The Limitations of Large Language Models for Understanding Human Language and Cognition.

Area of Science:

• Cognitive Science
• Computational Linguistics
• Evolutionary Biology

Background:

• Large Language Models (LLMs) are increasingly proposed as tools to investigate human language development and evolution.
• Debates persist regarding the balance of learning versus innate factors in language acquisition.

Purpose of the Study:

• To critically evaluate the extent to which LLMs can inform our understanding of human language acquisition and evolution.
• To differentiate between functional similarities and mechanistic differences between LLMs and human language.

Main Methods:

• Comparative analysis of LLM processing (text-based) versus human language acquisition (multimodal, interactive).
• Application of the ethological "four questions" framework to assess LLM capabilities.
• Examination of the scope and limitations of LLMs in representing the breadth of human linguistic behavior.

Main Results:

• Similarities between LLM output and human language are superficial and functional, not mechanistic.
• LLMs primarily process vast amounts of unimodal text, unlike humans' rich multimodal input.
• LLMs' design for imitating writing limits their capacity for naturalistic interaction and understanding diverse human language.

Conclusions:

• LLMs alone provide minimal insight into human language acquisition and evolution.
• LLMs are tools, not theories; their effective use requires specific research hypotheses.
• Understanding human language requires acknowledging the limitations of current LLM architectures.