Ghana, Nigeria, India Test Image Model Safety

Researchers partnered with universities in Ghana, Nigeria, and two regions of India to test the safety of text-to-image models. They focused on secondary urban centers within those countries, prioritizing localized perspectives often missing from current safety standards. The team ran workshops to understand local norms and collected over 26,000 examples of model failures.

Analysis shows the prompts differ greatly from existing, globally-sourced red-teaming data in terms of culture and language. They found unique ways models failed due to local customs, including violations of religious norms and use of ominous symbolism.

The study highlights gaps in current safety frameworks. Researchers note the need for more localized data collection and contextualization to improve text-to-image safety worldwide, but acknowledge the examples collected contain potentially harmful content. The work to broaden safety standards continues.