Knowing Ourselves Through Others: Reflecting with AI in Digital Human Debates

TL;DR

This study introduces Digital Human Debates (DHD) as a framework for exploring how observing debates between self-designed, self-projected AI agents can foster metacognition and a new form of AI literacy. Using a Research Through Design approach, nine students design digital humans with prompt engineering and Retrieval Augmented Generation (RAG) and observe their autonomous debates to understand how designers reflect on their own cognition and values. The authors identify Reflecting with AI as a novel AI-literacy competency that leverages AI as a mirror while maintaining a boundary between self and other, and they conceptualize AI Ludens as a paradigm where AI-driven cultural play occurs through observation. The work demonstrates broad engagement with generative AI literacy competencies, highlights the importance of boundary management, and offers a practical pathway for integrating reflective AI practices into education and AI-literacy curricula.

Abstract

LLMs can act as an impartial other, drawing on vast knowledge, or as personalized self-reflecting user prompts. These personalized LLMs, or Digital Humans, occupy an intermediate position between self and other. This research explores the dynamic of self and other mediated by these Digital Humans. Using a Research Through Design approach, nine junior and senior high school students, working in teams, designed Digital Humans and had them debate. Each team built a unique Digital Human using prompt engineering and RAG, then observed their autonomous debates. Findings from generative AI literacy tests, interviews, and log analysis revealed that participants deepened their understanding of AI's capabilities. Furthermore, experiencing their own creations as others prompted a reflective attitude, enabling them to objectively view their own cognition and values. We propose "Reflecting with AI" - using AI to re-examine the self - as a new generative AI literacy, complementing the conventional understanding, applying, criticism and ethics.

Figures (7)

• Figure 1: Vocal grooming is the origin of linguistic communication, which builds social relationships within groups. As LLMs can communicate with humans, we can expect them to build new relationships between humans and LLMs and create new social groups.
• Figure 2: Positioning of this research in conversation perspectives. First-Person Conversation: Direct experiential engagement in communication with others. Third-Person Conversation: Observing conversations from an external standpoint. This Study (Digital Human-Mediated Conversation): Designing digital humans that embody one's thinking patterns, then observing their dialogues from a third-person perspective—creating a unique hybrid of self-projection and external observation.
• Figure 3: Dialogue system evaluation scores across three experimental conditions. Participants rated various aspects of dialogue quality on a 5-point Likert scale (1=strongly disagree to 5=strongly agree) after interacting with (A) human partners, (B) partner's digital human, and (C) their own digital human. Error bars represent standard error. Asterisks indicate significant differences from neutral (=3) using one-sample tests (*p<.05, **p<.01). While partner-DHs received significantly positive evaluations for accurately reflecting personal characteristics, self-DHs showed no significant difference from neutral, suggesting a complex perception of self-projected digital entities.
• Figure 4: System architecture diagram. Blue areas indicate components customizable by students. The system integrates Google Cloud Text-to-Speech API, React Speech Recognition, and Wav2Lip hosted on Replicate API for lip-sync video generation. LangChain Framework orchestrates RAG and LLM operations.
• Figure 5: Contest flow diagram. One-on-one debates between digital humans proceeded with topics and positions determined by roulette immediately before each match. Debates alternated between affirmative and negative speakers. After closing arguments, three judges raised placards to determine winners.