Large language models are not about natural language

TL;DR

The commentary challenges claims that large language models can explain linguistic structure, processing, and learning by contrasting their analysis of externalized word strings with the human language system. It argues that LLMs do not form meaning-determining structures and cannot meaningfully inform the human language faculty in their current probabilistic form.

• Problem

  The commentary disputes the claim that LLMs are useful for understanding linguistic structure, processing, and learning because they work differently from the human language system.

  Source: p. 1

• Method

  The paper contrasts LLMs’ analysis of externalized strings with the human computational language system, drawing on child development and possible-versus-impossible language comparisons.

  Source: pp. 1–2

• Results

  LLMs simulate impossible languages as well as possible languages and do not form hierarchical structures that determine meaning.

  Source: pp. 2–3

• Takeaways & Limitations

  Given these differences, the authors conclude that current probabilistic LLMs cannot meaningfully inform the human language faculty.

  Source: p. 3

  The authors limit their conclusion to LLMs in their current probabilistic form.

  Source: p. 3