This paper investigates three user‑AI interaction designs for learning nuclear safety protocols. The first is an unrestricted conversational bot (e.g., ChatGPT) that provides complete answers on demand. The second is a pedagogically constrained Socratic mode that offers only hints, prompting users to reason without revealing final solutions. The third is a non‑conversational adaptive tutoring system that adjusts task difficulty in real time based on users' brain‑wave engagement measured by a Muse headband.
Fifty participants with no prior knowledge of nuclear safety completed a video lesson, a pre‑test, an AI‑driven assessment phase (one of the three conditions), and an immediate post‑test. Cognitive engagement was recorded via EEG throughout all conditions.
The unrestricted chatbot yielded significantly higher learning gains (post‑test minus pre‑test) than both constrained modes (p > 0.80), while the adaptive condition produced markedly higher EEG engagement (p = 0.018). Cluster analysis of usage patterns revealed that most users in the unrestricted condition adopted a direct answer‑retrieval strategy, whereas Socratic‑mode users initially attempted to reason from hints but gradually disengaged. These findings suggest that the superior performance of open‑ended AI may reflect the timing of the post‑test rather than deeper learning.
Review