AI Is Spreading Old Stereotypes to New Languages and Cultures

image via WIRED

April 23, 2025, 4:31 PM

• •Margaret Mitchell, an AI ethics leader at Hugging Face, spearheaded the SHADES dataset to test generative AI models for perpetuating stereotypes across multiple languages and cultures.
• •The SHADES dataset employs human translations to address the limitations of English-centric bias mitigation efforts.
• •Mitchell highlights how AI models risk propagating harmful stereotypes globally, especially when bias mitigation is limited to English.
• •Challenges in creating SHADES included navigating linguistic differences and developing a novel, template-based approach for bias evaluation.

Margaret Mitchell, now with Hugging Face, developed the SHADES dataset to assess how generative AI models propagate stereotypes across various languages and cultures, moving beyond English-focused bias mitigation. She emphasizes the risk of AI amplifying harmful stereotypes worldwide due to insufficient attention to multilingual biases. Mitchell notes the challenges of linguistic diversity in creating this dataset and the need for techniques to address deep-seated biases. The SHADES dataset provides a way to better evaluate and analyze AI biases by offering human translations.

Read original article