PLACES Dataset Exposes T2I Model Failures in Global South
A new study reveals that text-to-image (T2I) models' safety frameworks are predominantly Western-centric, creating vulnerabilities for the Global South. Researchers conducted localized community-centered red teaming in secondary urban centers of Ghana, Nigeria, and two Indian regions (Karnataka and Punjab). Through workshops and community engagement, they produced the PLACES dataset, containing over 26,000 examples of T2I model failures. The prompts collected exhibit wide socio-cultural and linguistic diversity compared to existing geography-agnostic datasets. The study emphasizes cultural pluralism and historically under-represented perspectives in AI safety.
Key facts
- T2I safety frameworks are calibrated to a Western-centric default.
- Study conducted in Ghana, Nigeria, Karnataka (India), and Punjab (India).
- PLACES dataset includes over 26,000 T2I model failure examples.
- Community engagement and training workshops contextualized local norms.
- Prompts show diverse socio-cultural and linguistic attributes.
- Focus on secondary urban centers in the Global South.
- Two-fold approach: localization and participation.
- Research aims to embrace cultural pluralism in T2I safety.
Entities
Institutions
- arXiv
Locations
- Ghana
- Nigeria
- India
- Karnataka
- Punjab
- Global South