Linguistic Markers of Emotional Attachment in AI Companion Chatbot Interactions: A Mixed-Methods Analysis of Dependency Patterns in Negative Emotion Conversations

Between 2022 and mid-2025, AI companion apps surged by 700%. As artificial intelligence increasingly mediates communication, therapy and companion chatbots have become leading uses of generative AI. However, these technologies are released globally without regulatory oversight or empirical research on mental health outcomes. This study investigates how AI companion chatbots influence emotional attachment patterns among users, focusing on linguistic markers of dependency in negative emotion conversations. Using mixed methods, we analyzed 6,030 English conversations from an AI chatbot dataset. VADER sentiment analysis classified 1,243 (20.6%) as negative emotion conversations. An emotional attachment framework across five categories (loneliness, need_support, emotional_bond, mental_health, dependency) quantified attachment language through keyword and phrase matching. Among negative conversations, 4.1% (n = 51) exhibited high attachment scores. Users wrote shorter messages during negative states (mean = 76.72 words) versus positive interactions (mean = 143.13 words), while AI responses remained consistently lengthy (mean = 187.09 words). Linguistic analysis through word frequency, bigram/trigram extraction, and word cloud visualization revealed systematic dependency patterns. A stratified subsample of 100 conversations underwent detailed severity classification. Analysis of high-attachment conversations revealed limited AI boundary-setting or professional referrals. A small but significant subset of users displays linguistic markers of emotional dependency during negative emotional states. Findings underscore urgent needs for design safeguards, regulatory frameworks, and longitudinal research to balance emotional support with user protection.