TL;DR

  • A University of Maryland and Google DeepMind team (StoryScope, arXiv 2604.03136) separated human fiction from AI fiction using only how each story was built: plot shape, character agency, how time moves. The authors set out to do it without relying on stylistic signals, and they report that structure alone kept over 97 percent of the performance of models that also saw style.
  • The patterns are specific. AI stories over-explain their themes and favor tidy, single-track plots. Human stories frame the protagonist’s choices as more morally ambiguous and move through time in more complicated ways. And AI stories cluster in one shared region of the feature space, while human stories spread out.
  • If it holds outside fiction (the paper does not test that), it is a structural explanation for why AI content reads generic, and for why a word-level cleanup (swap the giveaway vocabulary, vary the sentence length) would not remove the sameness on its own.
  • Scope, plainly: this is a study of fiction, roughly 5,000-word stories written from writing prompts. Its 93.2 percent figure belongs to that corpus and that classifier. It is not a rate at which anyone can detect AI in a blog post, and it says nothing about Google.
  • The non-fiction checks we apply in our own editing (a thesis stated three times, no counter-position, every section resolved) are our extrapolation, not the paper’s finding. They are cheap to run, which is the reason to run them.

Ask a business owner what is wrong with the AI-written posts on their site and the answer is usually some version of “they read like everyone else’s.” That is a fair description, and until recently it was mostly a matter of taste. A paper published in April 2026 makes part of it measurable, and the measurable part is not the part most people try to fix.

Most efforts to make AI content read less generic work on the words: banned-vocabulary lists, punctuation rules, rewritten sentence openings. The StoryScope study asked what is left if you throw that whole layer away. Its answer, on fiction, is that the machine still gives itself away through how the text is put together.

What did the StoryScope study measure?

The team (Jenna Russell, Rishanth Rajendhran, Chau Minh Pham and Mohit Iyyer at the University of Maryland, with John Wieting at Google DeepMind) built a pipeline that describes a story through 304 discourse-level features across ten dimensions: agents, plot, events, setting, temporal structure, perspective and so on. They ran it on a parallel corpus of 10,272 writing prompts, each answered by a human writer and by five large language models, for 61,608 stories of roughly 5,000 words each. Because every author wrote to the same prompt, topic was held constant across the six versions of each story.

Without relying on stylistic signals, the narrative features alone reached 93.2 percent macro-F1 (the unweighted average of the F1 score across classes) on human versus AI, and 68.4 percent on the harder question of which of the six authors wrote a given story. The authors report that this retains over 97 percent of the performance of models that also see stylistic cues, and that a compact set of 30 core features carries much of the signal, which, on our reading, suggests the result does not depend on a very wide feature space.

Every figure in this section is the authors’ own claim, read from the paper’s abstract and its public code repository. We did not work through the full PDF, so the method detail and the authors’ own stated limitations are not reflected here. Treat this as a report of what the paper says, not an independent check of its evidence.

Why does AI-written content read generic?

On fiction, the StoryScope authors’ answer is structural: AI stories over-explain their themes and favor tidy, single-track plots, while human stories frame the protagonist’s choices as more morally ambiguous and show more temporal complexity. Whether the same holds for business writing is not something the paper tested.

The second result matters more for a business. AI-generated stories cluster in a shared region of narrative space, while human-authored stories show greater diversity. Five different models, and their stories still landed in one region while the human stories spread out. On this corpus, sameness is not only an impression readers have of machine text. It shows up as a measurable property of how the stories are built.

The models are not identical to each other either. The six-way attribution works because each one has habits of its own: per the authors, Claude produces notably flat event escalation, GPT over-indexes on dream sequences, and Gemini defaults to describing characters from the outside. “AI-sounding” is plural, and an editor who has learned one model’s tells can miss another’s.

We have seen convergence from another angle. In our post on why AI keeps citing the same sources, eleven models from three vendors agreed with each other about which sources to prefer. The task and the paper are different, but the picture is similar: models from different companies behaving more alike than you would expect. Neither study explains the other, and we are not presenting one as proof of the other.

Why doesn’t a word-level cleanup fix generic AI content?

Our reading, and it is an extrapolation from fiction: a cleanup that only changes words works on the layer the StoryScope classifier did not need. This is where the paper becomes useful to someone buying or producing content. Most tools and services that promise to make AI text “sound human” work at the sentence level. They replace the vocabulary that has become a giveaway, break up the rhythm, remove the punctuation habits. Our own de-AI editing pass started there too, as a checklist of a few dozen word-level and phrase-level tells.

That work is not wasted: those tells are real and readers notice them. But if the structural signal survives the removal of style, as it did on this fiction corpus, then a draft can pass every word-level check and still be built like every other AI draft. The vocabulary would change and the shape would stay. The citation study linked above hit the same limit from the other side: paraphrasing the candidate sources until almost none of the original wording remained left the models’ preferences practically unchanged. In both studies, the result held with the wording taken out of the picture. One clarification, since that post advises against paying for restructuring: “structure” here means how an argument is built (what it commits to, concedes and leaves open), not headings, markup or layout.

What should an editor check in a non-fiction draft? (Our extrapolation)

The paper tests fiction only, and most of its feature space has no counterpart in a service page or a blog post. Of the 304 features, 265 sit in dimensions such as agents, social networks, plot, setting and revelation, which only exist inside a story. Only one dimension, labelled Style (39 features), describes something any text has, and the abstract does not say how it relates to the stylistic cues the headline result leaves out, so we do not lean on it. What follows is how we translate “over-explains its themes, favors tidy single-track plots” into four checks for business writing: a thesis stated, restated and concluded again; one track, with no counter-position allowed to have a point; every section resolved, with no tradeoff left standing; and the same weight everywhere. We use them because they are cheap and they match what experienced editors already say about AI drafts, not because the paper proves them for blog posts.

The human side of the finding has a non-fiction version too: a writer who commits to a position, says what it costs, and leaves something open because it really is open. That is harder to get from a prompt than vocabulary is, and it is usually where the expertise shows.

We added these four checks to our own editing in July 2026, as a document-level pass that runs above the word-level checklist, labelled there as an extrapolation from a fiction study.

What the study does not say

  • It is not an AI-detection rate. The 93.2 percent belongs to one classifier on one corpus: one pool of human writers, five models, one genre. It does not mean AI content is 93 percent detectable, and it says nothing about whether a reader, a tool or a search engine can spot AI in your blog post.
  • It says nothing about Google. The paper does not touch search. Google’s published guidance on AI-generated content says it focuses on the quality of content rather than how it was produced, while still treating content generated mainly to manipulate rankings as spam. Generic content has costs, but “Google detects the structure and penalizes it” is not something this paper shows.
  • It is not established for non-fiction. The four checks above are our translation, not a tested result.

What this means if you buy or produce content

If an agency or a freelancer writes your content with AI help (most do, and that alone is not the problem), the useful question is not “is it AI?” It is “what does your edit change besides the words?” A process that ends with a vocabulary sweep produces copy that passes a quick read and still looks like your competitors’ copy, because it was built the same way.

What a model cannot supply is the material that would pull a draft out of the cluster: what you have actually seen on the job, the position you would defend in front of a client, the tradeoff you would admit to. The four checks go looking for that material, and when it is missing, no rewrite will put it there.

Our SEO audit sprint reads existing content for this alongside the technical and local work, and it is the standard we hold our own GEO work to. To try it yourself first, take your last three posts and run the four checks on them. If all three are built the same way, start there rather than with the wording.

Fixing the words is the easy part. The one study behind this post is about fiction, and it found the machine still recognizable with the style set aside. How much of that carries over to a service page is an open question, which is why the four checks above are labelled as our bet and not as a finding.