The structural fingerprint checklist

What the study actually found

StoryScope tested long-form fiction. Its full benchmark contained 61,608 stories, including human stories and stories from five AI models. Using narrative features rather than surface style, its classifier reached 93.2 macro-F1 on that closed benchmark.

The humanization experiment was narrower: 278 Gemini-generated stories were rewritten with LAMP to remove seven surface habits. The narrative detector moved from 95.5 to 93.9 macro-F1, a drop of 1.6 points.

That is not a literal watermark. It does not prove that an email, caption, essay, or individual document is AI-written. The useful finding is that surface cleanup may leave larger narrative decisions intact.

What “shape” means

The shape is not the vocabulary. It is the chain of decisions underneath the sentences.

An AI-shaped draft often moves in one straight line: every paragraph supports one thesis, the meaning is explained after it already landed, conflict resolves into a polished lesson, and details that do not serve the conclusion disappear. Human writing is more likely to keep a specific awkward detail, real tangent, contradiction, change of mind, or unresolved question.

The goal is not to manufacture randomness or add fake mistakes. It is to return the real choices and uncertainties that the surface rewrite cleaned away.

The structural fingerprint checklist

Score each item yes, no, or unsure. A “yes” is an editing prompt, not proof of authorship.

1. One perfect thesis: Does every paragraph point to the same neat conclusion?
2. Meaning explained twice: Does the draft state the point, then interpret the point again?
3. No genuine detours: Is every aside, tension, or counterexample immediately pulled back into the main argument?
4. Conflict resolves too cleanly: Does every problem end with a polished lesson?
5. Generic specificity: Are there lots of details that sound concrete but could belong to anybody?
6. Safe causal line: Does every event lead smoothly to the next with no uncertainty, contradiction, or unresolved edge?
7. Specific awkward detail removed: Did the “humanize” pass delete the odd example, imperfect sentence, or real constraint that made the idea yours?
8. Conclusion repeats the opening: Does the final paragraph mostly restate the thesis?

The three-question final pass

Before you publish, answer:

• Did every paragraph serve one perfect thesis?
• Did I explain the meaning twice?
• Did the rewrite remove the awkward, specific detail that made this mine?

If the answer is yes, rewrite at idea level before touching sentence style.

What an idea-level rewrite looks like

Surface-only version

This launch taught me that consistency, adaptability, and resilience are essential. Despite the challenges, the experience underscored the importance of trusting the process.

The words can change forever while the structure stays generic: tidy lesson, three abstract virtues, perfectly resolved conflict.

Structural rewrite

I almost cancelled the launch on Tuesday because the checkout broke twice. The fix was not inspiring. I removed the clever upsell, emailed 14 people manually, and kept the ugly page live. Three people bought. I still do not know whether the simpler offer or the direct email made the difference.

The second version keeps a real decision, an odd number, an unresolved uncertainty, and a detail that would be difficult to borrow.

A five-minute editing workflow

1. Map each paragraph’s job. Write one label beside it: claim, evidence, example, qualification, decision, or conclusion.
2. Break the perfect line. Add the strongest real counterexample, constraint, or uncertainty. Do not add randomness for decoration.
3. Return the human-only detail. Put back the specific moment, opinion, number, source, or consequence that only you can supply.
4. Cut repeated interpretation. Trust the reader after the evidence lands.
5. Polish sentences last. Sentence rhythm cannot rescue an idea that is still vague.

Sources

• StoryScope paper, especially Table 2, Figure 2, and §4.2: https://jenna-russell.github.io/assets/pdf/storyscope.pdf
• StoryScope code and data: https://github.com/jenna-russell/storyscope

Use this checklist as an editorial tool. Do not use it to accuse a writer or certify authorship.