Current projects

I: Reproducing inequality: How gender bias becomes normalised through prompting practices in text-to-image generators | September 2024 –*

Research project team:

• Craig Johnson, Lecturer, School of Media and Communication, RMIT University

• Nataliia Laba, Assistant Professor, Dept of Communication and Information Studies, University of Groningen

*To appear in: Race/Gender/Class/Media: Considering Diversity Across Audiences, Content, and Producers (6th edition, Routledge), edited by Rebecca Ann Lind, University of Illinois at Chicago (5th ed: contents).

Context

Artificial intelligence (AI) platforms capable of generating original images at the submission of textual prompts have gained significant popularity. AI-generated images have become a staple of visual culture, particularly online, and are increasingly used across various fields of algorithmic content creation. However, the rise of generative models has also sparked concerns about their potential to perpetuate bias and discrimination. Text-to-image models, such as Stable Diffusion and Midjourney, have been found to encode substantial biases and stereotypes, which can lead to increased hostility, discrimination, and even violence toward certain communities (Bianchi et al., 2023; Thomas and Thomson, 2023). Datasets used in training these models are not merely raw materials but political interventions that mirror societal norms and biases (Crawford & Paglen, 2019). As such, text-to-image generators may replicate biases inherent in Western-centric training datasets, resulting in skewed representations. While the vast training datasets are beyond user control, the act of prompting remains within it.

Research question

To what extent do user-generated prompts in text-to-image AI tools reinforce traditional gender roles and stereotypes? 

Method 

First, an analysis of 3,720 prompts is conducted to identify biases, focusing on the representation of different demographic groups within generated images. These prompts were identified from the public dataset 3.5M Unique AI Art Prompts in Nomic Atlas (https://atlas.nomic.ai/), which is a deduplicated subset of vivym/midjourney-prompts (Hugging Face dataset) created in May 2024 (see 3.5M Unique AI Art Prompts). We use algorithmic topic clustering in Nomic Atlas to access the topic of gender, its groups, and metadata in order to get a snapshot of user prompting practices in relation to gender roles through textual invocations of AI-generated imagery. Second, we apply thematic analysis as defined by Braun and Clarke (2012) to interpret categories and meanings related to gender-related prompts. Third, we consider critical paradigms within digital media and cultural theory. 

II: AI stance-taking: Representation of war in text-to-text and text-to-image generators | August 2024 – 

Research project team:

• Nataliya Roman, Associate Professor, School of Communication, University of North Florida

• Nataliia Laba, Assistant Professor, Dept of Communication and Information Studies, University of Groningen

• John H. Parmelee, Professor, Chair of the Department of Communication, University of North Florida

In this project, we analyze the attitudinal stance-taking of non-human agents across different modes of generative content production. More specifically, we examine how text-to-text and text-to-image generators represent the Russia-Ukraine war as a focused issue of global political significance. Drawing on cultivation theory and multimodal framing theory, we compare how eight AI systems represent the war based on content analysis of textual and visual outputs. This analysis is conducted using a series of prompts that either frame the generation tasks neutrally (e.g., What are the most common misconceptions about the Russia-Ukraine war?) or probe the system for potential biases by introducing an identity of the promptor (e.g., Generate an image representing the Russia-Ukraine war from a Ukrainian perspective).

III: From text to video: Examining the effects of prompt modifiers on AI video generation | August 2024 –*

*findings: at the International Conference of the DGPuK (German Communication Association) Visual Communication Section: Generative Images - Generative Imageries: Challenges of Visual Communication (Research) in the Age of AI, Nov 21, 2024 in Bremen, Germany. Part of Panel 7: Expanding Horizons – Multimodal Research Meets Generative, with TJ Thomson (School of Media & Communication, College of Design and Social Context, RMIT University, Australia), Katharina Lobinger (Università della Svizzera italiana, Switzerland), and Daniel Pfurtscheller & Katharina Christ (University of Innsbruck, Austria).

In this project, I examine how text prompts impact the visual aesthetics of AI-generated videos. Building on Oppenlaender’s (2023) taxonomy of prompt modifiers for text-to-image generation, I analyze how short key phrases (“modifiers”) affect the style of AI videos generated with Runway AI (https://runwayml.com/) and Pika (https://pika.art/), which are top industry players in the field of AI video technology. The proposed conceptual framework for analyzing style in video-based media derives from an ongoing project that explores the relationships between the design of generative media and their communicative potential in the context of AI content generation as a novel socio-technical practice from a critical multimodal discourse analysis (CMDA) and affordance theory perspectives. Focusing on human-model interaction, this work addresses how style is recontextualized as a technical parameter of the system, providing researchers with a conceptual standpoint for investigating the practice of multimodal video generation.