Metacognition and pathway dynamics in generative AI supported design thinking: A process-sensitive evaluation with network-based cognitive mapping

Generative AI (GenAI) is embedded in design education, yet little is known about how different tools reshape metacognition and the pathways linking ideas to outcomes. We compare three conditions, ChatGPT-4o, DeepSeek-R1, and a no-AI control, in a virtual reality cultural heritage serious game task with 51 novice designers. Think-aloud, screen recordings, sketches, and AI logs were coded into network-based cognitive maps and event-anchored indicators around metacognitive nodes. Analyses combine expert ratings of design outcomes with thinking pathway profiles and window diversity after metacognitive events. Both GenAI groups show higher mean design-quality ratings than the control, but in complementary ways. DeepSeek broadens early ideation and shows higher novelty ratings, while ChatGPT strengthens metacognitive regulation and structured convergence, aligning with higher systematicity and interactivity. The findings introduce process-sensitive indicators and phase-aware guidance for configuring exploration-oriented and regulation-oriented assistance in human and AI co-creation.