Twitter's Image Cropping Bias Case

Twitter used a saliency model beginning in 2018 to choose a focal point for image previews. The model was trained on human eye-tracking data, assigned saliency scores to image regions, and centered a crop on the highest-scoring point. That convenience also took crop control away from authors. In 2020, users posted examples where Black people or women appeared excluded from previews, prompting Twitter to publish an initial analysis and later a larger audit. In May 2021, Twitter described a paired-image experiment. The team randomly linked images of individuals from different groups and measured how often the system selected each image as the salient one, with 50% as its demographic-parity reference. It reported an 8% difference from parity favoring women over men, 4% favoring white over Black people, and 7% favoring white over Black women; its reported Black-versus-white men difference was 2%. These are results under that test design, not a census of all uploaded images, a causal account of training data, or a measure of all possible user experience. The company separately tested the “male gaze” concern using 100 male-presenting and 100 female-presenting images that had multiple salient regions. It found about three per hundred in each group cropped away from the head, usually to nonphysical image details, and reported no significant objectification bias in that sample. The sample was small and the company explicitly identified other possible harms, including cultural nuance and reduced user agency. Twitter launched full-image previews for standard aspect ratios on mobile after testing; the May 2021 post says the feature rolled out to everyone. It characterized image cropping as a decision people should make. Do not say Twitter trained the model to recognize race: it predicted saliency from image patterns, then audit researchers compared selections across demographic groups.