Twitter’s Image Cropping Tool Allegedly Showed Gender and Racial Bias

Twitter deployed a saliency-based image cropping algorithm in 2018 to automatically crop images for consistent display in timelines. In September 2020, users discovered that the algorithm exhibited racial bias by consistently selecting lighter-skinned people over darker-skinned people when cropping photos containing multiple individuals. Twitter initially denied bias after internal testing but later conducted a comprehensive study of 10,000 image pairs that confirmed the bias. The research found a 4% difference from demographic parity favoring white individuals over Black individuals, with the largest gap being a 28% difference (64% vs 36%) when comparing Black men to white women. The algorithm also showed an 8% bias favoring women over men and a 7% bias favoring white women over Black women. Twitter researchers suggested the bias may result from training data containing images with dark backgrounds where lighter skin creates higher contrast, and from the algorithm being trained on eye-tracking data that may reflect existing human biases. Twitter responded by eliminating automatic cropping for single images on mobile apps in March 2021 and announced plans to remove it from the website, though it continues to be used for multi-image tweets and thumbnails.