Introduction
Modern AI systems are remarkably proficient at specific tasks, from image recognition to complex strategy games. However, they often lack a fundamental component of human intelligence: a sense of self. Most agents operate without an understanding of their own identity, purpose, or social role. This paper proposes a computational framework for "social self-awareness," enabling agents to maintain a mental model of themselves as both a subjective "I" and an objective "me" in the eyes of others.
The Problem: AI Without Meaning
The authors argue that for agents to work safely and effectively alongside humans, they must comprehend that they are distinct individuals. Without a model of "self," AI cannot truly grasp social nuances, trust, or "joint attention"—the state where two parties focus on the same object and are aware of the other's focus. For instance, to follow Asimov’s "Three Laws of Robotics," a robot must first understand that the word "robot" refers to itself.
A Framework for Self-Awareness
The proposed model achieves self-awareness by having the agent maintain a representation of itself in its own mind, treated similarly to other objects or people.
The Awareness Structure (AS)
At the core of this framework is the Awareness Structure, a graph-based representation in the agent's working memory. It consists of "Attention Units" representing objects, other agents, and the agent itself, linked by attributes and relations such as "perceive," "believe," or "distance."
Levels of Awareness
The framework distinguishes between different levels of self-representation:
- Subjective Self-Awareness: The agent as the subject of an experience (e.g., "I see the box").
- Objective Self-Awareness: The agent representing itself as an object in someone else's mind (e.g., "The user believes I am helpful").
System Architecture
The framework integrates three key components:
- Perceptual System: Identifies and tracks external objects and agents.
- Executive Process: Manages and updates the Awareness Structure in real-time.
- Autobiographical Memory: Stores past experiences to provide a continuous sense of identity over time.
Practical Application: The Eunice Chatbot
The framework was evaluated through a functional implementation of "Eunice," a conversational agent designed for Individual Cognitive Stimulation Therapy (ICST) for elderly patients with dementia.
In trial dialogues, Eunice demonstrated the ability to navigate social friction. By maintaining an "objective" view of itself, the agent could detect when a user was becoming frustrated or when "joint attention" was failing. For example, if a user reacted negatively to a topic, Eunice could pinpoint the break in the social interaction and proactively pivot to a more positive subject.
Furthermore, the agent used "strategic self-disclosure"—sharing its own simulated difficulties (e.g., "I also can't remember so well")—to build rapport and reduce user frustration, mimicking human-like empathetic support.
Key Insights and Future Work
Self-awareness acts as the "glue" connecting an agent's internal state with the external social world. By focusing on identity-centric modeling, the agent tracks who it is across different contexts, allowing for more resilient and meaningful interactions.
While the current results are qualitative and based on specific case studies, they highlight the potential for self-aware AI to provide better social support. Future work will focus on large-scale clinical studies, more robust human-like inference mechanisms, and exploring how this model can simulate the development of complex social personalities.
Summary updated