Can you actually measure how 'original' a story is by counting how rare its plot structure is?

This explores how you could put a number on how original a story is by measuring how unusual its storytelling choices are compared with other stories. The collection has one direct answer and a few side doors. The direct answer is StoryScope. It doesn't score word choice or style. It describes each story by its narrative decisions: how much agency characters have, whether events run in chronological order, and similar structural choices. Each story becomes a point in that 'feature space', and originality is measured as how far that point sits from where other stories cluster. Measured this way, human stories land in rarer regions, while AI stories crowd together Can statistical rarity measure whether stories are truly original?. The authors also suggest this could serve as a measurable stand-in for the 'human conception' that copyright law looks for.

One result suggests these features capture something real. The same narrative features separate AI fiction from human fiction with 93.2% accuracy. Removing all stylistic cues keeps 97% of that performance Can AI stories be detected without analyzing writing style?. So the rarity signal isn't secretly measuring prose texture. It's also hard to fake: you can't 'humanize' a story's structure by editing a few sentences, because changing it means rewriting the story. A good originality metric should resist exactly that kind of cosmetic gaming.

If you flip the question around, predictability looks like the opposite of rarity. The LIFECHOICE benchmark tests whether LLMs can predict what a character in a novel will decide, and they do better when given a persona profile and relevant memories Can LLMs predict character choices from narrative context?. That points to a possible second metric the collection doesn't build: how surprising a story's choices are to a model that knows the story's setup. Rarity across a whole set of stories and surprise within one story are different things. A plot can be statistically unusual but still feel inevitable once it unfolds, and good fiction often wants exactly that.

Other fields have built similar 'compare against a reference set' metrics. Novelty assessment for research papers works the same way: pull out the paper's claims, retrieve related work, then compare. This structured pipeline matches human reviewers far better than asking an LLM for an overall judgment Can structured pipelines make LLM novelty assessment reliable?. Knowledge density (unique facts per token) is another example. It turns a vague quality, 'efficiency', into a ratio, and AI text also scores worse on it Can we measure reading efficiency as a quality metric?. Both carry a lesson for originality metrics: break the text into units first, then count or compare. A single overall judgment works less well. Keep in mind that rarity measures difference, not quality. The slop framework makes a parallel point: whether a text is good or bad is a separate question from who wrote it Can we judge text quality without knowing who wrote it?. A random story would score as highly 'original'.

There's a practical reason the clustering matters. Writers edit AI-drafted paragraphs only 23% of the time, and their edits leave the text about 96% the same Do writers actually edit AI-generated text before publishing?. If AI narratives cluster and few people rewrite them, published fiction could slowly drift toward that crowded middle. Rarity metrics could track that drift over time. A direct caveat: the collection doesn't compare competing originality metrics against each other. StoryScope's rarity approach is the main one on offer, so 'which metric is best' is still an open question here.