"Because our first studies involved human subjects, we couldn’t use hundreds of faces to show the effect. So we turned to machine learning, reasoning that if Charlotte looks like a Charlotte, even a computer should be able to recognize her as one. We taught a computer what a Charlotte looks like by presenting a few Charlottes and what a non-Charlotte looks like by presenting an Amélie, a Claire, and so on. Then we fed the computer nearly 100,000 faces that it had never processed and, for each one, supplied two names—the real name of the person shown and a second possibility. The computer chose the correct name 54% to 64% of the time, which is significantly higher than the 50% chance level," said one of the study authors.