Abstract
As autonomous pedagogical agents, such as social robots, become increasingly integrated into educational settings, the need for transparency in their decision-making processes grows. This paper explores how Explainable AI (XAI) can be leveraged to foster logical thinking and trust in children. By moving beyond "black box" interactions, the proposed framework utilizes strategic sub-optimality and comparative explanations to help children build shared mental models with their robotic peers.
Introduction
Explainable AI (XAI) has emerged as a critical field for making complex machine learning models interpretable to human experts and adult users. However, a significant gap exists regarding pedagogical agents designed for children. Children learn most effectively through social interaction and constructivist processes—reflecting on their actions to deepen their understanding. When an AI agent acts as a "black box," children are unable to reflect on the agent's logic or learn from its strategic choices. This research addresses the challenge of minimizing the gap between a child's expectations and an agent's actual behavior to improve both trust and educational outcomes.
The Challenge of Pedagogical Transparency
In collaborative learning environments, an agent's utility is not just measured by its performance, but by its ability to facilitate a teammate-like relationship. If a child does not understand why a robot teammate made a specific move in a game, they cannot effectively cooperate or internalize the underlying logic. The goal is to transition from a tool-user relationship to a collaborative partnership where the agent's internal states are accessible and adaptable to the learner's needs.
Methodology: The Minicomputer Tug of War
To investigate these dynamics, the research utilizes a collaborative game scenario called "Minicomputer Tug of War." This game is based on Papy’s Minicomputer, a non-verbal tool used to teach binary and decimal arithmetic.
Decision-Making and Explanation Generation
The agent employs a minimax search algorithm to determine its moves. Crucially, the agent does not always play perfectly. To facilitate learning, it occasionally selects sub-optimal strategies. Explanations are then generated by comparing the "optimal" action (the best move identified by the search) with the chosen "sub-optimal" action. By articulating the difference between these choices, the agent provides strategic "hints" that help the child predict the agent's behavior and grasp the global strategy of the game.
Two-Way Transparency
The framework emphasizes a "two-way transparency" model. This involves the agent explaining its own logic while simultaneously remaining open to human feedback. This approach is designed to support a shared mental model, ensuring that the agent's explanations align with the child's evolving understanding of the problem.
Key Insights
Transparency as a Social Bridge
In a pedagogical context, transparency serves as a social tool rather than just a debugging mechanism. By revealing its reasoning, the agent establishes "human-likeness," which helps the child feel like part of a cohesive team. This social bond is essential for maintaining engagement in long-term learning scenarios.
The Strategic Value of Sub-optimality
An agent that consistently plays perfectly is often a poor teacher, as its logic may be too advanced for the learner to follow. By intentionally choosing sub-optimal moves and explaining why they are inferior to the optimal ones, the agent encourages the child to engage in "goal heuristics"—strategic shortcuts that simplify complex decision-making.
Matching Mental Models
For an explanation to be effective, it must resonate with the child's existing mental model. The research identifies that the most impactful "lightbulb moments" occur when the agent reveals specific parts of its heuristic algorithm that match the child's current level of logical development.
Preliminary Results
Initial studies (Tulli et al. 2019) have yielded promising results regarding the impact of agent transparency:
- Increased Trust and Identification: Transparency significantly enhances the level of trust children place in the system and increases their identification with the agent as a group member.
- Perception of Human-likeness: Agents that explain their reasoning are perceived as more human-like, which is a key driver for social learning.
- Cognitive Engagement: Observations suggest that by making the agent's planning process visible, children are better able to engage in mental computation and logical thinking regarding mathematical rules.
Future Directions
While the current framework provides a strong foundation, several areas for future research remain:
- Personalization: Future iterations must adapt explanations to the unique mental models of individual children, recognizing that educational needs vary widely.
- Interactive Task Learning: The research aims to evolve toward a bidirectional dialogue where the agent can also ask the child to explain their actions, creating a more robust educational exchange.
- Long-term Validation: Further longitudinal studies are required to quantify the long-term impact of these explanations on the actual acquisition of mathematical and logical skills.
Summary updated