Study of 8,000 AI stories finds scene-writing patterns vary by language
Four models favor sensory atmosphere in English fiction, but German results diverge. The analysis of 8,000 stories traces how settings enter a narrative.
Atmosphere takes more space in the English stories
A comparison of 8,000 generated stories finds a recurring emphasis on sensory atmosphere in English AI fiction. Katrin Rohrbacher and colleagues report that all four tested models devote a larger share of sentences to this category than the human baseline throughout the narrative.
The contrast is visible early. Among the opening 15 sentences, about 47% of GPT 4.1’s English output falls into the perceived-space category, against roughly 19% in human fiction. More atmosphere does not always mean less action: GPT 4.1 stays near or above the human baseline for character–environment interaction, while the other models show deficits.
Shared opening sentences, followed chapter by chapter
Submitted to arXiv on September 2, 2026, the study samples 1,000 literary works in each of two languages. Their first sentences and genres seed four-chapter continuations from GPT 4.1, LlaMA 3.3, Mistral 3.2 and Gemma 3: 1,000 stories for every model–language pairing.
Fine-tuned BERT classifiers distinguish action, sensory atmosphere, static viewing, external description and sentences without spatial content. The analysis compares openings and ten equal-length narrative segments, matching each human excerpt to its generated counterpart’s length.
German stories show a different distribution
GPT 4.1 favors perceived space in both languages. Other models approach or fall below the German human baseline in some or later segments. Chapter-based generation also concentrates atmosphere at chapter starts, without fully explaining the overall English gap.
The preprint separates setting into dimensions such as character action and sensory atmosphere, revealing differences between models and languages. It examines how scenes are constructed; reader responses to those patterns were outside this experiment.