 2021), cognitive scientists (Boden, 2011, 2017), and computer scientists (Hertzmann, 2018, 2020). The proliferation of computer-generated artistic works, produced not by human hands but by algorithms, is unprecedented. Notable examples include the painting Edmond de Belamy (Epstein et al., 2020; A. I. Miller, 2019; Still & dInverno, 2019), the songs from the LP PROTO (Hsu, 2019), the garments designed by ClothGAN (Wu et al., 2021), and even newly generated levels for the video game Doom (Giacomello et al., 2018).

While these artificial creations have captivated the public, they have also sparked a contentious debate regarding the computers capability to create art autonomously. Proponents of machine-generated artistic creativity, such as the artist Mario Klingemann, ^[2] computer scientist Ahmed Elgammal, ^[3] and philosophers Catherina Moruzzi ^[4] and Galit Wellner, ^[5] argue from their respective fields that AI transcends its role as a mere tool and should be regarded as a collaborator alongside human artists.

Conversely, other experts argue the opposite. Artists like Anna Ridler ^[6] and Jon McCormack, ^[7] computer scientist Aaron Hertzmann, ^[8] and philosopher Sean Dorrance Kelly ^[9] contend that AI is more akin to a tool that serves human purposes rather than a creative peer to human artists.

These opposing views stem from the different academic disciplines they represent, each offering distinct interpretations of creativity, art, and AI. In light of these divergent perspectives, this article aims to establish a standardized definition of AI-generated artwork. This definition serves two primary purposes: facilitating cross-disciplinary knowledge exchange and providing a framework to mediate debates within and across various fields.

The proposed definition rests on three foundational pillars: understanding creativity through the lens of cognitive science, defining artworks via art theory, and conceptualizing computational creativity within computer science. This transdisciplinary approach synthesizes insights from these diverse disciplines to offer a holistic examination of their interconnections.

2. Methodology

This study adopts a multidisciplinary approach to explore the interconnected concepts of creativity, artwork, and computational creativity within the context of AI-generated works. Sections 3, 4, and 5 present a conceptual analysis, with the findings synthesized in Section 6. Thus, by integrating Margaret Bodens (2011) seminal definitions of creativity with George Dickies (1969) philosophical concept of artwork, this research illuminates the convergence and divergence of these theoretical frameworks. Additionally, it delves into the complexities of computational creativity and its dynamic interplay with human evaluation.

To establish parameters for defining AI-generated artworks, three essential criteria are proposed: (1) the autonomous production of novel and surprising artifacts by AI; (2) the successful navigation of internal evaluative processes within the AI system; and (3) the potential for human recognition of the output as a viable aesthetic object. These criteria provide a rigorous framework for assessing the extent to which AI-generated artworks align with the proposed definition, ultimately yielding a comprehensive understanding of their nature and significance.

Drawing upon insights from cognitive science, art theory, and computer science, the study develops a transdisciplinary definition of AI-generated art. The investigation begins with a thorough analysis of creativity, followed by a systematic examination of artwork and computational creativity. This multifaceted approach facilitates a comprehensive understanding of the interrelationships and implications of these concepts.

3. The Definition of Creativity

To accurately define AI-generated artworks, it is essential to analyze the process by which these works emerge. Consequently, a conceptualization of creativity that can be effectively applied to such products is required. ^[10] Cognitive scientist Margaret Boden (2011) offers a pertinent framework, defining creativity as the ability to generate ideas or artifacts that are new, surprising, and valuable (p. 29). This definition implies that creativity and production are intrinsically linked, carrying significant implications for understanding the nature of AI-generated works:

(1) X is a creator if X can develop ideas or artifacts that are new, surprising, and valuable.

(2) Y is a producer if Y can develop ideas or artifacts that do not fulfil the criteria under (1).

(3) A is a creation if A is an idea or artifact that is new, surprising, and valuable.

(4) B is a product if B is an idea or artifact that do not fulfil the criteria under (3).

Corollary 1: every creator is also a producer, but the inverse is false.

Corollary 2: every creation is also a product, but the inverse is false.

Bodens definition of creativity is used in this article because, as will be shown, it addresses the challenge of determining where creativity lies. That is, whether creativity is a producers personal property or a product attribute. ^[11] Why is this a problem? Because in studying creativity, there has been a tendency to choose one of these two conceptions over the other. ^[12] Unifying those views will be especially important for this article since, as will be seen, it forces us to understand the functioning of creative algorithms without neglecting the characteristics of their artificial creations.

Next, these two perspectives on creativity will be examined to illustrate how Bodens definition serves as an eclectic position between them. Subsequently, its utility in understanding AI-driven artistic creativity will be demonstrated.

3.1. Creativity as Producers Property

The first viewpoint about creativity considers that it lies in the producer. Consequently, it focuses on analyzing the creators characteristics. In this position, the studies on creativity have been understood as a psychological analysis of great creators personalities. An example of this posture can be found in Sigmund Freuds psychobiography, a psychology subfield designed to investigate Leonardo da Vincis artistic creativity (Kővry, 2011). A more recent illustration of this view can be found in James C. Kaufman and Ronald A. Beghettos Four-C model of creativity (2009). As its name suggests, this setup considers four kinds of creativity: little-c, Big-C, mini-c, and Pro-c. While the majority of people possess little-c, only great masters and thinkers exhibit Big-C. The mini-c is a step for humans to improve their creativity towards Pro-c, the professional level of creativity. It can be noted that the Four-C model neglects the analysis of what is created by focusing mainly on studying the creators mind. As it turns out, setups such as psychobiography or the Four-C are devised considering that the object of study of creativity is the human mind. Unfortunately, this poses a problem for studying computational creativity since AI is a non-human creator and, by extension, lacks a human mind.

3.2. Creativity as a Product Attribute

The second viewpoint about creativity considers that it is a product attribute. It focuses more on analyzing creative works than studying great creators personalities. It can be noted that in this view, the public becomes a judge. In particular, it implies that the audience decides whether something is creative when faced with a specific product. Psychologist Morris I. Stein (1953) shares this view. He defines creative work as novel work that is accepted as tenable or useful or satisfying by a group in time (p. 322). Another example is philosopher Berys Gaut (2010), who, in a more recent article, states that a creative product is one that a group of experts in a particular discipline has approved as such. As it turns out, these stances neglect the study of the characteristics of the creator.

Additionally, the role of the audience in judging creativity poses a challenge when studying computational creativity. For instance, people may not perceive a particular product as creative simply because it was produced by an AI. Furthermore, such a judgment may be influenced by prejudice against computers, particularly in the case of computer-generated art (CG-art) (Arriagada, 2020).

3.3. The Eclecticism of Bodens Definition of Creativity

Bodens definition of creativity unifies the two previous viewpoints because it considers creativity an ability, which is a property of the creator. This aspect of her conceptualization is consistent with the first view, which focuses on the creative agent as the object of study. However, Bodens definition also requires creative products to be new, surprising, and valuable, thus requiring the evaluation of the producers creativity through the examination of the product. This aspect of the definition aligns with the second view, which focuses on the created product as the object of study. Therefore, Bodens definition of creativity has a dual domain, encompassing both the study of the producer and the product.

Additionally, this conceptualization of creativity has two crucial characteristics that distinguish it from other perspectives. Firstly, it does not exclusively focus on the human being, as it does not consider creativity as a property of the human personality. Secondly, it does not disregard the importance of human evaluation, as it recognizes creativity to be subject-dependent. The following discussion will delve into these two features in detail, highlighting their significance in the study of AI creativity.

First, Boden�s definition is not human-exclusive. By saying that creativity is an ability, any producer who exhibits that ability should be considered creative, regardless of the type of producer. For example, if an animal or a computer produces something new, surprising and valuable, then that animal or computer is creative. This conception of creativity is broader than those followed, for example, by psychobiography or the Four-C model. The openness exhibited by Boden�s definition broadens the study of creativity. It no longer focuses exclusively on the human mind and introduces new agents such as creative AI. ^[13]

Second, Boden�s definition is subject-dependent (Moruzzi, 2020c). This means that making a new and surprising product is not enough to demonstrate creativity. The actual product must also be considered valuable. If the latter requirement is not met, the producer is not creative. The value aspect can be understood as people�s external evaluation of a given product. ^[14] For example, according to the preceding point, an AI can be creative. Though, its creative ability is verified by judging its outputs. Specifically, an AI is creative only if its AI-generated products are valuable to the human audience.

In conclusion, this section has established the specific meaning of creativity that this article adopts. It has shown why Bodens conceptualization is preferred over other viewpoints on creativity. In particular, it has been pointed out that it exhibits the features of (1) not being human-centered and (2) being subject-dependent. It is now time to address the second aspect of the definition of an AI-generated artwork, namely, the concept of artwork.

4. The Definition of Artwork

The second step in establishing a transdisciplinary definition of an AI-generated artwork involves elucidating the essential qualities that distinguish a product as an artwork. To this end, it is necessary to formulate a conceptualization of artwork. One influential perspective is the institutional theory of art, as espoused by the philosopher George Dickie (1969). According to Dickie, an artwork is (1) an artifact (2) upon which some society or some sub-group of a society has conferred the status of candidate for appreciation (p. 254). This definition is notable for its status as the most refined version of institutional theories of art, as elucidated by philosopher Stephen Davies (1991), and it is thus highly valuable in achieving the goal of this article. Even in the contemporary era, Dickies definition continues to be invoked to recognize the artistic value of robotic productions (Mikalonytė & Kneer, 2021) and to engage with the status of AI-generated outputs (Moki-Reinikka, 2018).

The question arises: why is Dickies theory helpful in defining an AI-generated artwork? To address this query, it is necessary to understand that art theories can be grouped into functional and procedural categories (Davies, 1991; Dickie, 1997). ^[15] As noted earlier, Dickies institutional theory falls under the latter category. Consequently, it is pertinent to briefly compare and contrast these theoretical orientations, which will enable us to demonstrate the advantages of employing an institutional definition of artwork.

4.1. Functional Theories of Art

Functional theories hold that art has an essential function (Davies, 1991). It can be noted that different functional theories will consider different functions as essential. For example, philosopher Robin George Collingwood posited that art is the expression of emotions, while philosopher Susanne Langer argued that it symbolically represents human feelings (Dickie, 1997). It is important to note that, in these cases, the theorists task is not to agree on what an artwork is. Instead, they seek to discover what function a work must fulfil to be considered art. In other words, a functional theory considers that an artwork intrinsically fulfils the essential function of art. Therefore, a product that does not fulfil this function could not conventionally be declared art by a group of artists, experts, or an audience. For instance, if a group of people deems an AI-generated painting to be art, their mere consensus does not automatically bestow that status upon the AI output. Moreover, no institution, whether a museum, art movement or any other, has the authority to determine whether a particular piece should be considered as art.

It can be seen that discovering the essential function of art is a problem that undermines all functional theories. This particularly truncates the construction of the definition of an AI-generated artwork in our context. Later functional art theorists, such as philosopher Monroe Beardsley, understood that the function of art is to intentionally arrange conditions that offer experiences with a distinct aesthetic character (Davies, 1991; Dickie, 1997). In this case, the introduction of intentionality means that a computer output could hardly be considered art. For instance, consider defining the essential function of art as rooted in intentionality, a capacity inherent to the mind. In that case, an AI capable of producing artworks should have mental states. That implication is strongly debatable. ^{[ 16]} Thus, the problems mentioned above motivate the search for a suitable candidate for defining AI-generated works of art in procedural theories. ^[17]

4.2. Procedural Theories of Art

Dickies artwork definition is framed within procedural theories of art. In these, it is a specific kind of context in which something is embedded that makes it art, not something it functions to do? (Dickie, 1997, p. 20). This is closer to a social convention. Indeed, the emphasis is placed on the context in which a product is embedded instead of the product itself. In this case, no intrinsic function makes some product an artwork.

It has been criticized that a pre-existing art theory is needed to determine this context (Davies, 1991). Thus, if a creator does not have a theory justifying a given work, procedurally, art status cannot be reached. Nevertheless, from the perspective of this study, it could be argued that it is redundant to have a systematized theory for determining whether a product is an artwork, since what is essential is what a given context deems as such. It is important to note that procedural theories inherently view the status of art as an extrinsic characteristic. The context that delivers that status may be a theory, institution, or simple audience agreement. Thus, products can be considered artworks using an implicit theory or none at all. For example, if a group of people considers an AI-generated painting as art, their mere agreement confers that status to that AI output.

Shifting our focus now to Dickies concept of artwork mentioned earlier an artifact that a group of people considers a candidate of appreciation two distinct components can be identified. The first refers to the manufacture of a product, the artifact. It should be noted that this involves a task performed by the producer (P). The second refers to evaluating work, that is, the peoples agreement to consider a given artifact as a candidate of appreciation. Generally, the audience (A) performs this evaluative task. When A considers Ps work worthy of aesthetic appreciation, P becomes an artist, and his/her work becomes an artwork. At this point, there are two extreme options. If A evaluates Ps work negatively, P becomes a lousy artist. As opposed to that, if A evaluates Ps work positively, P becomes a good artist. Between these two extremes, there is a continuum of different values for P. The critical question to bear in mind is that it is not a positive evaluation that gives a product the status of art. It is enough that such a product is considered a candidate for aesthetic appreciation.

Thus, the first part of Dickies definition of an artwork pertains to the artist's creative process, while the second part emphasizes the audiences evaluation of the final product. At this point, it is necessary to recall Margaret Bodens definition of creativity. The review in Section 2 concluded that creativity is a capability that has two components: (1) a producer who generates new and surprising ideas or artifacts, along with (2) a human audience who finds those products valuable. The producer is creative only if (1) and (2) are met. Put in this way, Bodens definition of creativity and Dickies definition of artwork show some similarities. In particular, it is crucial to note that both are subject-dependent. That is, they require external valorization in the terms already outlined. Regarding AI-artistic creativity, this subject dependency exhibits the following two benefits:

Firstly, according to this perspective, art does not require the involvement of agents with mental states in its creation. As mentioned before, functional theories, such as Beardsleys, include intentionality as an intrinsic part of art creation. This requirement restricts artistic creation to human beings and animals that can demonstrate mental states. Non-living agents, such as forces of nature or machines, are left out of that cluster. Differently, using Dickies procedural artwork concept, if a group of people recognizes a given product as art, it will count as such. For example, suppose a community agrees to consider the sound of a flowing river as art. In that case, it will count as art, even though the forces of nature generated that river without any intentionality. Robots productions have also seen their status as art defended using Dickies definition (Mikalonytė & Kneer, 2021). Indeed, the robots products will count as art if people deem it so, regardless the robots have no mind.

Secondly, Dickies definition has proved useful in analyzing why two physically indistinguishable objects can have different art status (Mki-Reinikka, 2018). Let us further explore this idea. Historically, procedural theories emerged after philosopher Arthur Dantos publication of The Artworld? (1964). In particular, Dickie asserts that Dantos article introduces a new way of thinking within the philosophy of art, due to the inclusion of an argument that Dickie refers to as The Perceptually Indistinguishable Objects. ^[18] Dickie (1997) summarizes this argument as follows:

Consider a pair of visually indistinguishable objects . . . as Fountain and a urinal that exactly resembles it . . . In such a pair, one object is a work of art and the other is not, or would not be, an artwork, and, since they exactly resemble one another, it cannot be some visually discernible characteristic that makes the artwork art. Therefore, it must be some context involving at least some non-visible elements that makes the artwork art. (p. 20)

Accordingly, artworks are created through the combined efforts of artists and the community, with the latter providing the framework for interpreting a given work as art. The audience, the non-visible element referenced by Danto, plays a pivotal role in this process. In the context of AI-generated artworks, a procedural theory must address and substantiate why creations by machines can be classified as art. To illustrate this point, consider three representations of the Gioconda: the original masterpiece by Leonardo da Vinci, a copy created by a distinguished human artist, and an AI-generated 3D print. The argument here is that the identity of the creator of these works is not the determining factor in their classification as art; rather, it is the audiences perception that is crucial. The acceptance of these works as art is contingent upon the audience's evaluation. Different audiences may render different judgments about these paintings. It is important to note that there is no intrinsic property in the molecular composition or temporality of these works that inherently qualifies them as art; their designation as such depends entirely on the audiences recognition.

In summary, this section has elucidated the definition of artwork that this article adopts and explicated why Dickies procedural conceptualization is preferred over functional art theories. Specifically, it has been emphasized that Dickies definition possesses two critical attributes: firstly, it is not human-centered as it does not mandate the artist to possess a human mind; and secondly, it is subject-dependent as it necessitates external human approbation for a product to be deemed as art. The subsequent section will delve into the third aspect of the definition of an AI-generated artwork, which is AI creativity.

5. AI Creativity

The third step to constructing a transdisciplinary definition of an AI-generated artwork is to understand AI creativity. Therefore, a conceptualization of AI is needed. Recently, the researcher Mariusz Stanowski (2021) has pointed out that AI can be described as a way of implementing by technical methods such solutions or behaviors that, if they were human actions, we would call them intelligent (p. 139). From the standpoint of this study, this definition allows us to highlight the following three valuable points in order to achieve the goal of this article:

Firstly, Stanowskis description of AI can be extended to a broader concept of intelligence. Specifically, intelligence, whether human or machine, is defined as the ability to use ones resources to solve problems, provided that a human audience recognizes it as such. The focus will now shift to the first aspect of this definition, with the second aspect addressed in the following paragraphs. Therefore, intelligence can be understood as the capacity to solve problems. For instance, if a person possesses this ability, it demonstrates natural intelligence; similarly, if a pig exhibits this ability, it indicates another form of natural intelligence. By the same logic, if a machine has this ability, it can be considered to possess artificial intelligence. In line with Bodens definition of creativity, Stanowski suggests that intelligence is not an exclusively human trait; rather, anything or anyone capable of problem-solving can be regarded as intelligent.

Furthermore, Stanowski underscores that an intelligent agent, whether a machine or a human, must utilize its own resources to solve problems. This criterion implies a need for autonomy, rather than mere assistance. Consequently, a human relying on another for problem-solving would not be deemed intelligent if they cannot rely solely on their own resources. Although human autonomy is generally assumed, this aspect is particularly crucial for machines, which must demonstrate their autonomy to be considered intelligent. This issue will be further explored in relation to creative autonomy within computer science.

Secondly, with regard to human audience recognition, Stanowskis definition of intelligence suggests that problem-solving ability alone is insufficient for an entity to be considered intelligent. Public recognition is also a requisite. In other words, according to Stanowski, intelligence is subject-dependent. For instance, a human who is a misunderstood genius may not be recognized as intelligent despite possessing problem-solving capabilities. Similarly, a machine might excel in solving various problems, but if its outputs are not perceived as valuable by a human audience, the machine cannot be classified as intelligent.

Thirdly, within computer science, artificial creativity is a specialized subfield. Simon Colton and Geraint Wiggins (2012) define artificial creativity as the philosophy, science, and engineering of computational systems which, by assuming specific responsibilities, exhibit behaviors that unbiased observers would consider creative (p. 21). This perspective highlights that artificial creativity, like human creativity, is subject-dependent. This requirement parallels (a) Bodens definition of creativity and (b) Dickies definition of artwork. Consequently, the following sections will first elaborate on the subject-dependent nature of AI, identifying parallels with the concepts discussed earlier concerning Bodens creativity and Dickies artwork. Subsequently, the discussion will address the necessity of creative autonomy for AI to be recognized as a creative agent. This approach will facilitate the construction of a transdisciplinary definition of AI-generated artwork by integrating insights from cognitive science, art theory, and computer science.

5.1. The Subject-Dependent Character of AI

The subject-dependent character of AI means that an algorithm can be a skilled problem solver, but peoples opinions determine the final verdict on its intelligence. Intelligence is an ability that requires external evaluation. This requirement can be traced back to its very beginnings in computer science.

In Computing Machinery and Intelligence (1950), Alan Turing poses the question: can machines think? To avoid problems from using the word think, he replaces the question with his famous imitation game, with its questions and answers. Turings reasoning shows that the crucial point is not if the machine is thinking. Instead, what is important is if humans can be fooled and believe that. It should be noted that the Turing Test takes advantage of the subject-dependent characteristic of AI to test its performance. Following this articles aim, the next step is to ask if a Turing Test can be adapted to check the artistic creativity of AI.

5.1.1. Empirical Evidence on Human Audience Evaluation of AI-generated products

Recently, experiments have been conducted to present AI-generated outputs to human audiences. The conclusions of three of the most prominent studies are as follows:

In An investigation into peoples bias against computational creativity in music composition (2006), researchers David C. Moffat and Martin Kelly used a modified version of the Turing Test to evaluate the performance of machines in music composition. In particular, computer-generated songs from the Voyager (Lewis, 1993), ILLIAC ^[19] Suite (Hiller, 1989) and EMI ^[20] (Cope, 2003) machines were used. These were added to human pieces in similar styles (i.e., jazz, violin and Bach-like). The human listeners were grouped into musicians and non-musicians. The results showed that most of the computer-created songs passed the Turing Test. In particular, the musicians group was more likely to be fooled. Nevertheless, Moffat and Kelly also note that both musicians and non-musicians are biased against computers. Consequently, listeners tend to evaluate work they believe to be artificially-made worse than work they believe to be human-made.

In CAN: Creative Adversarial Networks, Generating Art by Learning About Styles and Deviating from Style Norms (2017), researchers Ahmed Elgammal, Bingchen Liu, Mohamed Elhoseiny, and Marian Mazzone also used an adaptation of the Turing Test to evaluate the performance of machines in plastic creation. In particular, they used paintings generated by Generative Adversarial Networks (GANs), ^[21] human-generated paintings obtained from Art Base^{l [22]} 2016 and a sample of abstract expressionism. The results showed that human observers could not distinguish which art is computer-made or human-made.

More recently, researchers Martin Ragot and Nicolas Martin again subjected GANs to a modified Turing Test in AI-generated vs Human Artworks. A Perception Bias Towards Artificial Intelligence? (2020). In this case, the human audience rated paintings labelled AI-made and human-made. Nevertheless, the labelling was manipulated. Thus, some AI-made paintings were labelled as human-made and vice versa. Next, the manipulation was reported to the observers. Then, they were asked to try to find out the origin of some of the paintings. The results showed that, depending on the perceived identity of the author (Human vs AI), the same artworks were evaluated differently. Moreover, the results show that real artworks made by humans are also evaluated more highly than real AI-generated artworks? (Ragot et al., 2020, p. 6).

In short, it can be seen that these machine-generated products pass the Turing Test. While there may be differences in the quality of their creations (i.e., whether they are considered good or bad art), this does not affect their status as art. On the other hand, it is important to consider the potential influence of anti-machine biases on negative human evaluations.

5.2. The Autonomy Requirement of AI in Computational Creativity

One common critique of computational creativity focuses on the issue of autonomy. While it is acknowledged that machines do produce outputs, there is debate over whether they do so autonomously, like humans. From my research perspective, this criticism overlooks the following facts:

First, a human performer cannot definitively demonstrate autonomy. Expanding upon Turings response to Jefferson, it becomes apparent that the question of autonomy is inherently unverifiable. When asserting that a machine lacks autonomy, it is equally valid to argue that human autonomy is similarly unknown. Both humans and AI are influenced by external stimuli that affect their creative outputs. This reasoning extends to artistic creativity, which, as a subset of computational creativity, follows the same logic.

Second, the very definition of CG-art refers to outcomes produced with zero or minimal human interference (Boden, 2011). This can be seen as an appeal to machine autonomy. Computer science has responded to this query. Indeed, Kyle E. Jennings has said that creative autonomy exists when a system not only evaluates creations on its own but also changes its standards without explicit direction? (Jennings, 2010, p. 489).

It can be noted that this requirement is covered on the very basis of machine learning (ML). Creative autonomy is, at its root, another way of saying that machines must learn through trial and error. Nevertheless, this does not mean that only ML-based machines are autonomous. In particular, two things are required: (1) self-evaluation and (2) changes founded on that evaluation.

According to the foregoing, autonomy is currently assured in certain types of AIs, namely GANs. Research in artistic computational creation focuses on these models (Wu et al., 2021). It is essential to know that they are a type of ANN. As summarized by researcher Arthur I. Miller (2019), GANs consist of two deep neural networks, the discriminator (D) and the generator (G). For example, if D is fed a dataset of tulip images, G will try to create new tulip images without access to Does dataset. That is, starting from scratch and only through trial and error, G will try to fool D into believing that this new image is a tulip. Since G starts with no information, its first attempts at deception will send out noise, a jumble of meaningless dots D rejects. Still, G picks up that what it created was not a tulip with each rejection. Eventually, after learning from its mistakes, G deceives D. This iteration process is unsupervised ML.

It must thus be noted that G evaluates its work and learns from its mistakes. On the other hand, the GAN tandem composed of G and D is autonomously critical of its work, as it incorporates an internal quality control of its creations. In this sense, Galit Wellner (2021) has mentioned that these algorithms operation is a variation of the Turing Test.

In practice, GANs have demonstrated their ability to create art that is indistinguishable from human art (Cetinic & She, 2021). Consequently, both approaches discussed address criticisms regarding computational autonomy. Nevertheless, as Wellner (2021) observes, the final appreciation and judgment of the outcome are performed by humans, even though GAN algorithms are involved and make initial judgments. Ultimately, works of art are intended to be viewed (or listened to) by humans.This observation brings the discussion back to the Turing Test. It is insufficient for an algorithm to exhibit creative autonomy as described by Jennings (2010); the human audience must also perceive the algorithm as creative.

6. AI-Generated Artwork: A Transdisciplinary Definition

The present study sought to establish a comprehensive, transdisciplinary definition of AI-generated artwork. Prior analyses revealed a shared, subject-dependent character among Margaret Bodens conception of creativity, George Dickies definition of artwork, and the computational creativity paradigm within computer science. Building on this foundation, it is proposed that an AI-generated artwork involves (1) an autonomous AI-production of a new and surprising idea or artifact, (2) which passes an internal evaluation mechanism embedded in the very same AI, and (3) is considered a candidate for appreciation by a human audience.

The adoption of this definition clarifies conceptual ambiguities prevalent in interdisciplinary discourse. By evaluating adherence to these criteria, more productive discussions can be facilitated. The absence of any criterion precludes classification as AI-generated artwork. For instance, fulfillment of only criteria (1) and (2) results in an AI-generated product, not art. Similarly, satisfying solely criterion (3) based on general artistic understanding does not constitute AI-generated artwork.