Paragraph over-fragmentation
- Id
- paragraph-fragmentation
- Status
- Contested
- Severity
- low
- Detection
- statistical
- Evidence grade
- community-observed
- Languages
- en
- Added
- 2026-08-15
- Updated
- 2026-08-15
Credible people dispute that this signals anything at all.
What it is
A body built almost entirely from paragraphs of one or two sentences. Pudasaini and colleagues took apart what detector features were actually keying on and found paragraph count dominating feature importance on one corpus and near-absent on another, which means those models had learned the corpus rather than the writer. On the corpus where it dominated, the association the classifier had learned ran the opposite way from this entry: single-paragraph texts read as machine-made, at a median paragraph count of 1 for the machine text it caught and 17 for the human text it cleared. The human text it flagged wrongly had a median of 1 as well. Paragraph shape was a corpus artifact there before it was a signal. That is why this entry ships without a number.
Why it reads as machine-made
The shape comes from a model writing for a scroll, and from a person pasting model output into a post box without joining anything back up. It also comes from every mobile-first style guide written since 2015. Alone it means very little. It earns its place because it travels with the other structural tells and because it makes a piece look longer than the thinking behind it.
Specimens
We shipped the new onboarding last week. It went well. The team learned a lot from the process. Everyone contributed. There is more to do here. We will keep going.
We shipped the new onboarding last week: five screens became two, and email verification now runs in the background instead of blocking the first login. Completion went from 61 to 78 percent over the following six days, the first movement in that number since February. The import step is next, and it still loses about a third of the people who reach it.
Six short paragraphs of nothing became one that carries two figures, a date and the next problem. Figures in this repair are invented for the specimen.
How it is detected
- Metric
- share-of-paragraphs-under-three-sentences
- Threshold
- unset
- Direction
- above
- Threshold basis
- No threshold ships. The error-analysis section of the Pudasaini paper found detectors keying on paragraph shape had learned the corpus they were trained on, with paragraph count carrying the largest gap in feature contribution between the two error types. The human text that analysis flagged wrongly also ran short, averaging 221 words against 421 for the human text the same model cleared. A cutoff here would inherit both failures. Report the share as context beside the other structural measures, as a FeedSquad reading cue and nothing more.
Who writes this way legitimately
Mobile-first editorial style guides mandate this shape, and newsrooms writing for phones have followed them for a decade. The broetry format on LinkedIn is a documented human invention that predates chat assistants by about five years, and it is written that way on purpose by people who know exactly what they are doing. Published error analysis shows detectors that keyed on paragraph shape were learning their training corpus, and the human writing they flagged wrongly had the same paragraph shape as the machine text they caught, so a hit here is a reading cue and never a conclusion.
Model attribution
No study attributes paragraph shape to a model family. Chat surfaces shape it more than models do: formatting defaults differ between a chat product and the same model called through an API.
Platform notes
- The one-sentence-paragraph post format on LinkedIn predates chat assistants and remains in wide human use. No published LinkedIn policy addresses line breaks or paragraph length.
Status history
| Date | Status | Rationale |
|---|---|---|
| 2026-08-15 | Contested | Ships contested on the strength of the published error analysis, which shows this exact feature behaving as a corpus artifact. Paragraph count carried the largest gap in feature contribution between the classifier's two error types, and the human writing it misfired on had the same paragraph shape as the machine text it caught. |
Sources
- 01
- 02vale-ai-tells, 111 machine-checkable rulescommunityaccessed 2026-08-14
- 03anti-slop-writing system prompt and pattern listcommunityaccessed 2026-08-14
Inverted source. Its own purpose runs counter to this index. It is cited for what it records about model output, and never as sole support for an entry.
CC BY 4.0 / The AI Tells Index, feedsquad.com/ai-tells