StoryScope reads plot, character, setting and time rather than word choice, so a paraphrase pass would not be enough to evade tools built on it.
Narrative traits measured purely at the discourse level are enough on their own to separate human-written fiction from machine-generated text, at 93.2% macro-F1, and they retain more than 97% of the performance of models that also read stylistic cues.
The system behind the result is StoryScope, introduced in an arXiv preprint, which automatically builds an interpretable feature space spanning 10 discourse-level dimensions.
Those dimensions cover how a piece of fiction develops its plot, describes its characters, and organises setting and time. To test them, researchers took 10,272 human-written stories, had Gemini 2.5 work backward from them to produce writing prompts, then ran those prompts through Claude Sonnet 4.6, DeepSeek V3.2, GPT 5.4, Gemini 3 Flash and Kimi K2.5. The resulting parallel corpus holds 61,608 stories of roughly 5,000 words each, with 304 features extracted from every one.
Machine-written stories over-explain their central messages and favour tidy, single-strand plots, while human-written ones give characters' decisions more moral grey area and handle time in more layered ways. AI narrators state the theme outright 77% of the time, against 52% for human authors, and AI dialogue becomes a vehicle for philosophical debate in 59% of cases, versus 34%. Plotted in narrative feature space, the machine-written stories cluster in a shared region. The human ones spread far more widely.
Those tells also vary by model, carrying the method from detection into attribution: naming the author among six candidates — the human writer plus the five systems — reaches 68.4% macro-F1. Claude Sonnet 4.6 lets tension climb hardly at all across a story, GPT 5.4 reaches for dream sequences well past what a plot needs, and Gemini 3 Flash draws its characters from the outside in.
Because the signal sits in structure rather than wording, detection products built on this approach could not be evaded by a thesaurus pass. Ghostwriters, marketing teams and students leaning on automated drafts would have to change how the story itself is built.
The human half of the corpus comes from Books3, a collection of about 183,000 books assembled from pirated ebooks that has been used to train large language models and is the subject of several lawsuits. The authors say they recognise the copyright concerns, do not endorse using the dataset to train models or generate commercial text, and confine their own use to academic research. The team is affiliated with University of Maryland, College Park and Google DeepMind; one author, Maryland researcher Jenna Russell, also holds an internship at the AI-detection company Pangram.
The arXiv preprint was first submitted on 3 April 2026. What the reported results do not settle is how the detector fares on hybrid drafts — where a person substantially rewrites an AI-generated first pass — or on genres beyond those examined.