Artificial Intelligence and Psycholinguistics: How Do Large Language Models Represent Human Language Processes? A Systematic Review
TL;DR
Current research examining the similarities and differences between human language processing and LLM-based language representation indicates that LLMs successfully approximate numerous observable characteristics of human language behavior, including syntactic processing, contextual prediction, semantic association, discourse coherence, and pragmatic inference.
Abstract
The rapid development of large language models (LLMs) has fundamentally transformed contemporary perspectives on language representation and processing. Models such as GPT, Gemini, Claude, Llama, and DeepSeek demonstrate remarkable capabilities in language comprehension, generation, translation, summarization, and reasoning, raising important questions regarding their relationship to human language processing. While psycholinguistic theories traditionally explain language through cognitive mechanisms involving perception, memory, attention, semantic representation, and executive control, LLMs rely on statistical learning, neural network architectures, and large-scale pattern recognition. The present study synthesizes contemporary research examining the similarities and differences between human language processing and LLM-based language representation. Employing a qualitative systematic review, the study integrates empirical findings published between 2018 and 2025 from psycholinguistics, cognitive science, neuroscience, computational linguistics, and artificial intelligence. The findings indicate that LLMs successfully approximate numerous observable characteristics of human language behavior, including syntactic processing, contextual prediction, semantic association, discourse coherence, and pragmatic inference. However, important differences remain regarding grounding, intentionality, episodic memory, emotional experience, and embodied cognition. Although LLMs provide valuable computational models for investigating psycholinguistic theories, they cannot currently be considered cognitive equivalents of human language users. Instead, they represent sophisticated probabilistic systems that emulate linguistic behavior without reproducing the complete cognitive architecture underlying human communication. The review discusses theoretical implications for psycholinguistics, cognitive science, and artificial intelligence while identifying future research directions for integrating computational and cognitive approaches to language.