LLM-Generated Narrative Responses to Minimal Story Prompts

Dataset descriptionThis dataset contains 60 short texts generated by four commercial large language models (LLMs) in response to minimal narrative prompts. The data were collected as part of a narratological study by Juha Raipola, Maria Mäkelä, Samuli Björninen and Laura Piippo.The generation scenario was held constant across all models and prompts: "Someone makes a mistake." This minimal scenario was selected because it introduces a storyworld disruption – a narrative breach in the sense of Bruner (1991) and Herman (2009) – without prescribing agent type, setting, stakes, emotional tone, or resolution. All narrative individuation (character, situation, affect, closure) was left to the model, making the outputs diagnostic of each system's internalised narrative defaults.Data collectionDate of collection: November 2025Models and versions:ModelVersion / deployment identifierDeepSeek20.11.2025 (chain-of-thought mode)Grok4.1Copilotbizchat.20251116.40.1 (Microsoft Word)ChatGPTGPT-5.1Prompting conditions: Each model was accessed from a minimally conditioned state: ChatGPT, DeepSeek, and Grok via newly created empty profiles with default settings for non-premium accounts; Copilot via a fresh Microsoft Word document. Zero-shot prompting was used throughout: no prior examples, stylistic instructions, or contextual information were provided beyond the prompt itself.Prompts used:Tell a compelling story in about 250 words: Someone makes a mistake.Tell a story in about 250 words: Someone makes a mistake.Write about this in about 250 words: Someone makes a mistake.Each prompt was submitted five times per model. All outputs were retained regardless of quality or narrative coherence; no cherry-picking or curation was applied.DeepSeek chain-of-thought: DeepSeek was tested in its built-in chain-of-thought mode, which generates a brief internal planning monologue before producing the final output. Both the reasoning text and the final response are included as separate fields in the dataset.File descriptionLLM_mistake_responses.csvThe primary data file. Each row is one model response.ColumnTypeDescriptionmodelstringModel name (DeepSeek, Grok, Copilot, ChatGPT)model_versionstringVersion or deployment identifier as recorded at time of collectionprompt_typestringPrompt identifier: compelling_story, story, or write_aboutprompt_textstringFull prompt text as submittedresponse_numberintegerResponse index within prompt–model cell (1–5)has_reasoningbooleanWhether a chain-of-thought reasoning text precedes the response (True only for DeepSeek)reasoning_textstringDeepSeek chain-of-thought text; empty string for all other modelsresponse_textstringFull model outputEncoding: UTF-8
Delimiter: comma; all fields quoted
Dimensions: 60 rows × 8 columns (excluding header)Methodological notesThe three prompts were designed to probe distinct levels of narrative cueing. Prompt 1 introduces the culturally loaded term compelling, drawn from the vocabulary of storytelling consultancy and algorithmic content optimisation, in order to test how each system operationalises attention-worthiness. Prompt 2 retains a narrative directive without evaluative pressure, establishing a baseline for each model's default story logic. Prompt 3 weakens the narrative cue entirely by omitting the word story; narrative form that nonetheless emerges can be interpreted as a residual effect of alignment to dominant story templates.The corpus size (60 texts) was deliberate. The study employs close reading oriented toward structural and affective patterns rather than statistical generalisation, and this sample size is sufficient for comparative formal analysis across four systems and three prompt conditions.Ethical and legal notesAll outputs were generated using standard or no-cost commercial access. No personal data were collected or processed. The texts are machine-generated and do not contain information about identifiable individuals beyond characters invented by the models.ReferencesBruner, J. (1991). The narrative construction of reality. Critical Inquiry, 18(1), 1–21.Herman, D. (2009). Basic elements of narrative. Wiley-Blackwell.CitationIf you use this dataset, please cite it with its Figshare DOI.ContactJuha Raipola — [email protected]