This dissertation investigates the potential promises and perils of AI-mediated communication (AI-MC), a subset of computer-mediated communication (CMC) where communication is augmented or generated by AI. While users are accustomed to some existing implementations of AI-MC, such as spell check and grammar correction, new systems display a much higher amount of intervention, such as smart replies in messaging and email. Smart replies offer algorithmically-generated suggested responses based on the conversation content. Despite the fact that smart replies are directly aimed at shaping the production of messages, we do not know how they are influencing conversational and interpersonal dynamics. To avoid unexpected social consequences, this dissertation focuses on examining the effects that smart replies have on human interactions. I argue that while smart replies increase efficiency by allowing users to respond to messages more quickly, they also alter language and impact receivers’ interpersonal perceptions. Our work shows that smart replies are different from conversation content across various linguistic measures of sentiment. Specifically, commercially-available smart replies offer a disproportionate amount of positive sentiment compared to everyday conversation, and, as a result, our communication contains more positive sentiment when mediated by AI than it would have without smart replies. We also find that the presence of smart replies serves to increase trust between communicators and that, when interactions are unsuccessful, the AI can act like moral crumple zone by taking on responsibility that would otherwise have been assigned to the other human communicator. Similarly, we find that while actual smart reply use leads to improved interpersonal perceptions, the idea of sending these impersonal smart reply messages in everyday conversation is perceived negatively. In other words, even though using smart replies makes language more positive and can improve interpersonal perceptions, the idea among users that smart replies are inherently negative has persisted throughout the studies presented in this dissertation. Overall, I argue that while AI-MC in the form of smart replies can save time, improve communication efficiency, and better interpersonal perceptions, users should be cautioned that these benefits are coupled with altered language and a potential loss of personal expression. Additionally, it seems that user perceptions about smart reply use are quite negative, although this does not match the reality of using smart replies. Taking these promises and perils into account, our work suggests that a more promising role for smart replies could be as a mediator that works to recognize when a conversation is going awry and affect the necessary reparative actions between communicators.