AI companion platforms such as Character.AI (CAI) allow people to create and interact with AI-generated characters for entertainment, role play, guidance or companionship. With 4+ million of user-created characters, ranging from fictional figures to representations of real people, CAI provides an opportunity to study the platform comprehensively. Previous work has analysed discourse from users in order to build a taxonomy of character archetypes [1]. In this project, you will scrape and analyse a large dataset of CAI characters to develop a taxonomy along the dimensions of:
- Who or what the characters represent (e.g., fictional characters, celebrities, original characters, relationship roles)
- Character attributes (gender, age, personality traits, occupations, relationships)
- Themes and scenarios (e.g., romance, friendship, fantasy, violence, family relationships)
- Popularity
- Potentially harmful representations, including stereotypes, sexual objectification, discrimination, violence and the normalization of harmful or illegal behaviours
- Relationships between characteristics and popularity, i.e. whether particular themes, identities, or scenarios are associated with greater user engagement.
A key resource for the project is CAIBotList (or the pro version), a search engine that indexes CAI characters. It provides information such as:
- the character’s name, title, description and greeting
- its creator
- date of creation
- its category (e.g., anime, original character, video games)
- number of interactions and likes
- automatically assigned tags, such as ‘male’, ‘arrogant’, ‘mafia’
This gives us the possibility of studying the platform at a scale that would not be possible by manually examining characters one by one.
Disclaimer: Some of the content may be disturbing.