Abstract: In the era of generative artificial intelligence, understanding the nature and origin of machines’ knowledge of language has become a critical inquiry across philosophy, linguistics, and artificial intelligence. This paper aims to explore the fundamental differences between advanced large language models and human children in core dimensions of language learning by comparing their language acquisition mechanisms. Children acquire language through an innate language acquisition device and universal grammar, supplemented by relatively limited postnatal language input alongside rich social interactions and embodied experiences. In contrast, machine learning is based on large datasets, advanced deep learning algorithms, and powerful computing resources. Despite their proficiency in language generation tasks, current language models primarily excel in pattern recognition, lacking deeper cognitive understanding. Similar to children, language models face challenges in extracting the intricate deep structure of language from imperfect input data.

Key Words: Generative artificial intelligence; Machine language; Child language; Knowledge of language;

Plato’s problem

This article can be downloaded here

https://jdn.ucas.ac.cn/public/uploads/files/67078b367a415.pdf