How good are humans at detecting AI-generated images? Learnings from an experiment