[Paper Review] Unmaking AI Imagemaking: A Methodological Toolkit for Critical Investigation

Luke Munn, Liam Magee+1 more|arXiv (Cornell University)|Jul 19, 2023

Aesthetic Perception and AnalysisNeuroscience3 citations

TL;DR

This paper introduces a three-pronged methodological toolkit—Unmaking the Ecosystem, Unmaking the Data, and Unmaking the Output—for critically investigating generative AI image models like Stable Diffusion. By analyzing production incentives, training data biases, and generative outputs through iterative prompting, the framework reveals how these models reproduce societal stereotypes and power imbalances, offering a foundation for politically and socially attuned AI research.

ABSTRACT

AI image models are rapidly evolving, disrupting aesthetic production in many industries. However, understanding of their underlying archives, their logic of image reproduction, and their persistent biases remains limited. What kind of methods and approaches could open up these black boxes? In this paper, we provide three methodological approaches for investigating AI image models and apply them to Stable Diffusion as a case study. Unmaking the ecosystem analyzes the values, structures, and incentives surrounding the model's production. Unmaking the data analyzes the images and text the model draws upon, with their attendant particularities and biases. Unmaking the output analyzes the model's generative results, revealing its logics through prompting, reflection, and iteration. Each mode of inquiry highlights particular ways in which the image model captures, "understands," and recreates the world. This accessible framework supports the work of critically investigating generative AI image models and paves the way for more socially and politically attuned analyses of their impacts in the world.

Motivation & Objective

• To address the lack of critical, accessible methods for investigating the sociotechnical operations of generative AI image models.
• To challenge the opacity and mythologizing surrounding AI image models by developing a multidisciplinary methodological framework.
• To enable researchers to uncover embedded biases in training data, model architectures, and generative outputs through systematic inquiry.
• To support a broader political and ethical engagement with AI image generation beyond technical evaluation.
• To provide a flexible, adaptable toolkit applicable across different models, especially open-source ones like Stable Diffusion, despite varying levels of transparency.

Proposed method

• Unmaking the Ecosystem: Analyzes the commercial, institutional, and incentive structures behind model development, including corporate ownership, funding, and organizational goals.
• Unmaking the Data: Investigates the composition, provenance, and biases of the training data—text-image pairs—highlighting how historical and cultural imbalances are encoded in model behavior.
• Unmaking the Output: Uses iterative prompting, reflection, and interface-based experimentation to reveal model logics, stereotypes, and aesthetic tendencies in generated images.
• Combines methods from critical AI studies, digital methods, and media studies to achieve a holistic, sociotechnical analysis of AI image models.
• Employs case study methodology using Stable Diffusion as a representative open-source model to demonstrate the toolkit’s applicability.
• Emphasizes methodological pluralism and 'interpretative pollution'—integrating diverse research traditions to avoid reductive technical or purely cultural analyses.

Experimental results

Research questions

• RQ1 How do the commercial and institutional structures surrounding AI image model development shape their design and deployment?
• RQ2 What biases and representational patterns are embedded in the training data of generative image models, and how do they reflect historical power imbalances?
• RQ3 How do prompting strategies reveal the model’s underlying logics, aesthetic preferences, and stereotypical tendencies in image generation?
• RQ4 In what ways do generative outputs reproduce or challenge existing gendered, racialized, and sexualized stereotypes in visual culture?
• RQ5 How can researchers develop a critical, multidimensional understanding of AI image models that accounts for technical, cultural, and political dimensions simultaneously?

Key findings

• Stable Diffusion generates predominantly white, youthful, and classically beautiful female nurses when prompted with 'nurse', revealing strong gendered and racialized stereotypes in its outputs.
• The inclusion of 'hyperrealistic' and '4k' in prompts produces a hybrid aesthetic that blends photography, digital art, and 3D rendering, indicating model-specific stylistic tendencies.
• Prompts for 'threatening face' or 'friendly face' consistently generated images reflecting existing societal biases, including racial and gendered caricatures.
• Non-cisgender identities were systematically misrepresented in generated images as less human, more sexualized, and more stereotyped, indicating deep-seated representational inequities.
• Iterative prompting and reflection revealed that models reproduce and amplify existing cultural norms rather than generating neutral or objective representations.
• The toolkit successfully exposes how AI image models are not neutral tools but reflect and reinforce structural inequalities embedded in their training data and development ecosystems.