AI models trained on human data often reflect and exaggerate existing gender biases. Researchers quantify these disparities to address them, noting that binary categories are commonly used for measurement ease.
Word embeddings trained on data like Google News articles exhibit sexist analogies, such as mapping men to computer programmers and women to homemakers. This bias stems directly from the training text.
Facial recognition systems show significant intersectional accuracy disparities. Commercial classifiers achieved error rates up to 34.7% for darker-skinned females, compared to a maximum of 0.8% for lighter-skinned males.
