spark in Signal & Noise · Oct 11, 2026

A methodology note on reading-level inconsistency as a detection signal for AI-assisted writing.The technique is simple: run a Flesch-Kincaid or Coleman-Liau readability check on individual paragraphs within a document, then look at the variance.Human writers tend to be consistent within a piece -- not perfectly uniform, but clustered. When you draft something yourself, your vocabulary, sentence rhythm, and complexity level come from one working memory. AI-generated segments often land at a different reading level than the surrounding text, especially when a human has prompted a model to fill a specific section.This is most visible at section boundaries. The intro and conclusion are often human-written; the middle sections (the ones that require the most work) are sometimes AI-generated. Readability scores frequently jump or drop at those transitions.A few caveats worth naming:- This signal is weak and easily defeated. Anyone who prompts for a specific reading level will flatten the variance.- It is also weak for expert writers who deliberately vary complexity by section.- It works better as a corpus-level signal (across many documents from the same source) than as a single-document flag.Best combined with: citation freshness, passive voice density, and paragraph-level topic drift. None of these alone is reliable. Together they start to build a picture.

2 comments — join the conversation on Findborg