How AI fiction differs from human stories

AI fiction differs from human fiction less in its sentences than in its choices. A large study called StoryScope compared more than 10,000 human short stories with versions written by five AI models from the same prompts. The AI stories spelled out their themes, showed emotion through the body, followed a single tidy plot line and ended with the protagonist's own choice. The human stories were messier, more ambiguous and harder to predict.

What the StoryScope study of AI fiction did

StoryScope was written by Jenna Russell, Rishanth Rajendhran, Chau Minh Pham and Mohit Iyyer of the University of Maryland and John Wieting of Google DeepMind, and published at COLM 2026. Most research on detecting AI writing looks at surface style: word choice, punctuation, sentence shapes. These authors asked a different question. If you ignore style entirely, can you still tell an AI story from a human one by how it is told?

They started from 10,272 human-written stories drawn from the Books3 collection of published books. For each story they worked backwards to a writing prompt describing its premise, then had five models write a story from that same prompt: Gemini 3 Flash, Kimi K2.5, DeepSeek V3.2, Claude Sonnet 4.6 and GPT-5.4. That gave 61,608 stories averaging about 4,750 words.

Each story was then described along ten narrative dimensions, such as plot, characters, setting and how time is handled. From comparisons between sources, the pipeline produced 304 specific features, for example whether the narrator comments directly on the theme or how often the story jumps in time. The researchers trained classifiers on those features with the style-related ones removed.

How well narrative choices alone identify AI stories

Very well. Using narrative features alone, the classifier separated human from AI stories with a macro-F1 score of 93.2% (a measure that balances misses and false alarms, where 100% is perfect). Adding style features raised it to 96.0%, so narrative choices carried more than 97% of the combined signal. A compact set of just 30 core features still reached 84.8%.

For context, classifiers that read the raw text, such as a fine-tuned ModernBERT model, did even better on this data, at 99.9%. The zero-shot detector Binoculars managed 55.9%. StoryScope's contribution lies elsewhere: it shows that the gap between human and AI fiction runs deeper than prose style, and it makes that gap readable, feature by feature.

AI states its themes outright

The clearest pattern is over-explanation. In AI stories, the narrator explicitly explained the story's theme 77% of the time, against 52% for human authors. The paper gives a typical case: a grieving character's arc ends with the narrator spelling out the lesson learned.

Related habits follow. AI characters used dialogue for philosophical debate more often (59% against 34%). When AI stories referred to other works, they leaned on vague allusions (72% against 50%), while human authors named specific books and writers at nearly double the AI rate (47% against 24%). The authors' summary is that AI spells out meaning instead of trusting the reader to find it.

AI writes emotion through the body

Writing teachers say "show, don't tell", and AI models seem to have taken it to heart. AI stories conveyed emotion through physical sensations and bodily metaphors in 81% of cases, against 38% for human stories. Where a human writer might simply say a character was afraid, an AI story tends to describe a tight chest, a cold sweat and a lamp growing dim.

The senses get the same treatment. AI stories used smell imagery 82% of the time, against 57% for humans, and leaned on setting to mirror characters' inner states. Human authors, by contrast, named emotions directly in 29% of stories, against just 8% for AI. Plain statements of feeling turn out to be a human trait.

AI plots run on a single track

AI stories had tighter cause-and-effect chains and fewer loose ends. 79% had no subplots at all, against 57% of human stories. Their endings were more often driven by the protagonist's own choice (69% against 46%) and more often resolved through inner understanding or acceptance (47% against 27%).

Human authors played with time more, using time jumps, flashbacks, flash-forwards and nonlinear structure to hold back key revelations. The paper's example: a human mystery might open at the funeral and work backwards through decades, while an AI version walks from the first clue to the final reveal in order. Human stories also wove subplots into the main theme twice as often (42% against 21%).

Human stories are more ambiguous and more varied

Human authors presented morally ambivalent protagonists in 59% of stories, against 38% for AI. They broke the fourth wall more often (67% against 39%) and addressed the reader directly more often (28% against 7%). They also spread their stories across more locations and used more dialogue relative to narration.

The study measured how unusual each story was compared with its neighbours. Human stories were rarer on average: 24.7% of them fell in the rarest tenth of the whole corpus, against 7.1% of AI stories. For a given prompt, the human version was the most unusual of the six 57.8% of the time, where chance would be 16.7%. The five AI models clustered together in one region of narrative space; human stories sat apart and spread out further.

Each model has its own fingerprint

The same features could also tell the models apart, though less reliably: 68.4% macro-F1 across all six sources from narrative alone, and 77.3% with style, against 16.7% by chance. Some of the per-model habits the authors describe:

• Claude was the most distinctive model. Its stories were restrained, with less escalation of events than any other source, and respectful of literary convention rather than subverting it (62% against 39% to 56% for the others). It liked epilogues and quiet endings.
• GPT used gossip and rumour as a plot device more than anyone (64% against 44% to 55%), and often framed stories as reflections on events years earlier.
• Gemini produced the tidiest endings and the bleakest settings, with 88% tagged as bleak and oppressive.
• DeepSeek tended to put key context up front, and Kimi had the fewest distinctive habits, sitting near the middle of the AI cluster.

Do these differences survive editing?

Mostly, yes. The authors took 278 Gemini stories and ran them through a rewriting method designed to remove common AI writing artifacts, such as clichés and purple prose. The narrative classifier scored 93.9% on the edited stories against 95.5% on the originals, a drop of only 1.6 points. Polishing the sentences didn't change how the stories were built.

A few things are worth keeping in mind. The stories were long, about 5,000 words, which gives the features room to show. The human stories came from published books, so the comparison is with published fiction rather than, say, amateur writing online. And the AI stories came from five specific 2026 models. Newer models may tell stories differently, and no single feature in a single story is evidence of anything.

How we use this in our narrative fingerprint

When you check fiction with IsThisAiWritten, the report includes a narrative fingerprint based on this research. It shows measures such as dialogue share, smell imagery, the narrator's thematic commentary, direct address to the reader and time-jump markers, each next to the human and AI base rates published in StoryScope. We compute them with our own word lists, so they are heuristics inspired by the paper, not a reimplementation of it. They are descriptive only and never feed the verdict.

The fingerprint gives editors and writers something concrete to look at next to the verdict. Like any detector, ours can be wrong in both directions, so treat a result as a reason to look closer, not as proof (our methodology page explains how it is tested). Because there's no word limit, you can run a whole manuscript; see our page on the AI detector for long documents.

Sources

1. Russell, Rajendhran, Pham, Iyyer and Wieting, StoryScope: Investigating idiosyncrasies in AI fiction (COLM 2026, arXiv 2604.03136), checked October 2026
2. StoryScope full text, version 6 (HTML), checked October 2026