New data set helps identify culturally specific biases in LLMs

 Global – A new dataset called SHADES is helping researchers uncover how AI models internalise and reproduce harmful stereotypes in multiple languages. Built by an international team led by Margaret Mitchell, chief ethics scientist at Hugging Face, the tool moves beyond English-only bias detection by mapping responses across 16 languages and 37 regions.

Developed with input from native speakers, SHADES includes 304 culturally specific stereotypes – from gender tropes to racial prejudice. When prompted with these, AI models often generated pseudoscientific explanations, fabricated facts or reinforced the bias (especially in essay-style outputs).

‘These stereotypes are being justified as if they’re scientifically or historically true,’ says Mitchell. ‘That runs the risk of reifying deeply harmful views.’

Presented at the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, the dataset is publicly available and designed to evolve with ongoing contributions from around the world.

Explore how human and machine intelligence are merging in our Synthocene Era report.