Stable Bias: Evaluating Societal Representations in Diffusion Models

By Alexandra Sasha Luccioni, Christopher Akiki, Margaret Mitchell, Yacine Jernite. Neural Information Processing Systems, 2023

345 citations recorded by the source corpus.

As machine learning-enabled Text-to-Image (TTI) systems are becoming increasingly prevalent and seeing growing adoption as commercial services, characterizing the social biases they exhibit is a necessary first step to lowering their risk of discriminatory outcomes. This evaluation, however, is made more difficult by the synthetic nature of these systems' outputs: common definitions of diversity are grounded in social categories of people living in the world, whereas the artificial depictions of fictive humans created by these systems have no inherent gender or ethnicity. To address this need, we propose a new method for exploring the social biases in TTI systems. Our approach relies on characterizing the variation in generated images triggered by enumerating gender and ethnicity markers in the prompts, and comparing it to the variation engendered by spanning different professions. This allows us to (1) identify specific bias trends, (2) provide targeted scores to directly compare models in terms of diversity and representation, and (3) jointly model interdependent social variables to support a multidimensional analysis. We leverage this method to analyze images generated by 3 popular TTI systems (Dall-E 2, Stable Diffusion v 1.4 and 2) and find that while all of their outputs show correlations with US labor demographics, they also consistently under-represent marginalized identities to different extents. We also release the datasets and low-code interactive bias exploration platforms developed for this work, as well as the necessary tools to similarly evaluate additional TTI systems.

Subfields: Sociology and Political Science, Statistical and Nonlinear Physics, Statistics and Probability

View Paper: https://www.semanticscholar.org/paper/9116c50561007cd69cc24a60c7593a5425f57fcc

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Sociology and Political Science

In Statistical and Nonlinear Physics

In Statistics and Probability

Citations in the corpus

Art and the science of generative AI (2023, methodology): “Without documentation of what data is used in the training of generative-AI models, it is more difficult to identify, quantify and mitigate the biases the models have [46, 47], although efforts have been made to overcome this issue [48, 49].”
Design Principles for Generative AI Applications (2024, background): “…property [47, 107], the circumvention or reverse-engineering of prompts through attacks [35], the production of hateful, toxic, or profane language [60], the disclosure of sensitive or personal information [83], the production of malicious source code [26, 28], and a lack of representation of…”
Unified Concept Editing in Diffusion Models (2023, background): “Diffusion image generation models have been found to be prone to social and cultural biases [7,24,43], even exaggerating and amplifying societal stereotypes beyond simple imbalances in the training data [4,11], although quantifying ampliﬁcation can be subtle [36].”