AI tell · grade B
What is AI slop?
Slop is not a feeling here. underslop defines it as a number: the count of scored AI-writing tells in a piece of text, divided by its word count and multiplied by 100, the exact per100words figure our own linter returns on every draft before it ships. A tell only counts toward that number if it is grounded in our own model logs, corroborated by at least two independent style guides, or confirmed in our social-sentiment review, and not since refuted by that same review. Everything else gets logged for reference and left out of the score.
What it is
underslop's linter runs two passes over any text. The first counts SCORED tells: patterns either observed directly in our own 12-model logs, present in at least two of three legacy style guides we cross-checked, or confirmed by our social-sentiment review, provided that same review did not separately refute them. Those scored hits become the slop number. The second pass logs CONTESTED patterns for reference only: greylist words, single em dashes, curly quotes, title-case headings, adverb density, and a handful of once-popular tells our social-sentiment review found did not survive scrutiny. Contested hits never move the number. They exist so a writer can see what a checker over-flagged without paying a cost for it.
How often it shows up
We measured this on our own baseline set: 41 unedited AI drafts scored against the SCORED linter (research/eval/results/baseline/cells.jsonl). The average draft carried 1.11 scored tells per 100 words, and the median carried 0.79. Twelve of the 41 drafts scored a flat zero, clean by this measure straight out of the model. The messiest draft in the set hit 4.82 per 100 words. The spread runs from clean to dense inside the same dataset, which is the whole argument for measuring instead of eyeballing.
Why readers flag it
'Slop' online means whatever the person using it wants it to mean, which makes it useless for an editor trying to fix a specific draft. A number changes that. Two readers can disagree about whether a paragraph feels generated and still agree on how many scored tells it contains, because the count comes from a fixed list of patterns rather than a mood. That also cuts the other direction: human writing that leans on padded phrasing, vague sourcing, or a stack of bold-header bullets scores the same way a machine draft would. The measure doesn't care who typed the sentence.
The fix
Run a draft through the count before arguing about whether it 'feels' AI-written. In our own baseline set, half the raw drafts landed under 0.8 scored tells per 100 words, and the roughest landed near 5. Anything comfortably under 1 reads clean by this measure. Past 2 or 3, a paragraph-by-paragraph pass is worth the time, starting with whichever tell fires most often in the hit list the linter returns.
Examples
Great question! In today's fast-paced world, choosing the right project management tool isn't just a decision, it's a testament to how seriously your team takes execution. This platform stands as a testament to what's possible when you streamline your workflow, and it's not only affordable, it's a genuine game-changer for teams of any size.
Choosing the right project management tool says a lot about how a team approaches execution. This platform is affordable and works for teams of any size, and it can meaningfully improve the way you handle your workflow.
Here's the thing: research shows that a solid morning routine isn't just a nice-to-have, it's the secret to unlocking your full potential. Studies show early risers are more productive, and honestly, that's the real reason successful people wake up before dawn.
Morning routines get a lot of attention, and the research supports it. People who establish a consistent start to their day tend to be more productive, which is why so many successful people wake up early.
Questions
What is slop writing?
Writing padded with the surface patterns of AI generation: hollow authority claims, repeated rhetorical hinges, formatting used as decoration instead of structure. underslop measures it as a rate rather than a label, scored tells per 100 words, so 'slop' becomes something you can compute on a specific paragraph instead of argue about in the abstract.
How do I identify AI slop?
Run the text through a tell count and look at what actually fired. A single hit rarely means much. A cluster in one paragraph, three or four scored patterns inside a hundred words, is the real signal, and the hit list tells you exactly which pattern to fix first instead of leaving you to guess.
What is an example of AI content slop?
A paragraph that opens by praising the reader's question, reaches for a bold-header list to explain something that was one sentence of actual content, and closes by asking if you would like it to go deeper. Every piece of that is decoration. None of it is the answer.
How do I avoid AI slop in writing?
Write the plain version first: the claim, the number, the source, in that order, and skip the framing that dresses a sentence up before it says anything. Most scored tells are framing. Cut the setup and the score drops on its own.
What does 'slop' mean?
On this site specifically, it means a measured rate of scored AI-writing patterns per 100 words, not a synonym for 'AI-written.' Plenty of unedited model output scores zero by this measure, and plenty of human copy, especially marketing copy, scores high.
underslop marks this tell and the rest of the consensus set in any draft you paste, using the same open lint that scores it here. The free tier caps at 500 words a run, and your text is held for 0 seconds.