StoryScope: why changing words is not enough to humanise AI content

StoryScope: why changing words is not enough to humanise AI content

A text can lose its most recognisable AI phrases while keeping the same underlying story. StoryScope asks a useful question: can researchers distinguish AI fiction from human fiction by looking at narrative construction rather than word choice? In their experiment, narrative decisions retain a strong signal. For teams trying to humanise AI content, the finding suggests reviewing what a text leads readers to understand, not just how its sentences sound.

That lesson needs a boundary. The study examines long-form fiction. It does not measure the quality of your blog, its performance in Google or the authorship of any text you put into a detector. Its editorial value lies in the mechanism it investigates.

What StoryScope actually measured

In version 6 of StoryScope: Investigating idiosyncrasies in AI fiction, Jenna Russell and her co-authors report a corpus of 61,608 stories built around 10,272 writing prompts: a human reference story and stories from five language models. The texts average roughly 5,000 words. The prompts were reconstructed from human fiction in Books3, rather than being fresh briefs given simultaneously to human writers and models.

The pipeline converts stories into structured descriptions and then interpretable features, covering character agency, event sequences, chronology and revelations, among other dimensions. Of 304 features in total, 257 are used for the narrative experiment after removing style-related features. A classifier learns to distinguish sources from those features, rather than from the raw prose.

The headline result is 93.2% macro-F1 for human-versus-AI classification on the test set. Macro-F1 summarises the balance between precision and recall while giving equal weight to each class. It is not a 93.2% probability that a particular text is AI-generated, nor a guaranteed success rate on arbitrary content. Prompts are kept separate across training and testing to prevent the same prompt appearing on both sides.

The signal goes deeper than familiar AI wording

The researchers find that AI stories more often explain their themes and spell out the lesson. They also favour single-track plots and tighter causal chains. Human stories in the corpus make greater use of chronological discontinuity and morally ambiguous choices. These are patterns within this dataset, not rules for judging every writer.

Changing a grand phrase into everyday language does not stop a narrator repeating a moral the story has already demonstrated. Shortening a sentence does not create a revelation that changes how an earlier scene should be read. Those revisions require changes to what the story contains and when it reveals it.

The paper also tests targeted editing with LAMP, a rewriting framework addressing artefacts including clichés, redundant exposition and purple prose. Across 278 Gemini stories, the narrative classifier’s macro-F1 moves from 95.5% before editing to 93.9% afterwards. That supports a specific conclusion: removing these prose artefacts did little to remove the measured narrative signal. It does not show that every human revision would fail, or that a tool can reliably establish the origin of a business article.

A business blog should not try to behave like literary fiction

A buyer’s guide does not need a nonlinear plot. A clear sequence and an explicit recommendation may be exactly what its reader needs. Adding digressions or withholding the answer to appear more human could make the article less useful. StoryScope offers no reason to make narrative complexity a marketing target.

The useful transfer is narrower: fluent prose can still carry an interchangeable argument. A draft introduces a problem, promises benefits, lists steps and concludes that success requires the right approach. Replacing its adjectives will not supply evidence, a serious objection or a decision tailored to the reader. This application to business writing is our editorial interpretation, not an outcome tested in the paper.

There are research limitations too. The findings concern particular model versions and human fiction drawn from Books3; language models also perform the feature extraction. The results do not validate detection for French blogs, short copy or mixed human–AI writing. They establish no Google ranking penalty for these narrative patterns.

Review the decision before editing the sentences

Rather than aiming to beat a detector, identify the decision the reader needs to make. Then inspect what the draft gives them to work with:

• The angle: which specific question does the article resolve? “AI improves productivity” leaves the choice open. “Which support replies can we automate without prior approval?” defines it.
• The evidence: does each important claim have a source, a scope and a limit? Invented examples must remain clearly labelled as illustrative.
• The objection: what would change the recommendation? If human approval is necessary, explain where and why instead of adding a vague warning at the end.
• The sequence: does each section add information that changes or refines the choice? Remove paragraphs that repeat the same promise under another heading.

Numbers need the same treatment. Turning an experimental score into a commercial promise breaks the argument, however natural the sentence sounds. Our five checks for numerical claims and their interpretation extend this review to presentations that combine text and data.

Consider an illustrative support article that recommends automating every enquiry to speed up responses. A wording pass can make the promise more appealing. A substantive revision separates routine questions from disputes and contractual commitments, then identifies who approves exceptions. The distinction helps the reader make a safer choice; a supposedly human vocabulary does not.

Humanising AI content means putting judgement into the argument

StoryScope makes a useful contribution by examining narrative decisions beyond the words on the page. For a business blog, the response is not to manufacture disorder or mistakes. It is to assign an editor responsibility for the choices a smooth draft can obscure: what needs proof, what should be cut and when the recommendation no longer applies.

AI can suggest wording and explore outlines. Before publication, someone needs to be able to defend the argument without relying on how confidently it reads. A final style edit can then do its proper job: make an already examined argument easier to follow.

Is your team using AI but still rewriting every draft? Say Digital can help scope an editorial workflow for one recurring content type: briefs, sources, objections, review criteria and approval. Tell us what your team needs to publish and which decision it should help readers make.

Sources and scope

Russell et al. — StoryScope, arXiv, version 6, 10 August 2026. Full paper: methods, Table 2 and sections 4.1–4.2. Research preprint consulted on 26 October 2026. The recommendations for business blogs are Say Digital’s editorial interpretation and were not evaluated in this study.