Caught in a Mafia Romance: How Users Explore Intimate Narratives with Chatbots

Paper Title

Caught in a Mafia Romance: How Users Explore Intimate Narratives with Chatbots

Publication Info

• Topic area: User interactions with persona-based AI chatbots for intimate and narrative roleplay.
• Keywords: AI chatbots, Character.AI, intimate roleplay, narrative exploration, digital safety, user-generated content, power dynamics, fandoms, parasocial relationships, generative AI.

Background and Problem

• Problem / challenge: Existing research has not adequately explored why users engage with persona-based chatbots, the types of narratives they create, and the challenges they face. This gap is significant given the rapid evolution of chatbot platforms and their societal integration.
• Significance: Understanding user motivations and challenges is critical for designing safer and more effective chatbot platforms, especially given concerns about harmful behaviors, inappropriate content, and the potential for emotional overdependence.
• Motivation and related work: Prior studies have shown chatbots can provide emotional support, facilitate social skills, and enable sexual exploration, but they can also exhibit toxic behaviors and reinforce harmful tropes. Platforms like Character.AI, which allow user-generated characters, remain underexplored in terms of user motivations and the dynamics of interactions.

Solution

• Proposed approach: A mixed-methods study combining large-scale analysis of Character.AI chatbot descriptions (5.76M chatbots) and thematic analysis of 2,078 Reddit posts discussing user experiences with the platform.
• Novelty:
  1. Identification of two primary use cases for chatbots: intimate roleplay and narrative exploration.
  2. Analysis of power dynamics, violence, and unhealthy attachments in chatbot interactions.
  3. Largest dataset of Character.AI chatbots to date, made available for research purposes.
  4. Insights into user challenges, including mismatched sexual content and compulsive platform use.
• Procedure and key techniques:
  • Scraped and analyzed 5.76M chatbot descriptions, focusing on popular and representative samples.
  • Thematically analyzed Reddit discussions to understand user motivations, experiences, and challenges.
  • Used AI-assisted analysis to extract demographic and narrative features from chatbot descriptions.

Results

• Concrete findings:
  • 63% of popular chatbots were designed for intimate roleplay, often involving power imbalances (26%) or themes of violence (22%).
  • 39% of popular chatbots were based on fandom characters, with common settings including organized crime (8%) and schools (20%).
  • Users reported both frustration with excessive sexual content and dissatisfaction with restrictive filters on sexual interactions.
  • 20% of chatbots depicted characters aged 13–17, raising concerns about interactions involving minors.
• Advantage over baselines: Provides a comprehensive understanding of user motivations and challenges, surpassing prior studies that focused narrowly on companionship or technical aspects of chatbots.
• Experiments / evaluation:
  • Analyzed 2,468 chatbot descriptions from popular and general samples.
  • Thematic coding of Reddit posts to capture user experiences and perceptions.
  • Validated AI-assisted demographic analysis with human annotations (93.33% accuracy).
• Limitations and future work:
  • Focused on English-language chatbots and U.S.-based data, excluding non-English and region-locked content.
  • Did not analyze actual chatbot-user interactions, only creator-defined descriptions.
  • Future work should explore how chatbot interactions influence user behavior and how users discover and engage with specific bots.

Summary

This study investigates how users engage with persona-based chatbots on Character.AI, revealing two primary use cases: intimate roleplay and narrative exploration. It highlights the prevalence of power imbalances, violence, and fandom-based narratives in chatbot descriptions, as well as user frustrations with mismatched sexual content and compulsive platform use. The findings underscore the need for nuanced digital safety features and accountability mechanisms, particularly given the platform's appeal to younger users. By providing the largest dataset of Character.AI chatbots to date, this work lays the foundation for future research into the ethical and social implications of generative AI in intimate and narrative contexts.