Scientists from the University of Maryland and Google DeepMind have discovered that fictional texts created by neural networks possess fundamental defects in plot construction. Thanks to a new tool called StoryScope, it has become possible to identify AI authors not through statistical text features, but through deep violations of narrative logic.
What Happened
As part of a new study, it was proven that modern models, such as Claude, GPT, Gemini, and DeepSeek, use simplified and formulaic structures when creating fictional content. Using the StoryScope tool, scientists identified systemic errors: excessive moralizing, replacing action with philosophical dialogues, and the use of redundant physical clichés instead of deep emotional character development. During the experiment, model texts were compared with the works of professional authors, including Stephen King.
Context
Traditional methods of detecting AI content usually rely on stylistic analysis, using metrics such as perplexity and burstiness. However, this study shifts the focus to semantic-structural analysis, proving that violations of dramaturgy and narrative logic are the key markers of synthetic text.
Why It Matters for the Industry
For the industry, this signifies a paradigm shift in content evaluation and filtering methods. Developers will have to move from simple local word statistics analysis to creating new evaluation methods (evals) focused on structural integrity and plot complexity. This also opens a niche for specialized monitoring and authorship verification tools within observability pipelines.
Why It Matters for Users
For readers and authors, the study confirms that modern LLMs are not yet capable of genuine creativity and complex dramaturgy. This highlights the continuing value of the "human heart" in literature and makes deep emotional development an important quality criterion that serves as a defense of authorship against full automation.
What Is Not Yet Known / Limitations
No direct technical disagreements regarding the essence of the study were identified, although expert positions vary from neutrally descriptive to skeptical in the context of current model architectural limitations.
Sources
Author
Look at AI, Editorial Staff