While researchers have long noted that people anthropomorphize computers, LLM-based chatbots represent a qualitative shift; they are the first technology that can engage in genuinely human-like conversations. In this talk, Jeremy Foote will discuss how the ongoing integration of LLMs into our social worlds is creating new risks and opportunities, using findings from recent work done by his group as examples. The first project argues that companion AI like Replika introduce new privacy risks, because in order to provide emotional support they rely on the sharing of sensitive information. Interviews with chatbot users find that they often employ heuristics based in interpersonal communication when deciding what to share, but also engage in fairly sophisticated strategies to safeguard private information from AI companies. The second project considers how the social affordances of chatbots can be used for good. This project invited users who had posted toxic content on Reddit to engage in a conversational intervention with a chatbot designed to elicit self-reflection. Results show that people are willing to engage with chatbots, often in good faith, but these conversations did not cause a measurable change in toxic behavior. The results suggest best practices for designing rehabilitative bots, as well as other opportunities for integrating helpful AI agents into other online social processes.