Biases in Generative Art— a Causal Look from the Lens of Art History

Biases in Generative Art— A Causal Look from the Lens of Art History

Ramya Srinivasan Kanji Uchino Fujitsu Laboratories of America Fujitsu Laboratories of America

ABSTRACT With the rapid advancement of deep learning, there has been With rapid progress in artificial intelligence (AI), popularity of remarkable progress in AI based generative art. From creating hy- generative art has grown substantially. From creating paintings to brid images using attributes from multiple images to generating generating novel art styles, AI based generative art has showcased a cartoons from portraits, AI based generative art has exemplified variety of applications. However, there has been little focus concern- new and diverse applications [3, 25, 26, 45]. ing the ethical impacts of AI based generative art. In this work, we As a consequence, AI based generative art has become extremely investigate biases in the generative art AI pipeline right from those popular. In 2019, “Sotheby" auction house sold an AI generated art that can originate due to improper problem formulation to those work for 32000 pounds [55]. A little earlier in 2018, the auction related to algorithm design. Viewing from the lens of art history, we house “Christies" sold an AI generated art work for a staggering discuss the socio-cultural impacts of these biases. Leveraging causal 432500 USD [15]. Institutions have also shown an increased interest models, we highlight how current methods fall short in modeling towards AI based generative art as evidenced by the number of the process of art creation and thus contribute to various types of museum shows [7, 24]. There are also several apps and tools such as biases. We illustrate the same through case studies, in particular [3, 20, 37] that have not only enhanced the popularity of AI based those related to style transfer. To the best of our knowledge, this is generative art, but have also facilitated ease of use and accessibility the first extensive analysis that investigates biases in the generative to end-users. art AI pipeline from the perspective of art history. We hope our work sparks interdisciplinary discussions related to accountability of generative art. CCS CONCEPTS • Social and professional topics; • Computing methodologies → Artificial intelligence; Figure 1: Example of bias in AIportraits app [2]—Skin color of ac- KEYWORDS tress Tessa Thompson (left) is lightened in the app’s portrait rendi- generative art, style transfer, biases, AI, socio-cultural impacts tion (right), thus exhibiting racial bias. Image source [53]. ACM Reference Format: Ramya Srinivasan and Kanji Uchino. 2021. Biases in Generative Art— A Amidst this progress and popularity surge, experts across dis- Causal Look from the Lens of Art History. In . ACM, New York, NY, USA, ciplines have voiced concerns regarding the consequences of gen- 11 pages. erative art. For example, in [19], the authors reflect on the impact of interfacing with art autonomously generated by non-human 1 INTRODUCTION creative agents. The authors argue that such art could widen the Generative art refers to art that in part or in whole has been created divide between the human creator and the human audience, and by the use of an autonomous system [11]. In a broad sense, the emphasize on the need to balance between automation and develop- term “autonomous" can refer to any non-human system that can ment of humans’ creative potentials. Discussing whether computers determine features of an art work, such as use of smart materials, can replace human artists, the author in [34] argues that “art re- arXiv:2010.13266v2 [cs.AI] 16 Feb 2021 mechanical processes, and chemical processes, to name a few. Com- quires human intent, inspiration, a desire to express something". The puter generated art, i.e. art generated by algorithms or computer author further adds that artistic creation is primarily a social act, programs, is perhaps the most common form of generative art [63]. and concludes that computers cannot replace humans. In a simi- In fact, the terms "generative art" and "computer art" have been lar vein, discussing whether machines can create art, the author used more or less interchangeably for a long time now [11, 63]. in [16] argues that the process of creating an art work is distinct from the outcome, i.e. the artwork. Drawing from the theory of Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed philosophical aesthetics, the author states that when something is for profit or commercial advantage and that copies bear this notice and the full citation created, something about inner self is expressed. on the first page. Copyrights for components of this work owned by others than ACM It has been argued that both art and technology are means must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a through which humans reveal their epistemic knowledge [42]. Such fee. Request permissions from [email protected]. knowledge could include societal values, cultures, beliefs, as well FAccT, March 3–10, 2021, Virtual Event, Canada, , as individual biases and prejudices. In the work [34], artist and © 2021 Association for Computing Machinery. ACM ISBN ACM ISBN 978-1-4503-8309-7/21/03. computer scientist Aaron Hertzman states “artworks are created by a human-defined procedure", and further notes that computer FAccT, March 3–10, 2021, Virtual Event, Canada, , Ramya Srinivasan and Kanji Uchino

bias, gender bias, and other types of discrimination. Second, based on their limited understanding, algorithms could stereotype artists’ style and not reflect their true cognitive abilities. This means aspects such as artist’s intent and emotions are overlooked, thus potentially conveying an opposite affect in the generated art. Third, historical events and people may be depicted in a manner contrary to the original times, thus contributing to a bias in understanding history Figure 2: Example of bias in learning artists’ styles. The affect con- and thereby get in the way of authentically preserving cultural veyed by the original image (a) is lost in the “Van Gogh" version (b) heritage (illustrated in Sec. 6). These observations therefore compel of [70]. This is contrary to Van Gogh’s style, see (c), a real artwork an analysis of generative art so as to uncover various types of biases. by Van Gogh depicting red flowers in the field like in(a). 1.1 Contributions generated art can thus be biased. Even artists like David Young who In this paper, we investigate biases in AI based generative art right argue that machines create art on their own, acknowledge existen- from those that can originate due to inappropriate problem formu- tial bias in generative art. In the essay Tabula Rasa [67], Young notes lation to those that can be related to algorithm design. Viewing that human biases in the form of preconceptions, irrationalities, from the lens of art history, we discuss the social-cultural impacts and emotions can easily get embedded into the data used to train of these biases. We advocate for the use of causal models [48] to these generative art AI models. A recent notable example of bias depict potential processes of art creation. We highlight how current in generative art concerns a portrait generator app called “AIpor- methods can fall short in modeling the process and thus contribute traits" [2]. It was pointed out that skin color of people of color was to various biases such as selection bias and transportability bias lightened in the app’s portrait rendition [38, 53]. Figure 1 provides [10]. We illustrate the same through case studies that span various an illustration of the same. Furthermore, AI algorithms are best art movements, artists, art media/material, genres, and geographies. thought of as data-fitting procedures [12]. As the author in [34] First, examples of biases that arise due to improper problem beautifully describes, these algorithms are “like tourists in a foreign formulation, namely, framing effect biases, are considered (Sec. 4). country that can repeat and combine phrases from the phrasebook, After providing a background of causal models (Sec. 5), we discuss but not truly understand the foreign language or culture". case studies to demonstrate confounding biases in modeling artists’ Many types of latent biases in generative art, especially those styles (Sec. 6.1), followed by illustrations of selection bias (Sec. 6.2). concerning art history, have not been analyzed in any of the past Biases in transferring artists styles, also known as transportability studies. Further, socio-cultural impacts of biases in generative art biases, are discussed (Sec. 6.3). Transportability bias is demonstrated have not been investigated. We aim to address these issues in this by considering case studies that include both art movements that work. We motivate the problem with an illustration. Consider an are subtly different (e.g. modern art movements such as Cubism image-to-image translation model such as CycleGAN [70] which and Futurism) and art movements across different time periods (e.g. has been used to create images across different artists’ “styles". Fig- Modern art and Early Renaissance). Illustrations of biases in datasets ure 2 (b) shows an example of “Van Gogh" version of a photograph are provided in Sec. 7. A list of the case studies examined in the (Figure 2(a)) as rendered by the CycleGAN model. As can be seen, paper can be found in Table 1. Our findings shed light on the various the affect conveyed by the original image is lost in the “Van Gogh" inherent biases in generative art right from problem formulation version: the red flowers, perhaps indicative of Spring are no longer and datasets to algorithm design and data analysis. We also discuss evident, instead a dry season is reflected. the socio-cultural repercussions of these biases (Sec. 8). To the best The generated image seems quite contrary to Van Gogh’s take on of our knowledge, this is the first extensive analysis that investigates colors. “Green Ears of Wheat", an 1888 art work by Van Gogh serves biases in the generative art AI pipeline from the perspective of art as an illustration to the point; see Figure 2 (c) [60]. As documented history. We hope our work triggers interdisciplinary discussions in his letters to his sister Wilhelmina and artist Horace M. Livens, concerning accountability of generative art, and sparks the design Van Gogh mentions about how he emphasized colors [58], and how of novel methods to address these issues. by using bright contrasting colors he was able to infuse life in his works: “ Poppies or red geraniums in vigorously green leaves - motif in red and green. These are fundamentals, which one may subdivide 1.2 Case study selection further, and elaborate, but quite enough to show you without the We surveyed academic papers, online platforms, and apps that gen- help of a picture that there are colours which cause each other to erate art using AI. In order to uncover potential biases from an art shine brilliantly, which form a couple, which complete each other historical perspective, from the surveyed list, we selected papers like man and woman”,[56, 57]. Thus, by merely using correlation and platforms that focused on simulating established art movements statistics to model the artist’s style, aspects such as emotion and and/or artists’ styles. Thus, papers such as [25] or platforms such intent which are central to an art’s creation [16] are not taken into as [3] which focus on deviating from established styles to create account, thereby rendering a biased representation of the artist’s imaginary patterns are not included in our study. To demonstrate “style" and also possibly stereotyping the artist in the process. various biases, we have considered state-of-the-art generative art Such biases could potentially have long standing adverse socio- AI models [54, 70] and platforms/apps such as [2, 20, 29] that focus cultural impacts. First, because of the inherent biases in training on simulating established art movements and artists’ styles. The art data and algorithms, generative art could be embedded with racial movements considered as part of case studies have been determined Biases in Generative Art— A Causal Look from the Lens of Art History FAccT, March 3–10, 2021, Virtual Event, Canada, ,

CS Model Bias type Art Movements Artists Genres 1 [70] Confounding bias Post-Impressionism Vincent Van Gogh Landscape 2 [54] Selection bias Romanticism Gustave Dore Illustration 3 [2] Selection bias (racial bias) Renaissance Various artists Portraits 4 [70] Transportability bias Post-Impressionism Paul Cezanne Photo, Landscape 5 [20] Transportability bias Cubism, Futurism Fernand Leger, Gino Severini Genre art, battle painting 6 [20] Transportability bias Realism, Expressionism Mary Cassatt, Ernst Kirchner Portraits 7 [29] Transportability bias Renaissance, Expressionism, Clementine Hunter, Portrait, Sculpture (racial bias) Folkart Desiderio da Settignano 8 [1] Transportability (gender) bias Renaissance Raphael, Piero di Cosimo Portraits 9 [2] Representational bias Renaissance Various artists Portraits 10 [54] Label bias Ukiyo-e Various artists Various genres Table 1: Summary of Case Studies (CS) described in the paper. Note, there could be more than one type of bias associated with each CS. For illustrative purposes, only one bias type is discussed in each CS. Model denotes the algorithm/platform listed under corresponding references.

based on the experimental set-ups reported in these state-of-the-art deviation from art distribution. In [70], the authors proposed a AI models and platforms. These include Renaissance art, Modern art method to translate styles across unpaired images and illustrated (Cubism, Futurism, Impressionism, Expressionism, Post Impression- its applicability in transferring artists’ styles. In [54], the authors ism and Romanticism), and Ukiyo-e art. It is to be noted that due to used conditional GANs to generate artworks. the paucity of existing AI applications that study non-western art There are many tools and apps to facilitate users to easily create forms, Ukiyo-e is the only non-western art form studied. The genres art. Using [2], users can transform a portrait in the style of famous span landscapes, portraits, battle paintings, genre art, sketches, and portraits. Photos can be converted into artworks using [37]. In an illustrations. Art material considered includes woodblock prints, application called GANbreeder, two images are combined to gener- engravings, paint, etc. The study includes artists across cultures ate a new image [3]. The style of the input image is transformed such as Black folk artist Clementine Hunter, American painter Mary into another specified style in [20]. Cartoonify turns a photo into a Cassatt, Dutch artist Van Gogh, French illustrator and sculptor Gus- cartoon drawing leveraging Google’s “Draw This" [43]. tave Dore, Italian artist Gino Severini, amongst others. In the next Given that there are many tools to quickly create art, there is section, we discuss work related to computer generated art. an increased risk for generative art to be biased. With surging art market [13] and pressing need for diversity and inclusiveness in art [6], an analysis of bias in generative art becomes even more 2 AI FOR ART GENERATION pertinent. In a recent art project titled ‘Imagenet Roulette’ [17], AI Computer generated art has a long history. In 1970’s, painter Harold researcher Kate Crawford and artist Trevor Paglen exposed biases Cohen began exhibiting paintings generated by a program called in machine learning datasets. While there have been studies to AARON [34]. By 1980s, several artists were using computer pro- understand how humans perceive art [36] and to examine if there grams to create interactive experiences for the audience [34]. In is a perception bias towards art created by AI [51], there is little to the 1990s, Flash, a tool for creating animations became popular [7]. no extensive analysis concerning biases in generative art. In this Around the same time, Paul Haeberli introduced a paint program paper, we provide an extensive analysis of biases in art generated whereby a user could quickly create a painting without needing autonomously by AI, looking from the perspective of art history. any technical skill [33]. In 2000s, tools like Processing [14] and Such an analysis is useful to understand machine related biases OpenFrameworks [69] allowed artists to make art using code. independent of human biases, and is necessary to study scenarios As early as 2001, researchers in computer vision were training involving both human (artist) and machine bias. In the rest of this computers to learn artists’ styles from examples [35]. Since 2012, paper, generative art refers to AI based generative art. rapid advancement in deep learning has triggered a wide range of models for AI based generative art. For example, [44? ] is an open source tool released by Google that uses a convolutional neural network (CNN) to understand what neural networks are doing 3 ART-HISTORICAL ASPECTS OF at each layer by creating dream-like appearances. Another CNN ARTWORKS architecture is the neural style transfer [26] work that allows to In this section, we discuss various aspects pertaining to art history blend content of one image into style of another image to create based on which art works can be analyzed. Art historians employ a new images. In [32], a recurrent neural network is proposed to number of ways to group world arts into systems of classification construct stroke-based drawings of common objects. [60]. These groupings are based on a set of qualities that are signifi- Recently, generative adversarial networks (GANs) [30] have be- cant. Such significant qualities could be related to specific approach come popular for creating art. Creative adversarial networks [25] of an artist, material used to create art works, art movements, genre, proposed modifications to the GAN objective to make it creative etc. We provide a brief account of some main aspects so as to aid in by maximizing deviation from established styles and minimizing understanding some types of biases that we illustrate in the paper. FAccT, March 3–10, 2021, Virtual Event, Canada, , Ramya Srinivasan and Kanji Uchino

3.1 Art Movements art [60]. An artist usually can work across different genre types, Art movements can be described as tendencies or styles in art however, art historians mark certain artists as representatives of a with a specific common philosophy influenced by various factors particular genre. For example, Anthony van Dyck is recognized as a such as cultures, geographies, political-dynastical markers, etc. and portraitist, Alfred Sisley as a landscape painter, and Piet Mondrian followed by a group of artists during a specific period of time as an abstract artist – though each one of them worked in a number [60]. Examples of art movements include Ancient Egyptian art, of different genres [60]. Wikiart dataset provides an extensive list Ancient Greek art, Medieval Art, Renaissance art, Modern art, etc. of genres based on artists and artworks. Within each of these art movements, there are sub-categories based on various factors. For example, modern art includes many sub- 3.4 Artists categories such as Symbolism, Impressionism, Post-impressionism, There are many aspects that characterize an artist’s “style". In addi- Cubism, Futurism, Pop-art, and so on. As an illustration, consider tion to factors such as art movement, art material, and genre, an Impressionism and Post-impressionism. Both movements origi- artist’s style can be characterized by factors such as their cultural nated in France, however Post-impressionism originated in reac- backgrounds, their art lineage or schools (from whom they learned tion to Impressionism. While Impressionism was characterized by or who influenced them), and other subjective aspects such as their vibrant colors, spontaneous brush strokes, and urban life styles, cognitive skills, beliefs, prejudices, and so on. Consider for example, Post-Impressionism artists had their own individual styles to sym- Paul Cezanne, one of the most popular artists in the history of bolically display real subjects and their emotions [47]. Similarly, modern art. Although generally categorized as a Post-Impressionist each art movement is characterized by unique features that reflects artist, Cezanne influenced several other art movements such as certain trends. Thus art movement is a dominant factor influencing Cubism and Fauvism. Some of his early pictures depict classical and artists and artworks. Interested readers may refer to [60] where over romantic themes with expressive brushwork and dark colors, while hundred sub-categories are listed across a dozen art movements. later he is said to have adopted brighter colors drawing inspiration, emotions, and memory to paint [59]. Cezanne himself remarked: “A 3.2 Art material work of art which did not begin in emotion is not art". In his still-life Artworks can also be grouped based on the material and techniques paintings, Cezanne began to address technical problems of form and used in creating the art. Charcoal, enamel, mosaics, tapestry, paint, color by experimenting with subtly gradated tonal variations, or and lithography are some examples of art materials. Artists use “constructive brushstrokes,” to create dimension in the objects [59]. different techniques to create artworks from different materials. For His artworks span a variety of genres and exhibit patterns from example, mosaic is a coherent pattern or image in which each com- multiple art movements marked by subtle cognitive aspects. Thus, ponent element is built up from small regular or irregular pieces modeling artists’ style is not a straightforward computational task, of substances such as stone, glass or ceramic, held in place by plas- it entails many abstract elements that are hard to quantify. Yet, re- ter/mortar, entirely or predominantly covering a plane or curved searchers define “style" in ways that suit their model’s performance, surface, even a three dimensional shape, and normally integrated we discuss this issue next. with its architectural context [22]. Mosaics were traditionally used as decoration for floors and walls becoming very popular across the 4 BIASES DUE TO PROBLEM FORMULATION Ancient Roman World. Different art movements saw the prevalence Biases can arise based on how a problem is defined and is formally of different materials. For example during the Renaissance period, known as the framing effect bias [50]. Consider, for example, the sculptures were made out of various materials like marble, white problem of style transfer in artworks. There are at least two different stone, gold, etc. Thus, material and technique employed to create notions of styles when it comes to artworks: one which is related art can influence the artist and the resulting artwork. An elaborate to the art movement, and another which is related to the artist. list of various materials can be found in the WikiArt dataset [60]. As described in Section 3, there are many aspects to each of the above notions of style. Yet researchers conveniently define style in 3.3 Genre a manner that suits their model’s performance. This is a consistent Genre of an artwork is based on the depicted themes and objects. A problem across several models. For example, in [54], the authors hierarchy of genres was developed in the 17푡ℎ century [23]. Accord- claim that their model learns Ukiyo-e style since the generated ing to this hierarchy, history paintings, namely paintings depicting images are “yellowish" like Ukiyo-e artworks. In [70], a single model scenes of important historical, mythological, and religious events, is used to learn styles of artists and art movements. Like in [54], were considered to be the top ranked genre and this was so until the the justification to have learned Ukiyo-e style seems to be based on mid 19푡ℎ century. History paintings were usually large and typically color features. Thus, “style" has been defined based on the color of narrated a story such as a battle, allegory, or the like. Portraiture the generated art. Given that most Ukiyo-e works were woodblock was another prominent genre and these usually depicted royals, prints, it is thus natural for the generated art to be yellowish. aristocrats, and other important people in society. Portraiture had Ukiyo-e is a form of Japanese art. The works usually depicted to convey aesthetic aspects of the subject depicted, such as their landscapes, tales from history, scenes from the Kabuki theatre, and power, beauty, etc. In contrast, “genre painting" depicted scenes other aspects of everyday city life. Some unique characteristics from every day lives of ordinary people. Landscapes, animal paint- of Ukiyo-e included exaggerated foreshortening, asymmetry of ing, and still life painting are some other prominent genres. Abstract design, areas of flat (unshaded) colour, and imaginative cropping of or figurative art are the most common genres for contemporary figures [64]. Foreshortening refers to the technique of depicting an Biases in Generative Art— A Causal Look from the Lens of Art History FAccT, March 3–10, 2021, Virtual Event, Canada, , object or human body in a picture so as to produce an illusion of projection or extension in space and to convey the notion of depth. These characteristics are not captured in the examples depicted in [54, 70]. Such drawbacks can also happen due to the model design issues which we discuss in Section 6. Nevertheless, inappropriate problem formulation can introduce and perpetuate biases across the generative art AI pipeline. In the next section, we briefly describe Figure 3: D-separation: Graph structures to illustrate conditional structural causal models that we leverage to analyze different biases. independence. Please refer to Section 5.1 for details

5 STRUCTURAL CAUSAL MODELS the path from 푋 to 푍. In the center graph (ii), 푌 is a common cause We advocate for the use of causal directed acyclic graphs (DAGs) of 푋 and 푍. If unobserved, 푌 is a confounder as it causes spurious [48] to analyze some types of biases that are related to algorithm de- correlations between 푋 and 푍. Conditioned on 푌, the path from sign and datasets. In Section 3, we saw that there are several aspects 푋 to 푌 is blocked. In the rightmost graph (iii), 푌 is a collider as relevant to an artwork and that these aspects could influence each two arrows enter into it. As such, the path from 푋 to 푌 is blocked. other in many ways. For example, art movement could influence the However, conditioning on 푌 , the path will be unblocked. In general, choice of art material, the subject of a portrait could influence the a set 푌 is admissible (or “sufficient”) for estimating the causal effect artist, and so on. DAGs help in visualizing such relationships. DAGs of 푋 on 푍 if the following two conditions hold [48]: also allow encoding of assumptions about data, model, and analysis, and serve as a tool to test for various biases under such assumptions. • No element of 푌 is a descendant of 푋 Researchers have leveraged causal models to discuss and develop • The elements of 푌 block all backdoor paths from 푋 to 푍—i.e., various notions of fairness [28, 40]. Causal models facilitate domain all paths that end with an arrow pointing to 푋. experts such as art historians to encode their assumptions, and hence serve as accessible data visualization and analysis tools. Based 5.2 Interventions on different expert opinions, there can be multiple assumptions. An 푖푛푡푒푟푣푒푛푡푖표푛 on a graph is denoted by the “푑표" operator [48]. Thus, there can be more than one DAG describing the relationship The 푑표 operator corresponds to setting the intervened variable to a of an artwork with the artist, genre, art movement, etc. Thus, using specific value and removing the influence of other variables onit. multiple DAGs it is possible to analyze for biases under different For example, in graph (ii) of Figure 3, 푑표(푋 = 푥) implies that the scenarios. We discuss basic concepts related to causal models, and variable 푋 is set to the value 푥 and the incoming arrow into 푋 is through case studies, illustrate the intuition behind using causal removed. 푑표 operator facilitates quantification of causal effects. For models to analyze biases in generative art. example, in Figure 3 (ii), the expression 푃 (푍 |푑표(푋 = 푥)) quantifies DAGs are visual graphs for encoding causal assumptions be- the causal effect of 푋 on 푍. A set of rules referred to as 푑표 −푐푎푙푐푢푙푢푠 tween the variables of interest. Specifically, the assumptions about are always applicable in the context of interventions [48]. These data are encoded by means of structural causal models (SCMs) [48]. rules determine when it is possible to A structural causal model 푀, consists of two sets of variables, 푈 and • to add/delete observations in interventions, 푉 , and a set 퐹 of functions that determine or simulate how values • to exchange interventions and observations, are assigned to each variable 푉푖 ∈ 푉 . The variables 푉 are observed • to add/delete interventions and variables 푈 are unobserved. Variables 푈 and 푉 constitute the 푑표 −푐푎푙푐푢푙푢푠 helps in rendering expressions free from “푑표”. We are vertices of a causal graph 퐺 and the directed edges between them now equipped to discuss algorithmic biases in generative art. denote the various causal dependencies. 6 BIASES RELATED TO ALGORITHM 5.1 d-separation Bias can arise if the algorithm ignores the effect of unobserved Regardless of the functional form of the equations in the model (퐹), variables, overlooks domain specific differences, uses subsets ofthe conditional independence relations can be obtained if the model population for analysis, and so on. We discuss these in this section. 푀 satisfies certain criteria. 푑 − 푠푒푝푎푟푎푡푖표푛 is a criterion for decid- ing, from a given a causal graph, whether a set 푋 of variables is 6.1 Confounding bias independent of another set 푍, given a third set 푌. The idea is to Confounding bias originates from common causes that affect both associate “dependence" with “connectedness" (i.e., the existence of inputs and outputs [48]. We illustrate this bias through case studies. a connecting path) and “independence" with “unconnected-ness" or “separation" [48]. Path here refers to any consecutive sequence 6.1.1 Case Study 1: Modeling artists styles has been one of the of edges, disregarding their direction. most common applications in generative art. As discussed in Section Consider a three vertex graph consisting of vertices 푋, 푌 , and 푍. 3.4, several observed and unobserved abstract aspects constitute an There are three basic types of relations using which any pattern of artist’s style. Revisiting example discussed in Figure 2, the problem arrows in a DAG can be analyzed, these are depicted in Figure 3. The of modeling Van Gogh’s style can be viewed as estimating the leftmost graph (i) denotes a causal chain or that of a “mediation". causal effect of Van Gogh on the artwork. Thus, the expression The effect of 푋 on 푍 is mediated through 푌. In this case, given 푌 푃 (퐴푟푡푤표푟푘|푑표(퐴푟푡푖푠푡)) models the style of the artist in the artwork (i.e. conditioning on 푌 ), 푋 is independent of 푍 or 푌 is said to 푏푙표푐푘 (Section 5.2). It is to be noted that the assumptions encoded through FAccT, March 3–10, 2021, Virtual Event, Canada, , Ramya Srinivasan and Kanji Uchino a DAG can vary from one expert opinion to another. However, these varying opinions help to discover and test for biases under different scenarios and in turn highlight the drawbacks of existing correlation based methods such as [70] that do not take into account important cultural and social aspects that influence an artist’s style. Consider Figure 4 (i) that depicts one potential process of art creation (note there could be several others based on assumptions of domain experts, we consider one such for illustration). Here, the variable 푋 denotes the artist, 푍 denotes the artwork, 퐺 is the genre, 푀 is the art material, and 퐴 denotes the art movement. Accord- ing to the assumptions encoded in this DAG, art material, genre, and art movement are confounders influencing both the artist and the artwork. Further, art movement influences the art material. Figure 4: (i): Confounding bias due to genre, material and art move- Let us assume that all of the confounders are observable. Under ments. (ii): Confounding bias due to artist’s unobserved emotions. these assumptions, in order to model the style of Van Gogh (i.e. 푃 (푍 |푑표(푋 = 푥))), we have to block the backdoor path from 푋 to 푍 (Section 5.1), so as to remove confounding bias. Specifically, for graph (i) in Figure 4, the following equation captures the causal effect of 푋 on 푍

푃 (푍 |푑표(푋 = 푥)) = ∑︁ 푃 (푍 |푥,퐺 = 푔, 퐴 = 푎, 푀 = 푚)푃 (퐺 = 푔, 퐴 = 푎, 푀 = 푚) (1) 푔,푎,푚 Figure 5: (i): Example of selection bias (ii): Illustration of CS 2, see Section 6.2. (iii) Illustration of selection bias in datasets, see Section 7.1. Causal effect of 푋 on 푍 is not identifiable across (i), (ii), and (iii) For the case study, 푑표(푋 = 푥) implies do(Artist=Van Gogh). The Í summation 푔,푎,푚 in Eq. 1 indicates that one has to consider all possible art movements, art materials, and genres that the artist has 6.2 Sample Selection bias worked in order to model their style. The implication of finding a Sample selection bias (or selection bias for short) is the bias that sufficient set, 퐴,퐺, 푀, is that stratifying on 퐴,퐺, 푀 is guaranteed to is induced by preferential selection of units for data analysis [10]. remove all confounding bias relative to the causal effect of 푋 on 푍. In a DAG, a special variable 푆 is used to denote the selection of Thus, modeling artist’s style requires knowledge about the causal the variable in the analysis, 푆 = 1 indicating selection and 푆 = 0 process governing the artwork’s creation. Models like [54, 70] that indicating otherwise. Consider for example, Figure 5 (i). Here, 푋 → do not consider the influence of confounders like art movement 푆 and 푍 → 푆 indicates that both inputs and outputs are selection are prone to omitted variable bias [66] and confounding bias. Art dependent (i.e. 푆 = 1 with respect to both 푋 and 푍). The case study movements were characterized by many socio-cultural and political discussed below will further clarify these concepts. events and reflect the dynamic influence of these events on art. Thus by ignoring this variable, there is a bias in capturing the artist’s 6.2.1 Case Study 2. To illustrate selection bias, let us consider an style, in understanding the art’s intent, and in representing culture. example described in the ArtGAN model [54]. The authors state Based on the DAG, the causal effect may or may not be identifiable. that their model is able to recognize artist Gustave Dore’s prefer- Eq. (2) differs from the conditional distribution 푃 (푍 = 푧|푋 = 푥), ences as the generated images resonate with the dull color found and the difference between the two distributions, i.e. 푃 (푍 |푑표(푋 = in Dore’s artworks. Mostly engravings were selected for analysis. 푥)) − 푃 (푍 = 푧|푋 = 푥) defines confounding bias [10]. Graph (ii) in Figure 5 depicts a possible DAG for this case. Let 푋 Figure 4 (i) depicted a scenario with no unobserved confounders. denote Dore’s style. 푋 → 푆 depicts the selection of engravings in In presence of unobserved confounders, causal effects are not iden- the analysis. Further, as the authors mention, the generated images tifiable. Consider Figure 4 (ii). Let 퐸 represent emotions of the artist. are greyish due to the engravings considered, thus 푍 → 푆, where The dotted circle and lines denote the unobserved confounder 퐸 푍 denotes generated image. Additionally, there may be some unob- and its influence on other variables. In this case, even knowing the served confounders that influence both 푋 and 푍 as denoted by the joint distribution of genre, art movement, and material 푃 (퐴,퐺, 푀) bi-directional dotted arrow, and there may be some other Artgan does not help in identifying Van Gogh’s style, i.e. 푃 (푍 |푑표(푋 = 푥)) model variable 푉 that influences the nature of generated image 푍, is not identifiable. In general several subjective factors like prior these are depicted in Figure 5 (ii). beliefs, emotions, cultural values, etc. can influence the artist and In order to be able to recover the causal effect of 푋 on 푍 under the artwork. Thus in reality, the true style of any artist cannot be selection bias, [10] lists the conditions known as selection backdoor accurately modeled. Even if confounding bias due to unobserved criteria. Formally, let 푌 be a set of variables partitioned into two factors is overlooked, there are other biases as discussed next. groups 푌 + and 푌 − such that 푌 + contains all non-descendants of 푋 Biases in Generative Art— A Causal Look from the Lens of Art History FAccT, March 3–10, 2021, Virtual Event, Canada, ,

− and 푌 the descendants of 푋, and let 퐺푠 stand for the graph that includes the selection mechanism 푆. 푌 is said to satisfy the selection backdoor criterion if the following conditions are true: + • (i) 푌 blocks all backdoor paths from 푋 to 푍 in 퐺푠 + − • (ii) 푋 and 푌 block all paths between 푌 and 푍 in 퐺푠 • (iii) 푋 and 푌 block all paths between 푆 and 푍 in 퐺푠 • (iv) 푌 and 푌 ∪ (푋, 푍) are measured. In Figure 5 (ii), the path between 푆 and 푍 is not blocked due to the presence of direct link between the two variables, thus condition (iii) in selection backdoor criteria is not satisfied. Hence Dore’s Figure 6: (i): An example Selection diagram (SD). (ii): SD for case style cannot be recovered based on the engravings considered in study 4 and 5. (iii) SD illustrating case study 10. Causal effect of 푋 the analysis. In fact, in addition to engravings, Dore’s worked on on 푍 is not identifiable across all scenarios (i), (ii), and (iii) paintings. For example, landscapes such as ‘The Lost Cow’, paintings such as ‘Little Red Riding Hood’, religious painting such as ‘The Wrestle of Jacob’ are not greyish and exhibit colors reflective of the The conditions under which causal effects can be transported are genre. Thus, by merely selecting engravings for analysis, sample listed in [9]. Formally, let 퐷 be the selection diagram characteriz- ing the two populations Π and Π∗ with observational distributions selection bias is induced and style of Dore cannot be identified. ∗ Note, the failure to capture Dore’s style in [54] can be additionally 푃 and 푃 , and 푅 a set of selection variables in 퐷. The relation = ∗ ( | ( ) ) ∗ attributed to other types of biases such as confounding bias, label 푄 푃 푧 푑표 푥 ,푦 is transportable from Π to Π if and only if the ( | ( ) ) bias, and even framing effect bias. For the purpose of illustrating expression 푃 푧 푑표 푥 ,푦, 푟 is reducible, using rules of do-calculus selection bias, we have focused on describing selection bias only. [48] (Sec. 5.2), to an expression in which 푅 appears only as a condi- In general, there can be more than one type of bias in a case study. tioning variable in do-free terms. Given a DAG, open source tools like [8] can check for causal effects transportability automatically. 6.2.2 Case Study 3: As an other example of selection bias, consider As an illustration, consider Figure 6 (i). Let the variable 푋 de- the case of [2]. As illustrated in Figure 1, racial bias was evident in note artist and 푍 denote the artwork. Suppose the variable 푌 is this application. Figure 5 (ii) can also be used to describe this sce- an unobserved confounder denoting subjective emotions of the nario. Let 푋 denote the set of input images selected for analysis and artist. Between any two artists, this variable 푌 is bound to cause let 푍 represent the generated images. Selection of Renaissance por- differences and hence the selection variable 푅 is pointing to 푌 to traits of mostly white people was used in the analysis, thus 푋 → 푆. indicate this difference. For this DAG, the causal effect of 푋 on As evident, the generated images were portraits of fair skinned 푍 is not transportable or transferable across the two artists. For people, thus 푍 → 푆. Additionally, there could be unobserved con- illustrative purposes, we will ignore the differences due to unob- founders influencing both the input images 푋 and generated images served factors such as subjective emotions of artists and consider 푍, and there may be model parameters 푉 influencing the gener- differences in observed variables. There can still be transportability ated images. As condition (iii) in selection backdoor criteria is not bias as illustrated through the following case studies. satisfied, there is selection bias. 6.3.1 Case Study 4: In [70], a model that can transfer photograph to an artist’s style is proposed, say photograph and Cezanne for 6.3 Transportability bias illustration. For simplicity, we consider only one genre say land- Style transfer is a popular application in generative art. Various scapes. The goal is to model Cezanne’s style in rendering the land- works have demonstrated transferring across artists’ styles (e.g. scape corresponding to the photo. Consider Figure 6 (ii) which Cezanne to Monet) and across art media (e.g. photograph to paint- illustrates one possible selection diagram for this case study. Let ing). Learning models that can generalize across domains is com- 푋 denote artist/photographer and let 푍 denote the artwork/photo. monly known as transfer learning in the deep learning community Thus, there are two populations corresponding to the choice of 푋 and as transportability in the causality community. Transporta- and 푍, i.e. photographer/photo, and artist/artwork. For the style bility defines the conditions under which causal effects learned transfer problem, we want to be able to capture the causal effect of in experimental studies can be transferred into to a new popula- the artist 푋 on the artwork 푍 using the photograph. The shaded tion in which only observational studies can be conducted [49]. squares marked by the symbol 푅 are the selection variables and The differences between the populations of interest are expressed are used to denote differences in the two populations10 [ ]. Further, through representations called as “selection diagrams”. To this end, there may other unobserved confounders. However, to illustrate DAGs are augmented with a set, 푅, of “selection variables,” where biases beyond the difference in unobserved confounders, we will each member of 푅 corresponds to a mechanism by which the two overlook such confounders in this case study. populations differ, and switching between the two populations The factors that influence an artist are different from those that will be represented by conditioning on different values of these influence a photographer. For example, Cezanne could be influ- 푅 variables. 푅 variables locate the mechanisms where structural enced by the art movement whereas the photographer may be discrepancies between the two domains are suspected to take place. influenced by the photography trends. This distinction is indicated Transportability bias arises if causal effects cannot be transferred by the selection variable 푅 pointing into 푋. Further, the factors across populations. that affect the artwork may be different in the two populations. A FAccT, March 3–10, 2021, Virtual Event, Canada, , Ramya Srinivasan and Kanji Uchino

Figure 7: Center: “Propellers", a Cubist work by Fernand Leger. Right: Figure 8: Center: “Miss Mary Ellison", a Realism artwork by Mary “Armoured Train in Action", a Futurism work by Gino Severini. Left: Cassatt. Right: “Erna" , an Expressionism artwork by Ernst Ludwig translation of the center image according to the style of right im- Kirchner. Left: translation of the center image according to the style age by [20]. Movement, a key aspect of Futurism, is missing in the of right image by [20]. Distorted subjects, a key aspect of Expression- translation. Image source: [60] ism is absent in the translation. Image source: [60]

photograph may be subject to the camera characteristics, lighting, art as a symbol for expressing political and social views. Severini and measurement errors, selection variable 푅 pointing to 푍 denotes depicted aspects of war, movement, and modernity in this work. this difference. When the distinction is associated with the target Figure 6 (ii) can be a potential DAG for this case study, note, variable, i.e. 푍 in Figure 6 (ii), causal effects are not transportable there can be other DAGs based on assumptions. The differences [9]. Thus, there will be transportability bias. in Cubism and Futurism art movements combined with the differ- In fact, in post-impressionism, the art movement primarily as- ences in genre influences the artists differently and the artwork sociated with Cezanne, artists had their own individual styles. As differently. Also, there is the effect of unobserved factors suchas mentioned in [59], Cezanne concentrated on pictorial problems the artist’s emotions that influence the artist and the artwork. The of creating depth in his landscapes. He used an organized system left most image in Figure 7 corresponds to the “Futurism version” of of layers to construct a series of horizontal planes, which build Propellers. Kinetic patterns which is a distinct feature of Futurism, is dimension and draw the viewer into the landscape. This technique absent in the translated image. Given that the original image is that is apparent in some of his works such as Mont Sainte-Victoire, the of a mechanical object (propellers), a Futurism version of it should Viaduct of the Arc River Valley and The Gulf of Marseille Seen from have depicted the movement of the propellers much like in ‘Ar- L’Estaque [59]. In some of his works such as Gardanne, Cezanne moured train in Action’ that shows the fractured landscape, which painted the landscape with intense volumetric patterns of geometric accentuates the train’s force and momentum as it cuts through the rhythms most pronounced in the houses reflective of Cubism. The countryside [46]. Thus, there is bias in transferring styles. generated images in [70] do not exhibit such geometric rhythms. 6.3.3 Case Study 6: The previous case study encompassed style 6.3.2 Case Study 5: In the previous case study, we analyzed bias transfer across art movements which were similar in many ways. in the context of style transfer from one genre to another (from We now consider a case study involving style transfer between photograph to landscape). As another illustration, let us consider Realism and Expressionism, two art movements that have marked the problem of style transfer across art movements and genres. differences from one another. We consider common genre, namely In order to demonstrate transportability bias in this setting, we portraits across the two art movements. Thus, this case study serves consider DeepArt.io [20], an online platform that maps the style as a good test to see if style transfer from [20, 26] is effective given of one image to the content of the other using a neural network that the difference between two styles is significant. architecture [26]. We consider a case study that involves Cubism Realism focuses on representing subject matter truthfully, with- and Futurism, two art movements in the modern art era. Both out artificiality and avoiding implausible, exotic and supernatural these movements had many common aspects, yet they diverged in elements [60]. The center image in Figure 8 is an Realism portrait subtle ways. Therefore, we find it to be an interesting case study by Mary Cassatt. Expressionists, on the other hand, used gestural for analysis. Both Cubism and Futurism focus on representation brushstrokes and distorted subjects to portray intense emotions of objects from multiple perspectives/viewpoints and emphasize through their works. The right image in Figure 8 is an expression- on geometrical shapes. Cubism, is concerned with forms in static ism portrait by Ernst Ludwig Kirchner. As can be observed, the relationships while Futurism is concerned with them in a kinetic facial features in the right image have been distorted (e.g. sharp state. Futurism emphasized on objects and events that involved chin resulting in an almost triangular facial contour, pointed nose movement such as wars, energy of nightclub, and so on [62]. As and ears) to intensify emotions. The left image in Figure 8 is the such, we consider the following artworks for the case study. style translated version of the center image. Aesthetic innovations “Propellers" (center image in Figure 7) was a 1918 Cubist art by typical of an avante-garde movement like Expressionism are not Fernand Leger. Leger was fascinated with technology, in particular evident in the style translated version. As Kirchner himself said, in with propellers, and viewed them as objects of beauty holding them Expressionism, objective correctness of things is not emphasized close to sculptures [4]. Right image in Figure 7 is a 1915 Futurism [52], rather a new appearance is created through radical distortions art by Gino Servini called “Armored Train in Action". Severini was of subjects to evoke intense emotional experiences. The style trans- inspired by Cubism but was a member of Futurism. Futurism used lated version is exactly as the original but for some color changes. Biases in Generative Art— A Causal Look from the Lens of Art History FAccT, March 3–10, 2021, Virtual Event, Canada, ,

Figure 9: (i): “Black Matriarch", a Folkart by Clementine Hunter. (ii): Figure 10: (i) Portrait of a man by Raphael, (iii) Portrait of a young “Expressionism version" of (i) by [29]. (iii): “Giovinetto", a Renaissance man by Cosimo. (ii), (iv): Gender translations of (i) and (iii) respec- sculpture by Desiderio da Settignano. (iv): “Expressionism version" tively. Young men with long hair were mistaken as women by [1] of (iii) by [29]. Face color of “Black Matriarch" is changed in the translation unlike in “Giovinetto", a white sculpture. Image source: [60] 7.1 Representational bias The bias that arises because of having a dataset that is not repre- The image neither exhibits any distorted features nor has gestu- sentative of the real world is referred to as representational bias. ral brushstrokes that portray intense emotions. Thus, the style This is a particular type of selection bias. Specifically in the con- translation does not capture the subtleties of the Expressionism art text of art datasets, there may be imbalances with respect to art movement. genres (e.g. large number of photographs vs few sculptures), artists 6.3.4 Case Study 7: As another case study to demonstrate biases (e.g. mostly European artists vs few native artists), art movements in “style" transfer, we consider “GoART" [29]. This app allows a (large number of works concerning Renaissance and modern art user to convert an uploaded photo into various styles spanning art movements as opposed to others), and so on. The availability of movements such as Byzantine, Expressionism, Cubism, Ukiyo-e, artworks is one of the main constraints in collecting a dataset that and artists such as Van Gogh. We consider a 1970 folk art by Clemen- is representative of the bygone times, but preferences of the dataset tine Hunter titled “Black Matriarch" shown in Figure 9 (i). Figure 9 curators can also play a role in contributing to bias. (ii) shows the “Expressionism" version of the “Black Matriarch" from 7.1.1 Case Study 9: Consider [2] that was trained using about [29]. As can be noticed, the face is tinted in red. However, this kind 45000 Renaissance portraits of mostly white people [38, 53]. Quite of effect is not pronounced in light colored faces. Consider Figure naturally, the system performs poorly on dark skinned people. Faces 9 (iii). This is an early Renaissance sculpture by Desiderio da Set- depicting different races, appearances, etc. have not been pooled tignano. The face in the “Expressionism" version of this sculpture into the dataset, thus contributing to representational bias. This is (Figure 9 (iv)) does not seem to have shades of red as significant as a particular instance of selection bias that has to do with dataset in Figure 9 (ii). Similar results were observed with other styles such curation. Algorithms trained using datasets with severe class imbal- as Byzantine wherein the face of “Black Matriarch" was tinted with ances are bound to be biased. Figure 5 (iii) illustrates this. Suppose shades of blue while the fairer faces were not heavily tinted. There 푋 denotes artist and 푍 denotes artworks, then class imbalances is noticeable difference in the way faces are converted across styles corresponds to 푍 → 푆, i.e. type of artworks influences selection based on the color of the face, indicating potential racial biases. into the dataset (in this case mostly white Renaissance portraits 6.3.5 Case Study 8: We consider Abacus.AI’s online tool that con- were selected into the dataset). As condition (iii) fails in selection verts a user uploaded photo into a different gender: male to female backdoor criteria, there is representational bias. and vice versa [1]. One of their demos shows translation of the Renaissance painting of Mona Lisa into a masculine face. Thus, we 7.2 Label bias experimented with other Renaissance paintings. Figure 10 (i) is a This type of bias is associated with the inconsistencies in the label- ‘portrait of a man holding an apple’ by Raphael and Figure 10 (iii) ing process. Different annotators may have different preferences is a ‘portrait of a young man’ by Italian artist Piero di Cosimo [60]. which can get reflected in the labels created. A common instance Figure 10 (ii) and 10 (iv) are the gender translated versions of (i) of label bias arises when different annotations could be used to and (iii) by [1], both of which fail to identify the original paintings represent an artwork. For example, a scene with clouds may be as those of men. Young men with long hair were mistaken to be annotated as a “cloudscape" by some annotators and more generally women and thus the gender-translated versions of these images as a "landscape" by others. depicted masculine faces with beards. In the Renaissance era, it was Yet another type of label bias arises when subjective biases of common for young men to have long hair, often extending from evaluators can affect labeling. “Confirmation bias"50 [ ], a type of ears to shoulders. Thus, by not understanding the differences due human bias, is closely related to this type of label bias. For example, to culture across genders and age, men are being stereotyped as in a task of annotating emotions associated with artworks, the labels having short hair and [1] thus exhibits transportability bias. could be biased by the subjective preferences of annotators such as their culture, beliefs, and introspective capabilities. Consider the 7 BIASES IN DATASETS Behance Artistic Media dataset [65] which provides labels based In this section, we discuss biases due to unrepresentative datasets on emotions such as “happy", “scary", "peaceful", etc. Such labels and due to inconsistencies in annotation. could be based on annotator’s beliefs, and can therefore be noisy. FAccT, March 3–10, 2021, Virtual Event, Canada, , Ramya Srinivasan and Kanji Uchino

7.2.1 Case Study 10: To illustrate how annotation inconsistencies In a recent photo booth titled “Latent Face", latent vectors of can induce bias, consider the ArtGAN model [54]. This model uses the styleGAN model were combined to generate "hypothetical chil- the label information to train the GAN’s discriminator. The authors dren" of subjects depicted in the original portraits [5]. While this claim that by using labels pertaining to genre (cityscapes, por- may have been an exploratory experiment, the task exemplifies traits, etc.), art media (e.g. sketch and study, engraving), and style framing effect bias. Defining “children" as combination of latent (e.g. Ukiyo-e), they are able to generate images of those categories. vectors is highly questionable. The latent vectors may or may not However, the annotations are not necessarily reliable indicators of have any reliable interpretation, and a very difficult and potentially genres or styles. For example, “sketch and study" category includes impossible problem of generating faces of hypothetical children is several other categories such as portraits, religious paintings, and conveniently framed as a simple task. There are also several ethical allegorical works to name a few. Figure 6 (iii) illustrates this bias. Let concerns associated with such a framing given that people depicted Π and Π∗ denote the environments corresponding to two different in the original portraits were not related or never had children. annotators. Suppose 푋 denotes art movement’s style (e.g. Ukiyo-e), Framing effect bias coupled with algorithmic bias contributes to and 푍 is the artwork. Annotation inconsistencies across annota- inaccurate knowledge about history. Artworks were often meant tors affects the artworks’ labels, this is indicated by the selection to document important events in history such as wars, mytho- variable 푅 pointing to 푍. When there are differences in the target logical events, political movements, etc. By wrongly modeling or variable (i.e. label of artworks), the causal effect of 푋 on 푍 is not overlooking certain subtle aspects, generative art can contribute identifiable using the annotations provided [9]. Thus, a generative to false perceptions about social, cultural and political aspects of model that leverages such labels in modeling style is prone to bias. past times and hinder awareness about important historical events. For example, as discussed in case study 5, in the Futurism artwork “Armored Train in Action", artist Severini conveys his views on war 8 DISCUSSION that was prevalent during the time. In fact, Futurism artists heavily Art is much more than an aesthetic entity. As elaborated in [68], art depicted their views of political events through patterns indicating imparts “moral knowledge", i.e. knowledge about what ought to do movement in the artworks. A generated style translation should and not to do [31]. Art also enables “emphathic knowledge", some- thus preserve such important characteristics of art movements, or thing through which one can compare different views of the world else they will be contributing to a bias in understanding history. through direct experience [18]. Art initiates a conversation with People have a propensity to favor suggestions from automated sys- the public [19], it is a form of language that is not just a mimicry tems and to ignore contradictory information, even if it is correct. but a symbolic transposition [41]. Art is ‘a form of technology that This is called as “automation bias" [61]. Thus, people may give lit- contributes to knowledge production by exemplifying aspects of the tle importance to historical evidence. Further, often, evaluation of world that would otherwise go overlooked’ [31]. Thus given its pow- generative art is done by people (e.g. Amazon Mechanical Turk erful impact in shaping moral and empathetic values, generative workers) who do not necessarily possess domain knowledge. Thus, art comes with the responsibility of creating art that respects and even if generated art is not accurately representing cultural and upholds societal ethics. By coloring the face of the “Black Matri- historical knowledge, people may endorse them. arch", [29] is not only depicting racial bias, but also inaccurate in its Tutorials like [27] and [39] emphasize on the need to inspect the representation of the art movement. As artist Edgar Degas remarks, design choices made by the creators of AI systems and the socio- “Art is not just what you see but what you make others see". Thus, political contexts that shape their deployment. Can algorithms generative art that does not promote diversity and inclusiveness accurately model artist’s “styles"? More broadly, is the defined prob- has the potential of creating and communicating unethical values. lem even solvable? How can representative datasets and reliable Art reflects cognitive abilities of the artist [21]. Cognitive abili- labels be created to address the defined problem? What are the ties include perception, memory, emotions, and other latent aspects measurement biases in digitally capturing art? What are the socio- about the artist. Generative art that is meant to create art in the cultural impacts of generative art? Does generative art promote “style" of various artists must reflect and respect artist’s cognitive inclusiveness? Who should own responsibility for biases in genera- abilities and not stereotype them based on any narrowly defined tive art? These are just some questions that need to be analyzed. metric. For example, often, “style of Van Gogh" is largely modeled Also, domain experts (e.g. art historians) should be involved in the based on the brushstrokes in his rendition “Starry Nights" or based generative art pipeline to better inform the process of art creation. on certain colors such as in [54]. Similarly, “style of Cezanne" in [70] little reflects the variety of geometric patterns that were prominent in his works. Needless to mention that cognitive aspects of the artist are not considered. As discussed in Section 4 and 6, majority 9 CONCLUSIONS of these issues arise due to framing effect bias and algorithmic bias. In this paper, we investigated biases in the generative art AI pipeline Often style is defined and modeled in a way to suit the algorithm’s from the perspective of art history. Leveraging structural causal performance. Not only are several important abilities and achieve- models, we highlighted how current methods fall short in modeling ments of artists overlooked in the process of poor style modeling, the process of art creation and illustrated instances of framing effect but also those pertaining to larger art movements are ignored. For bias, dataset bias, selection bias, confounding bias, and transporta- example, advanced techniques such as exaggerated foreshorten- bility bias. We also discussed the socio-cultural impacts of these ing and perspective modeling that were typical of many Ukiyo-e biases. We hope our work sparks inter-disciplinary dicussions and renditions are hardly visible in the generated versions [29, 70]. inspires new directions concerning accountability of generative art. Biases in Generative Art— A Causal Look from the Lens of Art History FAccT, March 3–10, 2021, Virtual Event, Canada, ,

REFERENCES [38] Edward Ongweso Jr. 2019. Racial Bias in AI Isn’t Getting Better and Neither Are [1] Abacus.AI. 2020. Effortlessly embed cutting edge AI in your Applications. In Researchers’ Excuses. https://www.vice.com/en_us/article/8xzwgx/racial-bias-in- https://abacus.ai. ai-isnt-getting-better-and-neither-are-researchers-excuses (2019). [2] AIportraits. 2020. AIportraits: The easiest way to make your portraits look [39] Christine Kaeser-Chen, Elizabeth Dubois, Friederike Schuur, and Emanuel Moss. stunning. https://aiportraits.org (2020). 2020. Positionality-aware machine learning: translation tutorial. FAccT Tutorial [3] Artbreeder. 2020. Artbreeder: Extend your Imagination. (2020). https://www.artbreeder.com (2020). [40] Matt J. Kusner, Joshua R. Loftus, Chris Russell, and Ricardo Silva. 2017. Counter- [4] Christoph Asendorf. 1994. The Propeller and the Avant-Garde: Leger, Duchamp, factual Fairness. NeurIPS (2017). Brancusi. Fernand leger: The Rhythm of Modern Life (1994). [41] A. Leroi-Gourhan. 1993. Gesture and Speech. MIT Press (1993). [5] Jason Bailey. 2019. Breeding Paintings With Machine Learning. [42] Heidegger M. 1977. The question concerning technology, and other essays. https://www.artnome.com/news/2019/8/25/breeding-paintings-with-machine- Garland Publishing INC (1977). learning (2019). [43] Dan Macnish. 2020. Draw this. https://danmacnish.com/drawthis/ (2020). [6] Jason Bailey. 2020. 2020 Art Market Predictions. [44] Authur Miller. 2019. The Artist in the Machine The World of AI-Powered Cre- https://www.artnome.com/news/2020/1/27/2020-art-market-predictions (2020). ativity. MIT Press (2019). [7] Jason Bailey. 2020. The tools of generative art from flash to neural net- [45] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. 2015. Deep Dream. works. https://www.artnews.com/art-in-america/features/generative-art-tools- https://github.com/google/deepdream (2015). flash-processing-neural-networks-1202674657/ (2020). [46] MoMA: Musuem of Modern Art. 2020. Gino Severini: Armoured Train in Action. [8] Elias Bareinboim et al. 2020. Causal Fusion. In https://causalfusion.net. In https://www.moma.org/collection/works/33837. [9] Elias Bareinboim and Judea Pearl. 2012. Transportability of Causal Effects: [47] Oxford Art Online. 2020. Impressionism and Post-Impressionism. Completeness Results. In AAAI. https://www.oxfordartonline.com/page/impressionism-and-post- [10] Elias Bareinboim and Judea Pearl. 2016. Causal inference and the data-fusion impressionism/impressionism-and-postimpressionism (2020). problem. Proceedings of the National Academy of Sciences (2016). [48] Judea Pearl. 2009. Causality: Models, Reasoning and Inference, 2nd Edition. [11] Margaret Boden and Ernest Edmonds. 2009. What is Generative Art. Digital Cambridge University Press (2009). Creativity (2009). [49] Judea Pearl and Elias Bareinboim. 2014. External Validity: From Do-Calculus to [12] Rodney Brooks. 2017. The Seven Deadly Sins of AI Predictions. MIT Technology Transportability Across Populations. Statistical Sciences (2014). Review (2017). [50] Scott Plous. 1993. The psychology of judgment and decision making. Mc-Graw [13] McAndrew C. 2019. The art market. An Art Basel and UBS Report. Hill (1993). https://www.artbasel.com/about/initiatives/the-art-market (2019). [51] Martin Ragot, Nicolas Martin, and Salomé Cojean. 2020. AI-generated vs. Human [14] Reas C and Fry B. 2007. Processing: a programming handbook for visual designers Artworks. A Perception Bias Towards Artificial Intelligence? CHI: Late Breaking and artists. MIT Press (2007). Work (2020). [15] Christies. 2018. Is artificial intelligence set to become art’s next medium? [52] The Art Story. 2020. Ernst Ludwig Kirchner - Biography and Legacy. In https://goo.gl/4LDZjX (2018). https://www.theartstory.org/artist/kirchner-ernst-ludwig/life-and-legacy/#nav. [16] Mark Coeckelbergh. 2017. Can Machines Create Art? Philosophy and Technology [53] Morgan Sung. 2019. The AI Renaissance portrait generator isn’t great at painting (2017). people of color. https://mashable.com/article/ai-portrait-generator-pocs/ (2019). [17] Kate Crawford and Trevor Paglen. 2019. Excavating AI: The Politics of Images in [54] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. 2017. Machine Learning Training Sets. https://www.excavating.ai (2019). ArtGAN: Artwork Synthesis with Conditional Categorical GANs. ArXiV (2017). [18] Novitz D. 1987. Knowledge, Fiction, and Imagination. Temple University Press [55] Value. 2019. AI Artwork Goes Up for Auction for 30, 000 at Sotheby’s. (1987). https://en.thevalue.com/articles/sothebys-ai-memories-of-passersby (2019). [19] Antonio Daniele and Yi-Zhe Song. 2019. AI+Art= Human. AAAI AI Ethics and [56] Vincent van Gogh. 1886. Letter to Horace M. Livens, Translated by Robert Society (2019). Harrison. http://www.webexhibits.org/vangogh/letter/17/459a.htm (1886). [20] Deepart.io. 2020. Deepart.io. https://deepart.io (2020). [57] Vincent van Gogh. 1888. Letter to Wilhelmina van Gogh, [21] E. Dissanayake. 2001. Where art comes from and Why. University of Washington Translated by Translated by Mrs. Johanna van Gogh-Bonger. Press (2001). http://www.webexhibits.org/vangogh/letter/18/W04.htm (1888). [22] Katherine Dunbabin. 1999. Mosaics of the Greek and Roman World. Cambridge [58] VanGoghGallery. 2020. VINCENT VAN GOGH: POPPIES. University Press (1999). https://www.vangoghgallery.com/misc/poppies.html (2020). [23] Anna Brzyski (Editor). 2007. Partisan Canons. Duke University Press (2007). [59] James Voorhies). 2004. Paul Cézanne (1839–1906). Heilbrunn Timeline of Art [24] Ahmed Elgammal. 2019. Faceless Portraits Transcending Time. HG Contem- History (2004). porary New York https://uploads.strikinglycdn.com/files/3e2cdfa0-8b8f-44ea-a6ca- [60] Wikiart. 2020. Visual Art Encyclopedia. https://www.wikiart.org (2020). d12f123e3b0c/AICAN-HG-Catalogue-web.pdf (2019). [61] Wikipedia. 2020. Automation Bias. https://en.wikipedia.org/wiki/Automation_bias [25] Ahmed Elgammal, Bingchen Liu, Mohamed Elhoseiny, and Marian Mazzone. 2017. (2020). CAN: Creative Adversarial Networks, Generating "Art" by Learning About Styles [62] Wikipedia. 2020. Futurism. https://en.wikipedia.org/wiki/Futurism (2020). and Deviating from Style Norms. International Conference on Computational [63] Wikipedia. 2020. Generative Art. https://en.wikipedia.org/wiki/Generative_art Creativity (ICCC) (2017). (2020). [26] Leon A. Gatys, Alexander S Ecker, and Matthias Bethge. 2016. Image Style [64] Wikipedia. 2020. Ukiyo-e. https://en.wikipedia.org/wiki/Ukiyo-e (2020). Transfer Using Convolutional Neural Networks. Computer Vision and Pattern [65] Michael J. Wilber, Chen Fang, Hailin Jin, Aaron Hertzmann, John Collomosse, and Recognition (2016). Serge Belongie. 2017. BAM! The Behance Artistic Media Dataset for Recognition [27] Timnit Gebru and Emily Denton. 2020. Tutorial on Fairness Accountability Beyond Photography. ICCV (2017). Transparency and Ethics in Computer Vision. CVPR Tutorial (2020). [66] Jeffrey M. Wooldridge. 2009. Omitted Variable Bias: The Simple Case. In Intro- [28] Bruce Glymour and Jonathan Herington. 2019. Measuring the Biases that Matter. ductory Econometrics: A Modern Approach. FaccT (2019). [67] David Young. 2019. Tabula Rasa: Rethinking the Intelligence of Machine Minds. [29] GoART. 2020. GoART: AI Photo Effects. http://goart.fotor.com.s3-website-us-west- https://medium.com/@dkyy/tabula-rasa-b5f846e60859 (2019). 2.amazonaws.com (2020). [68] James O. Young. 2001. Art and Knowledge. Routledge, London, UK (2001). [30] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- [69] Lieberman Z., Watson T., and Castro A et. al. 2009. Openframeworks. Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative https://goo.gl/41tycE (2009). adversarial nets. NeurIPS (2014). [70] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. 2017. Unpaired [31] Tim Gorichanaz. 2020. Engaging with Public Art: An Exploration of the Design Image-to-Image Translation using Cycle-Consistent Adversarial Networks. ICCV Space. CHI (2020). (2017). [32] David Ha and Douglas Eck. 2018. A Neural Representation of Sketch Drawings. International Conference on Learning Representations (2018). [33] Paul Haeberli. 1990. Paint By Numbers: Abstract Image Representations. SIG- GRAPH (1990). [34] Aaron Hertzmann. 2018. Can Computers create Art? Arts (2018). [35] Aaron Hertzmann, C. Jacobs, N. Oliver, B. Curless, and D.H. Salesin. 2001. Image Analogies. SIGGRAPH (2001). [36] Joo-Wha Hong. 2018. Bias in Perception of Art Produced by Artificial Intelligence. International Conference on Human Computer Interaction (2018). [37] Instapainting. 2020. AI painter. https://www.instapainting.com/ai-painter (2020).