Building Reflection Competencies for Human-AI Collaboration: A Multi-Agent Training System

Generative AI (GenAI) tools like ChatGPT are widely adopted, with Singapore leading global usage at 4.6 times the per-capita average. While regulations mandate oversight, they do not equip workers with skills to critically evaluate AI outputs. This gap is urgent because 95% of GenAI projects fail to deliver impact, often due to overreliance on AI and declining critical thinking—a phenomenon termed “cognitive debt.” Current training focuses on technical skills like prompt writing but neglects reflection competencies, leaving organizations without scalable solutions to balance trust and scrutiny. Existing interventions either frustrate users or increase overreliance, creating a reflection–adoption paradox.​

​This research proposes a dual pathway model integrating Protection Motivation Theory and Self-Determination Theory to activate both threat-based and growth-based motivations for reflection. The approach includes developing a multi-agent reflection training system embedded in AI assistants, featuring components like persuasion detection, user state recognition, and reflection generation. A 2×2 factorial experiment will test interventions—Value Inquiry and Tactic Disclosure—designed to enhance reflection depth and breadth while improving adoption intent. Deliverables include an open-source system, empirical validation, publications, and practitioner engagement. By operationalizing oversight requirements and preventing cognitive debt, this project offers a scalable solution for Singapore’s workforce, ensuring safe and effective AI integration in a knowledge-driven economy.​

Research Impact: ​Overcoming AI users’ cognitive debt through reflection training for a resilient workforce