I work in the linguistics department at Stellar Dynamics. My job is to analyze conversations between our personal AI assistant MIRA and our clients, and flag any inappropriate responses from the AI. The end goal is to create better training data so we can eliminate those behaviors. As you can imagine I read a lot of strange, sometimes even disturbing, things in my work; most of the time it’s coming directly from the client, or is a response to a line of reasoning that the client has initiated. A couple of days ago I came across an extremely disturbing piece of conversation and this time it wasn’t the client prompting the weird responses. I have transcribed it in full below.