Semantic networks in humans and Large Language Models in Spanish

Semantic fluency tasks, like those traditionally employed in lexical availability research, provide a well-established method for analysing how lexical knowledge is retrieved and organised in humans. This study applies these tasks to Large Language Models (LLMs), comparing their lexical productions with those of human participants in Spanish for the category animals. One hundred university students and five LLMs (ChatGPT, Gemini, Claude, Grok, LLaMA) were tested, and their responses were analysed using both quantitative and qualitative data and network-based metrics. Results reveal a clear asymmetry between humans and LLMs. Human participants generated more diverse and heterogeneous lexical sets, producing networks characterised by lower connectivity but higher modularity. LLMs, by contrast, displayed convergent and cohesive outputs, forming denser networks with higher average degrees and clustering but reduced modularity and novelty. Disaggregated analysis showed systematic differences across the LLMs themselves. These findings confirm that LLMs approximate the underlying structure of semantic fluency but do not fully replicate the variability and contextual sensitivity characteristic of human lexical production. The study highlights both the strengths and limitations of artificial systems as models of lexical organization and points to new directions for cross-linguistic comparisons.

Abstraction · Fluency · Lexical item · Replicate · Semantic network · Vocabulary · Language Development and Disorders · Neurobiology of Language and Bilingualism · Second Language Acquisition and Learning