Abstract Despite being a well-studied problem, bias continues to affect modern Computer Vision (CV) datasets, models and systems, including Generative AI models that have been found to replicate and amplify harmful biases and stereotypes. In this work, we focus on framing bias : a type of bias that arises from the way images convey meaning. We introduce a semiotic perspective that treats image datasets as texts and argues that framing bias is not a statistical accident due to sampling noise or imbalance, but an inherent property of image datasets as meaning-making systems. We show that co-occurrence of visual elements is the primary mechanism through which image datasets frame concepts, and that not only do individual images contribute to dataset framing through co-occurrence, but dataset-level framing also shapes the interpretation of individual images. As a consequence, every image dataset embodies a point of view and cannot be fully unbiased . We illustrate these claims through a case study of the Visual Genome dataset, revealing a framing of human activities that over-represents leisure, systematically excludes everyday labour, and is characterised by a predominantly American/Western viewpoint. Finally, we reflect on the epistemological implications of analysing image datasets, highlighting the mediated and interpretive nature of knowledge produced through AI systems, and propose revisionism and source criticism as epistemological paradigms to address these issues.
Fabbrizzi et al. (Wed,) studied this question.