Recent studies have shown that Text-to-Image (T2I) model generations can reflect social stereotypes present in the real world. However, existing approaches for evaluating stereotypes have a noticeable lack of coverage of global identity groups and their associated stereotypes. To address this gap, we introduce the ViSAGe (Visual Stereotypes Around the Globe) dataset to enable the evaluation of known nationality-based stereotypes in T2I models, across 135 nationalities. We enrich an existing textual stereotype resource by distinguishing between stereotypical associations that are more likely to have visual depictions, such as ‘sombrero’, from those that are less visually concrete, such as ‘attractive’. We demonstrate ViSAGe’s utility through a multi-faceted evaluation of T2I generations. First, we show that stereotypical attributes in ViSAGe are thrice as likely to be present in generated images of corresponding identities as compared to other attributes, and that the offensiveness of these depictions is especially higher for identities from Africa, South America, and South East Asia. Second, we assess the ‘stereotypical pull’ of visual depictions of identity groups, which reveals how the ‘default’ representations of all identity groups in ViSAGe have a pull towards stereotypical depictions, and that this pull is even more prominent for identity groups from the Global South. CONTENT WARNING: Some examples contain offensive stereotypes.

Training and evaluation of language models are increasingly relying on semi-structured data that is annotated by humans, along with techniques such as RLHF growing in usage across the board. As a result, both the data and the human perspectives involved in this process play a key role in what is taken as ground truth by our models. As annotation tasks are becoming increasingly more subjective and culturally complex, it is unclear how much of their socio-cultural identity annotators use to respond to tasks. We also currently do not have ways to integrate rich and diverse community perspectives into our language technologies. Accounting for such cross-cultural differences in interacting with technology is an increasingly crucial step for evaluating AI harms holistically. Without this, the state of the art of the AI models being deployed is at risk of causing unprecedented biases at a global scale. In this tutorial, we will take an interactive approach by utilizing some different types of annotation tasks to investigate together how our different socio-cultural perspectives and lived experiences influence what we consider as appropriate representations of global concepts.