A new study reveals why AI-written stories are easy to spot. Researchers say artificial intelligence struggles with complex plots and subtle themes. Human authors still outperform AI in narrative diversity.
Media professionals and publishers tracking the impact of AI on content quality now have new evidence that artificial intelligence still falls short in creative writing. According to a preprint study from University of Maryland, College Park and Google DeepMind, fiction generated by large language models is easy to identify due to its predictable structure and lack of narrative complexity.
The research team analyzed over 50,000 AI-generated short stories and found that these texts often over-explain themes, stick to straightforward plots, and avoid the moral ambiguity that characterizes human writing. The study reported that models like Claude, GPT, and Gemini each display distinct narrative quirks-Claude produces flat event escalation, GPT relies heavily on dream sequences, and Gemini defaults to describing characters externally. In contrast, human-authored stories show greater diversity in plot, character, and temporal structure.
Instead of focusing on surface-level stylistic markers such as em-dashes or repeated words, the researchers developed a tool called StoryScope to detect deeper narrative features. Built on the NarraBench benchmark, StoryScope evaluates plot development, character description, setting, and time structure to distinguish between human and AI fiction. Jenna Russell, a University of Maryland researcher and Pangram intern, said the goal was to move beyond text detection and identify structural differences in storytelling.
To test StoryScope, the team selected 10,272 human-written stories, converted them into prompts using Gemini 2.5, and then generated new stories with Gemini 3 Flash, DeepSeek V3.2, Claude Sonnet 4.6, Kimi K2.5, and GPT 5.4. The dataset, which includes both prompts and AI outputs, is available on Hugging Face. The original stories were sourced from the Books3 dataset, a collection of 183,000 books obtained from pirated ebooks. This dataset has been the subject of multiple lawsuits and is not publicly released by the researchers due to copyright concerns.
The study disclosed that AI-generated stories tend to spell out their themes-narrators explicitly state the story's lesson 77% of the time, compared to 52% for human writers. AI dialogue is more likely to serve philosophical debate, and references to other works are often vague. The systems also avoid subplots, rarely use time jumps or flashbacks, and overwrite sensory details. Human authors, by contrast, create more complex narratives with multiple characters, locations, and specific references.
The researchers also noted that AI tools were used to assist with coding and editing during the study, with all AI contributions reviewed and edited by humans. Russell emphasized the importance of disclosing AI involvement in academic work, noting that many researchers do not fully report their use of AI tools.
For those interested in the intersection of AI and content authenticity, these findings highlight ongoing challenges in distinguishing human creativity from machine output. The debate over AI's role in creative industries continues, especially as educators and readers seek to understand the true source of a story's originality. This echoes concerns raised in other areas of digital content, such as the recent launch of a browser extension designed to filter questionable brands on Amazon, as reported here.
