PLACES dataset exposes T2I model failures in Global South cultures
26,000 red-teaming examples reveal Western bias in image generation safety.
Deep Dive
Researchers created PLACES, a dataset of over 26,000 text-to-image model failures collected via localized community-centered red teaming in partnership with universities in Ghana, Nigeria, and two regions of India. They found unique adversarial patterns enabled by local cultural and linguistic nuances, and structural contextual gaps in existing safety frameworks showing normative dissonance. The work argues that expanding T2I safety requires localized, participatory methodologies.
Key Points
- Dataset includes 26,000 T2I failures from Ghana, Nigeria, Karnataka, and Punjab.
- Novel harms uncovered: violating religious norms, ignoring local customs, ominous symbolism.
- Existing safety frameworks are calibrated to Western norms, causing systematic blind spots in the Global South.
Why It Matters
Forces AI developers to adopt localized safety testing, preventing cultural harm in global deployments.