Machine learning personalisation methods have the ability to bias the content of a news feed, and reinforce filter bubbles. Can they also be used to do the opposite?