Home › Glossary
What is AI slop?
Machine-flavored prose that reads fluent but says little. Here's the full taxonomy — and how to remove it without making it worse.
AI slop is writing that carries the statistical fingerprints of a language model: predictable phrases, structures, and rhythms that appear far more often in generated text than in considered human writing. It is grammatically clean and superficially confident, which is exactly what makes it hard to spot — the sentences are shaped like arguments without making one.
The tell is rarely a single word. It's an accumulation: a hedge to open, a listicle stem to organize, a dead transition to glue, a manufactured stake to inflate, a tidy summary to close — paragraphs that survive deletion without the reader losing anything. The test for slop isn't "does it sound polished?" It's "if I deleted this sentence, would anything be lost?"
The taxonomy of tells
Grouped by family. Each is a surface pattern a detector can flag as a candidate — the judgment of whether a real claim is buried underneath stays human.
Hedging & filler
- Empty hedging — avoids committing to the claim it's about to make. "It's worth noting that caching helps."
- Stacked hedges — two qualifiers where zero or one belong. "It might possibly be a memory leak."
- Vague quantifiers — a phrase standing in for a real number. "a wide variety of options"
- Filler intensifiers — adverbs as flavor, not measurement. "a truly elegant abstraction"
Structure & scaffolding
- Listicle stems — announce structure instead of making a point. "There are several key factors to consider."
- Dead transitions — "moreover / furthermore" gluing two non-claims.
- Wrap-up scaffolding — "in conclusion / the key takeaway is."
- Meta-commentary — narrating the writing instead of writing. "Let me walk you through…"
Performed voice
- Performed candor — fake intimacy to sound human. "Let's be honest, nobody reads docs."
- Assistant voice — chatbot sycophancy and email pleasantries. "Great question! I'd be happy to help."
- Manufactured stakes — borrowed urgency the content didn't earn. "In today's fast-paced world…"
- Rhetorical openers — engagement-bait with no information. "Ever wondered why your build is slow?"
Buzzwords & lexicon
- LLM lexicon — "delve, tapestry, realm, a testament to."
- Corporate uplift — "empowers, seamless, cutting-edge, best-in-class."
- Copula inflation — "boasts / serves as a testament" dodging plain "is / has."
- Business-jargon idioms — "circle back, move the needle, low-hanging fruit."
- Rule-of-three triplets — three buzz-adjectives in a row. "fast, reliable, and scalable"
The louder slop (over-correction)
The classic failure of "humanizer" tools is trading AI-slop for a noisier slop. Mechanically applying rules like "be punchy, drop every em-dash, go second-person, add a hot take" produces its own recognizable voice — the edgy-LinkedIn cadence. These are slop in a different costume, and a good detector catches them too:
- Binary-contrast cadence — "The answer isn't more servers. It's caching."
- Negative listing — "It wasn't the code. It wasn't the config. It was DNS."
- Corrective reveals — "You've been told X. Here's the truth: Y."
- Transformation chains — "Friction becomes flow. Flow becomes speed."
- Em-dash theatrics — dashes manufacturing emphasis the content doesn't earn.
When a humanizer cures timid slop by manufacturing this louder kind, that's over-correction — the two failure modes compared. This taxonomy folds in and extends the corpus from stop-slop (MIT) and its community pull requests — rendered as weighted, idiom-anchored rules rather than a blunt block-list, so honest technical prose stays silent.
How to remove AI slop (without faking it)
Two rules carry the whole job:
1. Fidelity over flair. Preserve the original meaning and claims exactly. Subtract the hedging and filler, surface the claim that was buried, and stop. Don't inject stance, edginess, or first-person the content never earned.
2. Flag hollow spans, don't fabricate. Some prose is weak because it has no point to make — rewording can't save it. Flag those for a human; never invent a hot take to fill the hole.
Before: "It's worth noting that, in many cases, caching can often lead to significant improvements in performance for a wide variety of applications."
After: "Caching improves performance for many applications."
The claim was always there. The fix was subtraction.
FAQ
- How do I make AI text sound less like AI?
- Cut the hedging and filler, state the buried claim plainly, and keep the meaning identical. Resist the urge to add forced opinions or dramatic punctuation — that's just louder slop.
- Is using more em-dashes a sign of AI?
- Dense em-dashes (several per sentence) manufacturing emphasis are a tell. But banning the em-dash outright is over-correction — a single, earned dash is fine. Context and density matter more than the character.
- Can a tool detect AI slop automatically?
- A detector can flag surface tells — the patterns above — as candidates. It cannot see hollowness (a clean-looking paragraph that makes no claim). That's why the right workflow flags rather than auto-rewrites, leaving the judgment call to a human or a careful model loop.