> For example gender targeting at 42%? That means they somehow did 8% worse than random chance. This is a lot more plausible than you might think. Simply build

> For example gender targeting at 42%? That means they somehow did 8% worse than random chance. This is a lot more plausible than you might think. Simply build a machine-learning model that is accurate 58% of the time (maybe because it takes the "gender" field people manually entered on their online profiles and adds noise) and accidentally invert its output.