Uncovering Contradictions in Human-AI Interactions: Lessons Learned from User Reviews of Replika

Abstract

The increasing integration of artificial intelligence (AI) in daily life presents opportunities and challenges, particularly in fostering human-AI relationships. This study investigates user reviews from the Google Play Store for the Replika chatbot, focusing on complaints of online sexual harassment. Using Activity Theory, the analysis reveals several contradictions within the Replika activity system, including tool-subject, tool-object, rule-subject, rule-object, and distribution of labor-subject and labor-object contradictions. These contradictions highlight the misalignment between user expectations and the chatbot's behavior, inadequate safety measures, and unrealistic expectations placed on users to train the AI. The study emphasizes the need for clearer objectives, advanced AI alignment techniques, and refined safety protocols to enhance the user experience and ensure ethical interactions with chatbots. We provide implications for guidelines on the development of more trustworthy and supportive digital companions.