Can Large Language Models Teach Creativity? Cognitive Gains, Affective Gaps, and Distinct Neural Patterns

As creativity gains increasing importance in education amid rapid technological advancements, the capacity of large language models (LLMs), such as ChatGPT, to foster creative skills remains largely unexplored. This study investigates the efficacy of AI as a creativity instructor compared to a human-led instructed group and a no-intervention control group, examining cognitive and affective outcomes through behavioral, perceptual, and neural indicators. Employing a 3 × 2 mixed design, 95 participants were randomized into AI instruction, human instruction, or control conditions. Creativity was evaluated using the Alternative Uses Task, which quantifies fluency, originality, and flexibility in both acquisition and transfer. Functional near-infrared spectroscopy (fNIRS) captured brain activity in the prefrontal cortex (PFC) and right temporo-parietal junction (rTPJ), while self-reported measures captured perceptual affiliative bonding, interaction quality, and perceived AI impact. Results indicated that AI significantly improved creativity, particularly originality and flexibility in acquisition and transfer. Notably, AI and human instructors engaged distinct neurocognitive systems: greater localized superior frontal gyrus activation correlated with AI's cognitive gains, whereas human instructors induced stronger prefrontal functional connectivity. Although perceived as useful, AI exhibited a distinct affective gap, eliciting significantly weaker affiliative bonding than human tutors. These findings highlight AI’s cognitive advantages in creativity education and underscore deficits in socio-emotional engagement. Future work could address these affective limitations and explore hybrid human-AI instructional models to optimize both cognitive and affiliative learning outcomes. • Examined LLMs as creativity instructors via behavioral and neural data • AI instruction improved creativity, especially originality and flexibility • Human teaching showed stronger brain connectivity and emotional bonding • Identified distinct neural bases for AI and human instruction • Proposed hybrid human–AI teaching for balanced learning outcomes