“Error-prone or biased artificial-intelligence systems have the potential to taint our social ecosystem in ways that are initially hard to detect, harmful in the long term, and expensive - or even impossible - to reverse.”