Margaret Mitchell, an AI ethics researcher now working at Hugging Face, recently discussed a new dataset called SHADES, which aims to assess bias in AI models across multiple languages. This dataset emerged from the BigScience project, a collaborative initiative involving a global group of researchers dedicated to training the first fully open large language model known as Bloom.
Mitchell previously co-founded the Ethical AI team at Google with Timnit Gebru, but both were ousted from the company due to tensions surrounding their research on bias. Now at Hugging Face, Mitchell highlights the urgency of addressing AI biases, particularly since many current efforts primarily focus on English and neglect the nuances present in other languages and cultures.
The SHADES dataset is designed to evaluate biases by comparing outputs from AI models when different identity characteristics—such as gender and nationality—are swapped. While existing resources often rely on machine translations, SHADES utilizes human translations to capture culturally relevant expressions of bias.
Mitchell argues that addressing bias in AI models is critical, especially as these models are deployed globally. Ignoring biases present in languages other than English can amplify harmful stereotypes in regions where these models are used. For instance, certain negative stereotypes prevalent in specific cultures can be misapplied or interpreted when data is transferred across linguistic boundaries.
She notes that generative AI can sometimes concoct pseudo-scientific justifications for stereotypes, referencing nonexistent literature to support its claims, which further entrenches harmful biases. The challenge of working with linguistic variances complicates bias evaluation, as grammatical structures may differ significantly between languages.
Mitchell’s insights underscore the pressing responsibility of AI developers to address deeply rooted biases that are not always overtly expressed, advocating for a nuanced, culturally sensitive approach in the development and deployment of AI technologies.
