Measuring bias in multimodal image–text models

Contact person: Erik Velldal

Status: Project

Multimodal text–image models have rapidly advanced in their ability to generate rich textual descriptions of visual content or, conversely, to generate images from a given textual description. However, neither their textual nor visual depictions are necessarily neutral. For example, when a model maps an image into language, this “translation” between modalities may also introduce bias through particular framing, emphasis, omission, or evaluative language. This project will develop methods and datasets for analyzing ideological bias in multimodal text–image models, addressing methodological challenges in multimodal benchmarking as well as broader questions about bias in how AI systems represent social and political realities.