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

1 INTRODUCTION

• Generative art uses autonomous systems, most commonly algorithms or computer programs, to create artworks.Its applications include hybrid images, portrait-to-cartoon generation, and style creation.Source: §1 · p. 1
• Rapid adoption, commercial sales, museums, and accessible applications have made AI-based generative art increasingly prominent.The paper notes major auction sales and growing institutional and user interest.Source: §1 · p. 1
• Training data and algorithms can embed racial and gender bias, stereotype artists’ styles, overlook intent and emotion, and distort historical or cultural understanding.The paper illustrates these concerns with lightened skin in AIportraits and lost affect in a generated “Van Gogh” image.Source: §1 · p. 2

1.1 Contributions

• The paper investigates biases from problem formulation through algorithm design, using causal models to represent art-creation processes and case studies across art-historical settings.Its cases cover art movements, artists, materials, genres, and geographies, including framing, confounding, selection, and transportability concerns.Source: §1.1 · p. 2

1.2 Case study selection

• The case studies focus on established art movements and artists’ styles, including Renaissance, Modern, and Ukiyo-e art across varied genres, materials, and artists.Ukiyo-e is the only non-western art form included because existing AI applications studying non-western forms are scarce.Source: §1.2 · pp. 2–3

2 AI FOR ART GENERATION

• Computer-generated art has a long history spanning AARON, interactive computer experiences, animation tools, paint programs, Processing, and OpenFrameworks.These systems enabled artists to create paintings, interactive works, animations, and code-based art.Source: §2 · p. 3
• Deep learning expanded AI art through neural style transfer, CNN-based visualization, recurrent stroke-based drawing, and GAN-based generation.Neural style transfer blends one image’s content with another’s style, while other systems generate drawings or artworks.Source: §2 · p. 3
• Applications let users transform portraits, convert photos into artworks, combine images, transfer styles, and create cartoons with limited technical effort.The paper names AIportraits, GANbreeder, and other tools as examples.Source: §2 · p. 3
• The paper argues that widespread tools, a growing art market, and demands for diversity make bias analysis in generative art particularly pertinent.It positions its work as an extensive art-history-based analysis addressing a gap in prior studies.Source: §2 · p. 3

3 ART-HISTORICAL ASPECTS OF ARTWORKS

• Art historians classify artworks using significant qualities such as artistic approach, material, movement, and genre.These aspects provide the framework for analyzing generative-art biases.Source: §3 · p. 3

3.1 Art Movements

• Art movements reflect shared philosophies shaped by culture, geography, politics, and historical period, while individual movements contain distinct stylistic subcategories.Impressionism and Post-Impressionism illustrate how related movements can differ in colors, brushwork, subjects, and emotional expression.Source: §3.1 · p. 4

3.2 Art material

• Materials and techniques influence artists and artworks, with different periods and movements favoring materials such as marble, paint, wood, or mosaics.The paper treats material and technique as art-historical dimensions relevant to artwork classification.Source: §3.2 · p. 4

3.3 Genre

• Genres organize artworks by depicted themes and objects, including history painting, portraiture, genre painting, landscapes, still life, and abstract or figurative art.Artists may work across genres even when art historians associate them primarily with one.Source: §3.3 · p. 4

3.4 Artists

• An artist’s style combines movement, material, genre, cultural background, artistic lineage, cognition, beliefs, and prejudices.Cezanne’s changing colors, genres, movements, emotions, and constructive brushstrokes exemplify this multidimensionality.Source: §3.4 · p. 4
• Because artists’ works span genres, movements, and subtle cognitive aspects, style is difficult to quantify and is often defined to suit model performance.The paper identifies this modeling choice as a source of bias examined in the following section.Source: §3.4 · p. 4

4 BIASES DUE TO PROBLEM FORMULATION

• Framing-effect bias arises when researchers define an artistic problem in a way that suits model performance.Style may refer to an art movement or an individual artist, and each notion contains multiple aspects.Source: §4 · p. 4
• Models have treated Ukiyo-e style primarily as yellowish coloration, even though Ukiyo-e also includes foreshortening, asymmetry, flat color, and imaginative cropping.The cited examples therefore do not capture several characteristics of the movement.Source: §4 · pp. 4–5
• Inappropriate problem formulation can introduce and perpetuate bias throughout the generative-art AI pipeline.The paper distinguishes this issue from additional biases arising from model design.Source: §4 · p. 5

5 STRUCTURAL CAUSAL MODELS

• Causal DAGs encode relationships among artwork variables, expose assumptions about data, models, and analysis, and support bias analysis under multiple expert-specified scenarios.The paper frames DAGs as accessible visualization and analysis tools for art historians.Source: §5 · p. 5

5.1 d-separation

• D-separation determines whether variables X and Z are independent given Y by treating connected paths as dependence and separated paths as independence.The framework analyzes mediation, confounding, and collider structures, with admissible adjustment sets blocking all backdoor paths without descendants of X.Source: §5.1 · p. 5, Figure 3 · p. 5

5.2 Interventions

• The do operator sets an intervened variable to a chosen value while removing other influences, allowing expressions such as P(Z|do(X = x)) to quantify causal effects.Do-calculus provides rules for manipulating observations and interventions and eliminating do operators from expressions.Source: §5.2 · p. 5

6.1 Confounding bias

• Ignoring art movements and other confounders can introduce omitted-variable and confounding bias, misrepresenting artistic intent, culture, and style.The causal effect P(Z|do(X=x)) can differ from the observational distribution P(Z=z|X=x); their difference defines confounding bias.Source: §6.1.1 · p. 6
• Causal modeling of artist style requires representing the artwork-creation process, because unobserved emotions, cultural values, and art movements can confound the artist–artwork relationship.When emotions are unobserved, even the joint distribution of genre, material, and art movement cannot identify the artist’s causal style effect.Source: §6.1.1 · p. 6

6.2.1 Case Study 2.

• Selecting only engravings for ArtGAN analysis makes Gustave Dore’s style non-identifiable because the selection mechanism remains directly connected to generated-image characteristics.Dore’s paintings include non-greyish colors absent from the selected engravings, so the sample does not represent his broader style.Source: §6.2.1 · pp. 6–7

6.2 Sample Selection bias

• Selection bias also affects portrait generation when predominantly white Renaissance portraits are used, producing fair-skinned outputs because the selection criterion is associated with both inputs and outputs.The selection backdoor criterion fails because the path between selection and generated images is not blocked.Source: §6.2.2 · p. 7, §6.2.1 · p. 7

6.3 Transportability bias

• Style transfer is transportable only when domain differences do not disrupt the causal effect; differences in artists, genres, media, or unobserved factors can therefore create transportability bias.Selection variables encode structural differences between populations, and effects are not transportable when those differences are associated with the target artwork.Source: §6.3 · p. 7, §6.3.1 · pp. 7–8
• Across case studies, style-transfer systems omit defining features of Cubism, Futurism, Expressionism, and Cezanne’s style, while gender and racial translations exhibit culturally dependent stereotypes.Examples include missing kinetic patterns, absent subject distortion and gestural brushstrokes, missing geometric rhythms, red or blue facial tinting, and long-haired Renaissance men rendered as women.Source: §6.3.2 · p. 8, §6.3.3 · pp. 8–9, §6.3.1 · p. 8

7 BIASES IN DATASETS

• The dataset-bias discussion covers both unrepresentative datasets and annotation inconsistencies as distinct sources of bias in generative art.These sources correspond to representational bias and label bias, respectively.Source: §7 · p. 9, §7.1 · p. 9, §7.2 · p. 9

7.1 Representational bias

• Representational bias occurs when art datasets do not reflect the real world, including imbalances across genres, artists, art movements, and other categories.Dataset availability and curator preferences can both constrain representativeness.Source: §7.1 · p. 9
• A portrait generator trained on about 45,000 mostly white Renaissance portraits performs poorly on dark-skinned people because racial diversity was not represented.The paper models this imbalance as selection bias in which artwork type influences dataset inclusion.Source: §7.1.1 · p. 9

7.2 Label bias

• Label bias arises when annotators use inconsistent categories, such as labeling a cloudy scene as either “cloudscape” or “landscape.”Subjective preferences, culture, beliefs, and introspective capabilities can also make emotion labels noisy.Source: §7.2 · p. 9
• ArtGAN’s genre, media, and style labels may be unreliable because categories such as “sketch and study” contain heterogeneous artworks.Across annotators, inconsistent target labels make the causal effect of style on artwork unidentifiable, making label-driven generation prone to bias.Source: §7.2.1 · p. 10

8 DISCUSSION

• Because art conveys moral, empathetic, and symbolic knowledge, generative systems that fail to promote diversity and inclusiveness can communicate unethical values.The discussion treats representation as consequential beyond aesthetic similarity.Source: §8 · p. 10
• Generative art that models style through narrow proxies can overlook artists’ cognitive abilities and movement-specific techniques, producing stereotyped representations.Examples include reducing Van Gogh to brushstrokes or colors, Cezanne to limited geometric patterns, and obscuring Ukiyo-e foreshortening and perspective.Source: §8 · p. 10
• Defining latent-vector combinations as “hypothetical children” frames a difficult and ethically questionable face-generation task as a simple operation.The paper also notes that the depicted people may not have been related or had children.Source: §8 · p. 10
• Framing and algorithmic biases can produce inaccurate historical knowledge by omitting subtle artistic features that convey social, cultural, and political contexts.The paper argues that style translation should preserve characteristics such as Futurist visual patterns expressing views on war.Source: §8 · p. 10
• Automation bias and evaluation by people without relevant domain knowledge can lead audiences to endorse historically inaccurate generated art.The paper therefore calls for art historians and other domain experts to participate in the generative art pipeline.Source: §8 · p. 10

• Structural causal models reveal framing, dataset, selection, confounding, and transportability biases in generative art AI pipelines.The analysis highlights shortcomings in how current methods model the process of art creation.Source: §9 · p. 10
• The paper connects these biases to socio-cultural consequences and calls for interdisciplinary discussion about generative-art accountability.Source: §9 · p. 10