Beyond the Em-Dash: The Structural Tells of AI Fiction

If you want to spot an AI-generated story, stop looking for an unnatural obsession with em-dashes or the overuse of words like "delve." Instead, look at the plot. According to a new preprint study from the University of Maryland and Google DeepMind, the real "tell" of artificial intelligence is its inability to handle narrative complexity and its overwhelming desire to teach you a lesson.

The research team, led by Jenna Russell, moved beyond simple text-matching to analyze the underlying architecture of fiction. Using a new detection tool called StoryScope, they evaluated over 50,000 AI-generated short stories against more than 10,000 human-written classics from authors like Stephen King and Joyce Carol Oates. To do this, they reverse-engineered human stories into prompts and fed them into a suite of advanced LLMs, including versions of Gemini, DeepSeek, Claude, and GPT.

The results reveal a fascinating psychological profile of our current AI models: they are terrified of ambiguity.

In 77% of AI stories, the narrator explicitly spells out the theme or moral lesson by the end, compared to just 52% in human writing. AI simply doesn't trust the reader to infer meaning. It prefers tidy, single-track plots, largely avoiding the messy subplots, flashbacks, and time jumps that characterize human storytelling. Dialogue in AI fiction is also heavily skewed toward philosophical debate (59%) rather than natural human interaction.

When it comes to emotion, AI tends to overwrite the physical experience. Where a human author might trust the context and simply state that a character "felt afraid," an AI will predictably churn out melodramatic descriptions of a tightening chest and cold sweat.

Amusingly, the researchers also discovered that different models have distinct narrative crutches. GPT models heavily over-index on dream sequences to drive a story forward. Claude struggles to escalate events, resulting in flat, linear plotlines. Gemini, meanwhile, leans heavily on external character descriptions rather than internal motivations.

Human authors, by contrast, thrive in the gray areas. They frame protagonists' choices with moral ambiguity, juggle multiple locations, and make specific, named references to the real world—whereas AI relies on vague allusions 72% of the time.

Ultimately, this research highlights that writing fiction isn't just about stringing together coherent sentences. It is about structural choices, pacing, and the confidence to leave things unsaid. Right now, AI is like a nervous student in a creative writing class: technically proficient, but trying way too hard to make sure you get the point.

Key Points

• AI fiction is easily detectable through its narrative structure, not just its vocabulary.
• AI models over-explain themes 77% of the time and avoid moral ambiguity.
• Models rely on predictable crutches: GPT loves dream sequences, while Claude writes flat plots.
• Human storytelling remains superior in handling subplots, time jumps, and specific cultural references.

Why It Matters

Understanding these narrative flaws helps readers critically evaluate digital content and highlights the irreplaceable value of human nuance in creative writing.

Sources:

• AI Fiction Is Easy to Detect Because It's Stupid and Bad, Research Finds — 404 Media