Large language models are not about natural language

Area of Science:

• Computational Linguistics
• Cognitive Science
• Psycholinguistics

Background:

• Large Language Models (LLMs) analyze externalized word strings using probabilistic methods and vast datasets.
• Human language acquisition and processing are theorized to involve a mind-internal, recursive computational system.
• A key characteristic of human language is its generative capacity and ability to discern grammaticality.

Purpose of the Study:

• To critically evaluate the utility of Large Language Models (LLMs) in the field of linguistics.
• To contrast the operational principles of LLMs with the theoretical underpinnings of the human language faculty.
• To highlight the limitations of data-driven probabilistic models in explaining core linguistic phenomena.

Main Methods:

• Comparative analysis of LLM architecture and function against established linguistic theories.
• Examination of the data requirements for LLMs versus the minimal input needed for human language acquisition.
• Assessment of LLMs' capabilities in distinguishing grammatical from ungrammatical or impossible language constructions.

Main Results:

• LLMs, as probabilistic models, necessitate extensive data for analyzing externalized language.
• Human language is supported by an internal computational system capable of recursive generation of hierarchical structures.
• The human language system demonstrates robust growth with limited external input and innate grammaticality judgments.

Conclusions:

• LLMs are fundamentally limited in their ability to model the core mechanisms of human language.
• The generative and recursive nature of the human language faculty is not adequately captured by current probabilistic LLM approaches.
• Linguistic research requires models that account for the internal, computational basis of language, not just surface-level statistical patterns.