Recent advances in large language models (LLMs) have intensified interest in using ChatGPT for educational dialogue. This study examined whether ChatGPT can function as an instructional partner in dialogic teaching and how it differs from a human instructor in learning outcomes and neurocognitive processes. Thirty-four undergraduates were randomly assigned to a ChatGPT group or a human-instructor group while learning histogram equalization in a digital image processing course. Both groups achieved comparable retention scores, but the ChatGPT group scored significantly lower on transfer. EEG analyses showed broadly higher theta power in the ChatGPT group. After region-level averaging and false discovery rate correction, gamma differences were more selective, with higher posterior gamma but lower temporal gamma in the ChatGPT condition. Rather than indicating unequivocally better learning, this pattern is more consistent with greater visually driven cognitive effort during text-based interaction. Post-dialogue microstate analysis further suggested reduced speech-related processing and attentional switching after ChatGPT-based dialogue, although these findings should be interpreted cautiously because they were derived from a short post-dialogue resting window. Overall, ChatGPT appeared capable of supporting basic knowledge acquisition, but it was less effective than the novice human instructor in promoting knowledge transfer and cognitively efficient deep learning. The findings support a hybrid instructional model in which AI tools assist with foundational learning while human instructors provide multimodal scaffolding for higher-order understanding.
Zhang et al. (Tue,) studied this question.