Does GPT-5 dream of (il)literate minds? A network analysis of semantic fluency

Abstract

Research comparing human and machine language abilities has largely relied on WEIRD (Western, Educated, Industrialized, Rich, Democratic) populations, neglecting the cognitive diversity introduced by literacy variation. This study investigates how literacy and text-based training shape semantic organization by comparing illiterate and literate Turkish adults with ChatGPT-5. Participants and the model completed a 1-minute semantic fluency task in three categories (animals, fruits, household objects). Responses were analyzed using network science metrics: average shortest path length (ASPL), clustering coefficient (CC), and modularity (Q). Results showed a stepwise pattern across groups. Illiterate speakers produced smaller but locally dense networks (short ASPL, high CC, low Q). Literate speakers demonstrated more balanced, small-world-like structures (moderate ASPL and CC). GPT-5, while generating the largest vocabulary, produced fragmented networks with long paths, low clustering, and high modularity. These findings suggest that literacy promotes efficient integration of semantic knowledge, whereas text-only learning yields over-segmented “hyper-literate” structures. The results highlight how embodied experience and print exposure jointly shape semantic memory organization in humans and GPT-5.