[Paper Review] Quantifying Confounding Bias in Generative Art: A Case Study

Ramya Srinivasan, Kanji Uchino|arXiv (Cornell University)|Feb 23, 2021

Generative Adversarial Networks and Image Synthesis24 references4 citations

TL;DR

This paper proposes a causal metric using directed acyclic graphs (DAGs) to quantify confounding bias in AI-generated art when art movements are not modeled during style transfer. Applied to cycleGAN, the metric outperforms state-of-the-art outlier detection in identifying bias from unmodeled socio-cultural influences, revealing that ignoring art movements leads to stereotypical, incomplete representations of artists' styles.

ABSTRACT

In recent years, AI generated art has become very popular. From generating art works in the style of famous artists like Paul Cezanne and Claude Monet to simulating styles of art movements like Ukiyo-e, a variety of creative applications have been explored using AI. Looking from an art historical perspective, these applications raise some ethical questions. Can AI model artists' styles without stereotyping them? Does AI do justice to the socio-cultural nuances of art movements? In this work, we take a first step towards analyzing these issues. Leveraging directed acyclic graphs to represent potential process of art creation, we propose a simple metric to quantify confounding bias due to the lack of modeling the influence of art movements in learning artists' styles. As a case study, we consider the popular cycleGAN model and analyze confounding bias across various genres. The proposed metric is more effective than state-of-the-art outlier detection method in understanding the influence of art movements in artworks. We hope our work will elucidate important shortcomings of computationally modeling artists' styles and trigger discussions related to accountability of AI generated art.

Motivation & Objective

• To investigate how unmodeled art movements introduce confounding bias in AI-generated art.
• To develop a quantitative metric for measuring bias due to the absence of art movement influence in style modeling.
• To assess whether current generative models like cycleGAN fail to capture socio-cultural nuances of artists' styles.
• To explore the implications of such bias for authenticity, valuation, and art historical understanding of AI-generated works.

Proposed method

• Constructs a directed acyclic graph (DAG) to model causal relationships between art movements, artists, and artwork characteristics.
• Defines a confounding bias metric based on the deviation of predicted style features from expected values when art movement is unobserved.
• Applies the metric to cycleGAN-generated images across genres, using DAG-structured assumptions about causal dependencies.
• Uses stratified loss minimization during training as a baseline to compare bias scores.
• Validates the metric against state-of-the-art outlier detection methods to assess its effectiveness in identifying confounded outputs.
• Proposes the metric as a tool for authenticity assessment, pricing, and art historical analysis by quantifying cultural and stylistic fidelity.

Experimental results

Research questions

• RQ1 How does the absence of art movement modeling introduce confounding bias in AI-generated art?
• RQ2 To what extent does the proposed metric detect bias more effectively than existing outlier detection methods?
• RQ3 Can the metric serve as a proxy for authenticity or cultural fidelity in AI-generated artworks?
• RQ4 How do different art genres (e.g., Impressionism vs. Post-Impressionism) differ in their susceptibility to confounding bias?
• RQ5 In what ways can the bias score inform art valuation or support art historical inquiry?

Key findings

• The proposed confounding bias metric outperforms state-of-the-art outlier detection methods in identifying the influence of unmodeled art movements on generated artworks.
• Artists' styles are systematically misrepresented when art movements are excluded from the modeling process, leading to stereotypical outputs based on superficial features like color and brushwork.
• The bias metric reveals that cycleGAN-generated art exhibits higher confounding bias when art movement is not accounted for, especially in genres like Impressionism and Post-Impressionism.
• The metric can be used as a complementary tool in art authentication, where lower bias scores correlate with higher likelihood of genuine artist-style generation.
• The metric provides a quantitative basis for assessing the cultural and historical fidelity of AI-generated art, supporting more accountable AI in creative applications.
• Art historians can use the metric to compare and validate differing perspectives on art movements by encoding their assumptions in DAGs and analyzing resulting bias scores.