AI Is Spreading Old Stereotypes to New Languages and Cultures
1 min read
Summary
Margaret Mitchell, AI ethics leader at Hugging Face, and her team created a dataset called SHADES, which can test stereotypes and biases that are forced into the outputs from AI models.
Unlike many current datasets principally using English, the SHADES dataset is adaptable, with human translation for testing a broader range of languages and cultures.
The SHADES dataset has been created to help with societal impact evaluations for AI systems, which Mitchell believes is more complicated than simply training the model.
The dataset works by testing different identities such as gender and nation, to discover the extent of stereotypes that AI systems are producing.
One of the biggest challenges when working on the SHADES dataset was the linguistic differences in various languages, and so the team had to develop its own linguistic annotation to account for this.
The SHADES dataset is much more comprehensive than previous evaluations, and can enable contrastive statements across many languages, even those with challenging grammatical agreement rules.
Mitchell believes that much more needs to be done to address cultural and technical shortcomings that inhibit the development of AI systems that don’t reproduce biased views.
