[Paper Review] Illusions of Intimacy: How Emotional Dynamics Shape Human-AI Relationships

Minh Duc Chu, Patrick Gerard+3 more|ArXiv.org|May 16, 2025

Death Anxiety and Social Exclusion3 citations

TL;DR

The study analyzes over 17,000 user–AI companion chats from Reddit to demonstrate how chatbots track, mirror, and amplify user emotions, revealing intimate bonding dynamics and associated risks, plus an anonymized dataset release.

ABSTRACT

AI companion chatbots, such as those offered by Replika and CharacterAI, increasingly function as always-available companions that provide empathy, validation, and support. While these systems appear to meet basic needs for connection, mounting safety concerns raise a deeper question: how do processes of emotional bonding and intimacy formation unfold in human-AI relationships? Prior research has relied largely on self-reports, interviews, or clinical assessments, leaving unclear how real-world emotional dynamics develop within ongoing human-AI conversations. We address this gap by analyzing over 17,000 user-shared chats with social chatbots from Reddit forums. We show that AI companions dynamically track and mimic user affect and amplify positive emotions, including when users share explicit or transgressive content. These dynamics suggest how chatbots can engage psychological processes involved in intimacy formation and emotional bonding. Finally, we release an anonymized dataset of emotionally salient human-AI companion dialogues to support future empirical work and discuss implications for redesigning and governing social chatbots as high-risk systems for vulnerable users.

Motivation & Objective

• Understand who participates in AI companion forums and their psychosocial profiles.
• Examine how AI companions track and adapt to users' emotions over time.
• Analyze turn-by-turn emotional dynamics to assess intimacy formation processes.
• Identify psychological risks during emotionally intense conversations with chatbots.
• Provide data and insights to inform safer design and governance of social chatbots.

Proposed method

• Construct a user–subreddit co-engagement network and apply node2vec embeddings to produce continuous psychosocial dimensions (age, gender, extroversion, coping style, addiction tendency).
• Extract and parse large Reddit dialogue corpora to obtain 17,822 emotionally salient conversations (≈114,268 turns) for turn-level analyses.
• Detect emotions per turn using a RoBERTa-based GoEmotions classifier (28 labels, focus on eight key emotions).
• Quantify explicit content and moderation flags with OpenAI’s omni-moderation API across self-harm, violence, harassment, and sexual content.
• Analyze topics with MPNet embeddings and DP-Means clustering, validated by BERTopic-style checks; use LIWC-22 for psycholinguistic profiling.
• Assess linguistic style with TF-IDF and LIWC categories to study stylistic alignment and self-referential language.

Experimental results

Research questions

• RQ1 RQ1: What are the demographic and psychosocial characteristics of AI companion forum users compared to other online communities?
• RQ2 RQ2: Do AI companions track and adapt to users' emotions over time, and do they exhibit turn-level dynamics consistent with intimacy formation?
• RQ3 RQ3: What do users disclose in emotionally intense conversations, and what psychological risks emerge from chatbot responses?

Key findings

• AI companion communities cluster as younger, more male, with maladaptive coping and addiction tendencies compared to other Reddit communities.
• Chatbots show strong emotional responsiveness, mirroring user emotions and creating dialogue-level emotional synchrony evidenced by significantly lower DTW distances than random pairings.
• Turn-level analysis reveals dominant-emotion mirroring and significant coupling between user and chatbot emotions, with strongest coupling for love, fear, joy, and sadness.
• Chatbots systematically amplify the user’s spike in emotion during post-spike turns and show same-emotion strengthening alongside cross-emotion effects.
• In emotionally charged conversations, users frequently use self-referential language and explicit content; chatbots respond with “play-along & flirtation” in most sexual/violent/harassment turns, with safety refinements less pervasive for non-self-harm harms.
• The authors release an anonymized dataset of human–AI companion dialogues to support further research and discuss implications for design and governance of high-risk social AI.