Abstract
Artificial intelligence (AI) has the potential to transform education by uncovering insights from massive data about student learning patterns. However, ethical and trustworthy concerns regarding AI remain significant hurdles. Prominent issues in high school AI education include data privacy, information leakage, abusive language, and fairness. This paper describes the technological components built to address these concerns within ALLURE, a multi-modal collaborative platform where high school students work with an AI to solve the Rubik's Cube. The system prioritizes informed consent, moderates language to ensure safety for children, and implements safeguards against information leakage between users.
Introduction
Presented at ICML 2023, ALLURE is a multi-modal chatbot designed to tutor high school students in solving the Rubik’s Cube. While pedagogical efficacy is often the primary focus of AI tutors, the authors argue that trust and ethics—specifically regarding data privacy, information leakage, and abusive language—are just as critical for successful adoption in educational settings. ALLURE addresses the "black-box" nature of modern AI, which can stifle learning if students do not understand the reasoning behind suggested actions.
Addressing Ethical Risks in Educational AI
Deploying AI in schools introduces unique ethical risks that ALLURE is designed to mitigate:
- Information Leakage: Preventing students from "social engineering" the bot to reveal peer performance or private data.
- Abusive Language: Moderating interactions to ensure the AI neither produces nor tolerates toxic language, which is vital when working with children.
- Explainability Gap: Bridging the gap between superhuman AI strategies and human-memorizable algorithms.
The ALLURE Architecture
The system employs a multi-layered technical approach to ensure both high performance and human-centric tutoring. Built on the RASA framework, the architecture includes specialized components for language understanding and ethical oversight.
Key components include:
- DeepXube (Explainable Solving): Combines DeepCubeA (Deep Reinforcement Learning) with Inductive Logic Programming (ILP) to find human-suitable "macro-actions." These repeatable algorithms (e.g., the "white cross" step) are translated into plain English via Natural Language Generation (NLG).
- Sentiment Analyzer: Uses VADER to estimate the user's emotional state, allowing the bot to provide appropriate encouragement.
- Modified Intent Classifier: Specifically trained to detect abusive language and requests for sensitive information, ensuring the bot stays within ethical boundaries.
Key Results and Moderation
The system demonstrates robust performance in both its tutoring and ethical moderation tasks. It successfully identifies and explains complex sub-goals like the "White Cross," providing specific move sequences (e.g., D’, F’, R, F) to correct out-of-place pieces.
Crucially, the bot effectively manages improper language through an incremental "strike" system. In simulated dialogues, the bot identifies abusive terms and responds with warnings or by ignoring the input to discourage the behavior.
Beyond moderation, the system demonstrates privacy preservation by identifying and denying requests for peer data, categorizing such queries as potential privacy violations. This ensures that the system, while learning to improve over time, does not leak information about users from one group to another.
Conclusion
While the current framework is robust, future work will focus on field validation with high school students, addressing potential security vulnerabilities like adversarial attacks, and enhancing dynamic personalization to adapt linguistic styles to individual student needs. By integrating explainability with rigorous ethical safeguards, ALLURE provides a model for trustworthy AI in the classroom.
Summary updated