Siobhan Mackenzie Hall

Siobhan Mackenzie Hall, Fernanda Gonçalves Abrantes, Hanwen Zhu et al. · cmu, oxford

We introduce VisoGender, a novel dataset for benchmarking gender bias in vision-language models. We focus on occupation-related biases within a hegemonic system of binary gender, inspired by Winograd and Winogender schemas, where each image is associated with a caption containing a pronoun relationship of subjects and objects in the scene. VisoGender is balanced by gender representation in professional roles, supporting bias evaluation in two ways: i) resolution bias, where we evaluate the difference between pronoun resolution accuracies for image subjects with gender presentations perceived as masculine versus feminine by human annotators and ii) retrieval bias, where we compare ratios of professionals perceived to have masculine and feminine gender presentations retrieved for a gender-neutral search query. We benchmark several state-of-the-art vision-language models and find that they demonstrate bias in resolving binary gender in complex scenes. While the direction and magnitude of gender bias depends on the task and the model being evaluated, captioning models are generally less biased than Vision-Language Encoders. Dataset and code are available at https://github.com/oxai/visogender

Hugo Berg, Siobhan Mackenzie Hall, Yash Bhalgat et al. · oxford

Vision-language models can encode societal biases and stereotypes, but there are challenges to measuring and mitigating these multimodal harms due to lacking measurement robustness and feature degradation. To address these challenges, we investigate bias measures and apply ranking metrics for image-text representations. We then investigate debiasing methods and show that prepending learned embeddings to text queries that are jointly trained with adversarial debiasing and a contrastive loss reduces various bias measures with minimal degradation to the image-text representation.

Jabez Magomere, Shu Ishida, Tejumade Afonja et al.

We introduce the World Wide recipe, which sets forth a framework for culturally aware and participatory data collection, and the resultant regionally diverse World Wide Dishes evaluation dataset. We also analyse bias operationalisation to highlight how current systems underperform across several dimensions: (in-)accuracy, (mis-)representation, and cultural (in-)sensitivity, with evidence from qualitative community-based observations and quantitative automated tools. We find that these T2I models generally do not produce quality outputs of dishes specific to various regions. This is true even for the US, which is typically considered more well-resourced in training data -- although the generation of US dishes does outperform that of the investigated African countries. The models demonstrate the propensity to produce inaccurate and culturally misrepresentative, flattening, and insensitive outputs. These representational biases have the potential to further reinforce stereotypes and disproportionately contribute to erasure based on region. The dataset and code are available at https://github.com/oxai/world-wide-dishes.

Samantha Dalal, Siobhan Mackenzie Hall, Nari Johnson

The demands for accurate and representative generative AI systems means there is an increased demand on participatory evaluation structures. While these participatory structures are paramount to to ensure non-dominant values, knowledge and material culture are also reflected in AI models and the media they generate, we argue that dominant structures of community participation in AI development and evaluation are not explicit enough about the benefits and harms that members of socially marginalized groups may experience as a result of their participation. Without explicit interrogation of these benefits by AI developers, as a community we may remain blind to the immensity of systemic change that is needed as well. To support this provocation, we present a speculative case study, developed from our own collective experiences as AI researchers. We use this speculative context to itemize the barriers that need to be overcome in order for the proposed benefits to marginalized communities to be realized, and harms mitigated.

Elizaveta Semenova, Alisa Sheinkman, Timothy James Hitge et al.

Surrogate models are widely used to approximate complex systems across science and engineering to reduce computational costs. Despite their widespread adoption, the field lacks standardisation across key stages of the modelling pipeline, including data sampling, model selection, evaluation, and downstream analysis. This fragmentation limits reproducibility and cross-domain utility -- a challenge further exacerbated by the rapid proliferation of AI-driven surrogate models. We argue for the urgent need to establish a structured reporting standard, the Surrogate Model Reporting Standard (SMRS), that systematically captures essential design and evaluation choices while remaining agnostic to implementation specifics. By promoting a standardised yet flexible framework, we aim to improve the reliability of surrogate modelling, foster interdisciplinary knowledge transfer, and, as a result, accelerate scientific progress in the AI era.

Brandon Smith, Miguel Farinha, Siobhan Mackenzie Hall et al.

Vision-language models are growing in popularity and public visibility to generate, edit, and caption images at scale; but their outputs can perpetuate and amplify societal biases learned during pre-training on uncurated image-text pairs from the internet. Although debiasing methods have been proposed, we argue that these measurements of model bias lack validity due to dataset bias. We demonstrate there are spurious correlations in COCO Captions, the most commonly used dataset for evaluating bias, between background context and the gender of people in-situ. This is problematic because commonly-used bias metrics (such as Bias@K) rely on per-gender base rates. To address this issue, we propose a novel dataset debiasing pipeline to augment the COCO dataset with synthetic, gender-balanced contrast sets, where only the gender of the subject is edited and the background is fixed. However, existing image editing methods have limitations and sometimes produce low-quality images; so, we introduce a method to automatically filter the generated images based on their similarity to real images. Using our balanced synthetic contrast sets, we benchmark bias in multiple CLIP-based models, demonstrating how metrics are skewed by imbalance in the original COCO images. Our results indicate that the proposed approach improves the validity of the evaluation, ultimately contributing to more realistic understanding of bias in vision-language models.