Can AI stories be detected without analyzing writing style?

Most AI-text detection rides on surface signatures: word choice, syntactic structure, the overused em-dash, "delve," "tapestry." These cues are discriminatory but fragile — GPT 5.4 cut em-dash usage, and fine-tuning to mimic human style drops detection on creative writing from 97% to 3%. StoryScope asks a different question: can AI stories be told apart without stylistic signals, using only discourse-level narrative choices like character agency and chronological structure? Across a parallel corpus of 10,272 prompts (each written by a human and five LLMs, 61,608 stories of ~5,000 words), narrative features alone reach 93.2% macro-F1 for human-vs-AI detection, retaining over 97% of the performance of models that include stylistic cues.

The consequential part is the durability argument. Surface style is a post-hoc edit away from concealment; discourse-level narrative structure is not. Changing whether a protagonist's choices are morally ambiguous, or whether a plot runs on a single tidy track versus a nonlinear one with flashbacks, requires structural rewrites rather than find-and-replace. So the features that survive humanization are precisely the ones tied to how a story is conceived, not how its sentences are dressed.

Why it matters: this reframes AI detection from a stylometric arms race into a structural one, and it relocates the question of authorship. If models keep closing the surface-style gap while their narrative choices stay distinct, then detection — and, downstream, the legal question of originality — should attach to discourse structure. The counterpoint is that narrative features are themselves learnable targets; nothing prevents future training from diversifying discourse-level choices, which would erode this signal too, just more slowly than style erodes.