# Specificity Is What Makes a Sentence Attributable ## Claim A sentence is attributable to its author, rather than interchangeable with anyone else's, when it carries a specific number, mechanism, or small surprise — and this is the same property that makes writing read as human to statistical AI detectors, which score predictability rather than authorship. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for first-person, experience-grounded writing where the point is to convey something only the author observed or did. - Weaker for reference material, where interchangeability across authors is a feature rather than a defect — a documentation sentence that could have been written by anyone is often the correct sentence. - Does not claim specificity is sufficient for good writing, only that it is a cheap, operational test for one specific failure mode: genericness that reads as replaceable. ## Evidence R Dilip Kumar proposes reading a candidate idea aloud and asking "could someone else have written this exact sentence?" — "AI helps creators save time" fails (anyone could have written it); "I got back three hours a week by writing one boring setup file" passes (it has a number, a mechanism, and a small surprise unique to the author's situation). > [!quote] > Read it out loud and ask: could someone else have written this exact sentence? Separately, AI detectors work by scoring how statistically predictable each next word is — long runs of high-probability word sequences flag as machine-generated, regardless of who wrote them. A generic sentence and a high-probability sentence are, in practice, the same object described from two different angles: genericness is exactly what makes a sequence of words unsurprising to a probability model. Neither Kumar's article nor the AI-detector material makes this connection explicitly, but the mechanism is the same. ## Links - [[3. Resources/Permanent Notes/AI Editing Toward the Probable Is What Detection Detects]] — because extends; that note explains the detector-side mechanism (predictable word order flags as machine-made), and this note supplies the writing-craft-side diagnostic (the specificity test) that predicts the same property from the author's side, before any detector is involved. - [[3. Resources/Permanent Notes/Re-Deciding Pays the Full Cost of Deciding]] — because applies-in; the specificity test is cheap enough to run at idea-selection time, which is exactly the decision point that note identifies as costly when repeated per-session rather than batched. - [[3. Resources/MOCs/Personal Knowledge Management MOC]] — because applies-in; a usable filter for the vault's own Permanent-note claims, not just external content — could this claim have been written by anyone who read the source, or does it carry something specific? ## Tests / falsifiers - Score a sample of sentences on both the specificity test (human-judged: could someone else have written this?) and a statistical AI-detector's predictability score; the claim predicts strong correlation between "generic" judgments and "high-probability/flagged" scores. - Check whether sentences that pass the specificity test but are still AI-generated (a sophisticated model could, in principle, produce a specific-sounding fabricated detail) evade detection at a higher rate than generic AI output — this would test whether specificity is a genuine authorship signal or just a detector-evasion technique. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/The Prompt That Turns One Idea Into a Week of Posts]] — reports Kumar's specificity test for content ideas, read here against the AI-detector material's predictability mechanism.