Tell glossary

The AI writing tells underslop detects

One concept per page, with the measurement behind it.

A tell is a structural pattern that shows up far more often in machine text than in human text. underslop scores a consensus set of them and reports a second, contested set it refuses to count. Each page below covers one tell: what it is, how often we measured it, why readers react to it, and how to fix it without flattening the sentence.

The glossary

Negative parallelism
The reframe that denies a plain reading to assert a grander one is the most-cited AI writing tell. What it is, how often models produce it, and how to fix it.
The "not only... but also" frame
The additive frame that stacks a concession onto a grander claim is a high-frequency AI tell. What it is, how often models use it, why it reads as generated, and the fix.
The em-dash chain
A single em dash is not an AI tell. A chain of them, several in a short span, is. What our logs show, why readers flag density, and how to fix it.
Assistant tells
Validation openers and engagement sign-offs are the clearest proof text came straight from a chatbot. What our logs show, why they stand out, and the fix.
The bold-header list
Turning prose into bullets that each start with a bold colon label is a high-frequency AI formatting tell. What our logs show and how to fix it.
Emoji as section decoration
Using an emoji to decorate a heading or bullet is a formatting tell. What our logs show, why it reads as generated, and how to fix it.
The throat-clearing pivot
The colon-label pivot that announces a revelation before delivering one is a tell grounded in our own model logs. What it is and how to fix it.
Vague attribution
Citing research or experts with no actual source is a recognized AI tell. What it is, why readers distrust it, and how to fix it.
The frequency-ranked checklist
Every AI-writing checklist lists the same dozen signs with no sense of which matter. Ours ranks them by frequency across twelve models in our own logs.
The three-slot response macro
In one of twelve models we logged, ChatGPT replies followed a three-slot macro that recurred across five unrelated topics, not any single sentence shape.
Lexical fixes vs. behavioral reflexes
Most advice for sounding less like ChatGPT targets vocabulary already absent from our model logs, while the near-universal reflexes go untouched.
Five sliders, not an adjective
Asking ChatGPT to write like you swaps a few words on the same shape. Running the same draft through five calibrated sliders shows what actually changes.
The em-dash density ceiling
A single em dash never trips underslop's scorer. What does: two or more inside the same 150-word stretch. Here is the number, and what a chain looks like.
Slop, defined as a density
'Slop' gets used as an insult with no fixed meaning. underslop defines it as a density: scored tells per 100 words, from the linter that grades every page.
The clipped one-line beat
AI-writing advice treats the one-line paragraph as a universal chatbot habit. Our own model logs show it in three of twelve models, not all twelve.
The graded checklist
Most 'AI writing style guide' pages are essays you can't check. This is the checklist underslop runs, every entry graded, some tracked but unscored.
'Let's dive in' and 'deep dive'
Two phrases get treated as the same red flag. One is a chatbot handoff we measured directly. The other is workplace speech our own research cleared.
The rule of three (tricolon)
Tricolon overuse showed up in a minority of models. Our social research found the pattern indefensible, so underslop tracks it but never scores it.
The bold colon lead-in: one tell, two forms
One model in our logs opened sentences with a bold colon label 35 times in 3,500 words. Same tell as bold-header-list, at its most concentrated.
Slush-pile sameness and the comp-title pitch
A novel has no headings, no bullet points, often no em-dash chains. The tells that survive are corpus-level: editors comparing hundreds of submissions.
The closed-connective stack ("in conclusion," "furthermore," "moreover")
Twelve models, zero uses of "in conclusion," "furthermore," or "moreover" as connective stacking. The top AI cliché mostly flags human essay writers.

Lint a draft against the whole set