Incident: A crop that always chose one face
14 min read·
Module 13 · Incident file · About 35 minutes · Needs: Lesson 1 (ranking the spare vans, NDCG), Lesson 2 (domain-specific evaluation)
Everything below about the case comes from a paper written by the team that owned the system, quoted where it matters. It deals with how an image-cropping model treated people of different races and genders; the paper's own words are used for those findings.