Authors: Hongming Li, Shan Zhang, Anthony F. Botelho Source: Computers and Education: Artificial Intelligence, Vol 11, 100646 — Open Access (CC BY 4.0)
Key Findings
Proposes VETTING — a dual-LLM architecture where a generator LLM produces responses and a separate verifier LLM checks them against safety policies. Deployed with 151 middle school students, achieving F1=0.928 for safety violation detection and reducing inappropriate content exposure by 91.2%. Documents a taxonomy of student boundary-testing behaviors. Open-source Python implementation available.