Large Language Models and Associative Fields in AI-based Creative Writing and Literary Interpretation

Abstract

Artificial intelligence-supported Large Language Models (LLMs) such as GPT-4, Claude, and Gemini have made remarkable progress in seemingly understanding and generating natural language texts. However, their performance in interpreting and regenerating the associative field of the text, especially in terms of short, poetic, or aphoristic texts, remains deeply limited. By associative field, we refer to the semantic, emotional, cultural, and symbolic network evoked by a text beyond its literal meaning. This paper investigates the key shortcomings in LLMs’ grasp of associative fields, explores the underlying causes, and evaluates the implications for literary analysis, creativity, and human-machine communication. We hypothesise that the smaller the corpus and the more complicated the (human) associative field around the sample poem, the less likely we will expect anything valuable from the LLMs. That is, the fewer points of contact the model has to map the associative field, the more difficult it will be to meet expectations and create something similar to the original and of a similar standard. We provide some relevant examples of these shortcomings by (the lack of) generating poems analogous to some famous English writers, including Shakespeare, and some poems of Hungarian authors Sándor Weöres and László Nagy.