This tool is built from processed impression ratings in the OMI (One Million Impressions) dataset, where people judged faces.

The outputs DO reflect human stereotypes and social biases.