On top of that they say that these sorts of behaviors only arise when the models are “stressed”, and the article also mentions “threats” like being unplugged. What kind of response do they actually expect from a fill-in-the-conversation machine when the prompt it’s been asked to continue from is a threat?