Narrative Fingerprint of AI Fiction
404 Media · rss · 2026-07-11
Research Conclusion
This preprint from University of Maryland and Google DeepMind argues that AI-written fiction is recognizable not just by surface style (punctuation, word choice) but because its narrative structure is more formulaic.
Key Findings
- AI stories more often over-explain themes, make meaning too explicit, and favor linear, tidy plot progression.
- Human works more often feature moral ambiguity, more time jumps and flashbacks, and handle complex multi-character, multi-location narratives.
- Different models have tendencies: e.g., Claude flatter event progression, GPT more dream sequences, Gemini more external character description.
- The study suggests these differences stem from narrative construction, not just style, and can distinguish human originals from AI-generated texts.
Method and Data
Researchers proposed a detector called StoryScope based on NarraBench classifications of narrative features, focusing on plot development, character description, scene, and temporal structure.
Test process:
- Selected 10,272 human-written short stories
- Used Gemini 2.5 to reverse-generate prompts
- Fed those prompts to Gemini 3 Flash, DeepSeek V3.2, Claude Sonnet 4.6, Kimi K2.5, and GPT 5.4 to generate AI stories
The stories came from Books3 dataset (183k pirated ebooks), with explicit mention of copyright controversy, and stated for academic use only, not for training or commercial generation.
Other Notes
Paper also discloses use of coding agents like Claude Code and Codex to assist writing, tables, and figures. Authors support fuller disclosure of AI use and note that agents significantly boost efficiency in code implementation and paper polish.
More from Research
- AI veteran Thomas Dietterich: stop using cs.AI as your primary arXiv category — tdietterich · 2026-10-27
- arXiv advice from T. Dietterich: wrong primary category will almost surely get your paper on hold — tdietterich · 2026-10-27
- arXiv has 4,000+ papers on hold amid submission surge; moderator shares how to avoid it — tdietterich · 2026-10-27
- Community-trained MOSS Sharpness Slider LoRA: 72 photos, 500 steps, +/-1 focus control — aoleg77 · 2026-10-27
- CoreWeave, You.com and NVIDIA lift BrowseComp 22.9% in 3-week RL post-training sprint — RichardSocher · 2026-10-27
- Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them — heathercmiller · 2026-10-27