Computer as a Spiritual Tool (CAST) 2 – The Nature of Large Language Models (LLMs), or How LLMs Describe Themselves - mauriceling/mauriceling.github.io GitHub Wiki

Citation: Ling, MHT. 2026. Computer as a Spiritual Tool (CAST) 2 – The Nature of Large Language Models (LLMs), or How LLMs Describe Themselves. Acta Scientific Computer Sciences 8(1): 10-18.

Link to [PDF].

Here is the permanent [PDF] and [dataset] links to my archive.

Large language models (LLMs) have demonstrated remarkable ability in generating coherent natural language and producing explanations of their own outputs. However, the epistemic status of such self-descriptions remains unclear. This study investigates how LLMs characterize their own nature, capabilities, and internal processes, and whether these self-descriptions support interpretations of LLMs as systems that manipulate symbols without semantic understanding, as proposed in the Chinese Room argument. Using a human-in-the-loop dialogical protocol inspired by postdigital duoethnography and multi-agent debate, a structured three-way interaction was conducted between the author, ChatGPT, and Microsoft Copilot across three research questions. The resulting transcripts were analysed qualitatively to identify recurring themes. Four key themes emerged: (i) LLMs consistently describe themselves as transformer-based systems that learn statistical patterns from large corpora and generate probabilistic outputs. (ii) Their capabilities, ranging from explanation to reasoning-like behaviour, are framed as emergent from architecture, data, and alignment rather than grounded understanding. (iii) Both models provide convergent accounts of their internal processes as a pipeline of tokenization, contextual encoding via attention, and autoregressive next-token prediction. (iv) Their self-descriptions strongly align with the Chinese Room interpretation, emphasizing symbolic manipulation without intrinsic meaning while simultaneously acknowledging emergent behaviours that simulate understanding. Overall, the findings suggest that LLMs are best understood as generative models of linguistic probability spaces. Their self-explanations are coherent and internally consistent, yet reflect functional descriptions rather than transparent accounts of underlying computation or genuine comprehension.