A new paper by Dr. Katrin Rohrbacher, Björn Nieth, Dr. Emmanuelle Salin, Prof. Dr. Björn Eskofier and Prof. Dr. Michaela Mahlberg, accepted to the main conference of EMNLP 2026 in Budapest (24 to 29 October), asks how large language models build fictional worlds. The authors compared 8,000 stories generated by four LLMs in English and German with human-authored fiction, classifying each sentence by type of narrative space. The paper finds that human fiction relies mainly on action space, where characters move through and interact with their surroundings. The models consistently overproduce perceived space, which foregrounds atmosphere and mood. LLM storytelling thus appears rich in ambience and affect but less anchored in embodied action and spatial concreteness. The findings point to narratology-informed spatial metrics as a new way of evaluating AI-generated narratives.
The preprint is available at https://doi.org/10.48550/arXiv.2609.02482
Conference Website: https://2026.emnlp.org/