A deep learning model they built to analyze the queries suggested that the warnings should have appeared between 29 and 58 times as often as they did in practice. For motivated conspiracy peddlers, they were easy to evade: just adding quotation marks or a single letter to a query was typically enough to make the banner disappear.