Image Cropping on Twitter: Fairness Metrics, their Limitations, and the Importance of Representation, Design, and Agency

Twitter uses machine learning to crop images, where crops are centered around the part predicted be most salient. In fall 2020, users raised concerns that automated image cropping system on favored light-skinned over dark-skinned individuals, as well woman's bodies instead of their heads. order address these concerns, we conduct an extensive analysis using formalized group fairness metrics. We find systematic disparities in and identify contributing factors, including fact based single salient point can amplify because effect term argmax bias. However, demonstrate metrics quantitative own insufficient for capturing risk representational harm automatic cropping. suggest removal saliency-based favor a solution better preserves user agency. For developing new sufficiently related harm, our critique motivates combination qualitative methods include human-centered design.