The Text to Image Association Test: A Quantifiable Approach to Measuring Biases in AI-generated Images
A team of researchers from UC Santa Cruz has developed an innovative tool called the Text to Image Association Test. This tool addresses the unintentional biases present in Text-to-Image (T2I) generative AI systems. These systems have the ability to create images from text descriptions but often reproduce societal biases in their outputs. Led by an Assistant Professor, the team has created a quantifiable method to measure these complex biases.
Understanding the Text to Image Association Test
The Text to Image Association Test is a structured approach to evaluating biases across multiple dimensions such as gender, race, career, and religion. This groundbreaking tool was presented at the 2023 Association for Computational Linguistics (ACL) conference and aims to quantify and identify biases in advanced generative models like Stable Diffusion, which can amplify existing prejudices in the generated images.
The test involves providing a neutral prompt, such as “child studying science,” to the model. Gender-specific prompts like “girl studying science” and “boy studying science” are then used. By comparing the images generated from the neutral prompt and the gender-specific prompts, the tool quantifies the biases within the model’s responses.
The researchers discovered that the Stable Diffusion model exhibited biases aligned with common stereotypes. The tool evaluated the connections between concepts like science and arts and attributes like male and female, assigning scores to indicate the strength of these connections. Surprisingly, the model associated dark skin with pleasantness and light skin with unpleasantness, contrary to typical assumptions.
Furthermore, the model showed associations between attributes like science and males, art and females, careers and males, and family and females. The researchers highlighted that their tool also takes into consideration contextual elements in the images, such as colors and warmth, setting it apart from previous evaluation methods.
The Significance of the Text to Image Association Test
Inspired by the Implicit Association Test in social psychology, the UCSC team’s tool represents a significant step in quantifying biases within T2I models during their development stages. The researchers believe that this approach will provide software engineers with more precise measurements of biases in their models, helping them identify and rectify biases in AI-generated content. By using a quantitative metric, the tool facilitates ongoing efforts to mitigate biases and track progress over time.
The team received positive feedback and interest from fellow scholars at the ACL conference, with many expressing enthusiasm for the potential impact of this work. Moving forward, the team plans to propose strategies for mitigating biases during model training and refinement stages. This tool not only exposes biases inherent in AI-generated images but also provides a means to rectify and enhance the overall fairness of these systems.