Have we stopped to think about what LLMs model?

Recent discussions surrounding large language models (LLMs) highlight a critical examination of what these technologies actually represent in terms of language and cognition. A peer-reviewed paper by Abeba Birhane and Marek McGann critiques the prevalent claims that LLMs can "understand" language, arguing that such assertions misinterpret the nature of human linguistic capabilities. The authors emphasize that LLMs, which are built on vast datasets and statistical techniques, do not replicate the complexities of human language, which is inherently social, contextual, and embodied. They argue that LLMs operate under flawed assumptions about language completeness and data representation, failing to capture the nuances of human interaction, such as emotional context and social participation. The paper warns that treating LLMs as language understanding machines can lead to misguided policies and social implications. Furthermore, the authors express concern over the deployment of LLMs in critical sectors like education and healthcare without adequate testing and regulation, highlighting the potential risks of misinformation and unreliability. While the AI industry continues to promote the economic benefits of LLMs, the authors call for a more cautious and skeptical approach to their development and application.

- The paper critiques the exaggerated claims about LLMs' understanding of language.

- LLMs are based on flawed assumptions about language and data representation.

- Human language is complex and cannot be fully captured by LLMs.

- There are significant risks in deploying LLMs in critical sectors without proper regulation.

- A more cautious approach to LLM development is advocated by researchers.