Text-to-Image (T2I) models such as DALL-E and Midjourney are gaining popularity across various global communities, where they are used to generate visual representations of distinct cultural identities. Evidence suggests that, when prompted without a specific cultural identity, the outputs of T2I models frequently reflect Western cultural norms and perspectives. Moreover, while recent research has focused on the biases and stereotypes these models can reinforce (e.g., misrepresentation, erasure, stereotyping), there has been less attention on how well these models capture the full diversity and complexity of different cultures. In this project, we are interested in examining missing perspectives regarding how culture is represented and evaluated in AI-generated images.
We're looking for participants to join our workshops in March and April!
Register your interest here: CulturAI workshop study
Photo by AbsolutVision on Unsplash
Evaluation of models will be done based on a participatory definition of culture. To be informed by our stakeholder engagement activities.
Photo by Towfiqu barbhuiya on Unsplash
We are a multidisciplinary group of researchers working together as part of the UKRI Responsible Ai UK research programme.
Gisela Reyes Cruz
University of Nottingham
Sarah Kiden
University of Southampton
Maira Klyshbekova
King's College London
Damian Eke
University of Nottingham
Aislinn Bergin
University of Nottingham
Tayyaba Azim
University of Southampton
Oriane Peter
King's College London
Maria Waheed
University of Nottingham
Sena Choi
University of Nottingham
Contact gisela.reyescruz@nottingham.ac.uk to get more information about the project