AI Slop Checker

The scorer starts every post at 100 and takes points away. Thirty-odd checks run against the text: twelve hard checks worth 25 points each, eighteen soft checks worth 10 each, plus one statistical test on sentence rhythm. What is left is the score. Nothing is sent anywhere to compute it, because none of it needs a model. Every check is a rule.

The three bands are the same ones VoicePost uses on its own drafts. Seventy and above is a hard pass, the draft ships. Fifty five to sixty nine needs editing. Below fifty five goes back for a rewrite. Those thresholds were not picked to be flattering: they were tuned against a corpus of twenty posts known to be good and twenty known to be bad, and the engine has a test suite that fails the build if a change moves either group across its line.

The three families of tell

Vocabulary tells. A short list of words and phrases that models reach for constantly and people almost never do. One of them in a post is a coincidence. Three is a signature. The full list is printed below, because it is the single most useful thing on this page to copy into your own editing checklist.

Structure tells. Uniform sentence length is the biggest one. Models produce sentences that all land within a few words of each other, while human writing lurches: a long qualifying clause, then four words. The checker measures the standard deviation of your sentence lengths and flags anything too even. Stacked line breaks, thread numbering on a standalone post, hashtags, and raw links live in the same family.

Stance tells. This is the subtle one. Models hedge (“I think”, “perhaps”, “might”), prefer the passive voice, avoid specific numbers, and end on a question to seem engaging. Individually, each is forgivable. Together they read as text written by something with no stake in the claim, which is exactly what happened.

The 22 phrases that scream AI

This is the kill list the scorer matches against, rendered straight from the engine so it can never drift from what is actually checked. Each one is matched case-insensitively anywhere in the text. Hitting any of them costs 10 points once, not per phrase, so the penalty is a nudge rather than a punishment. Take it as an editing checklist:

• delve
• leverage
• seamless
• game-changing
• innovative
• groundbreaking
• cutting-edge
• it's worth noting
• in conclusion
• furthermore
• moreover
• in today's
• in the realm of
• I am thrilled
• I am excited to
• Let's dive in
• comprehensive
• robust
• scalable
• synergy
• paradigm shift
• unlock potential

A few deserve a note. “Leverage” and “robust” are real words with real uses, and they still cost you, because on social platforms they read as consultancy register rather than a person talking. “Delve” is the most-quoted example of the whole category, to the point that seeing it now reads as a draft nobody edited. And if a post opens with “I am thrilled” or “I am excited to”, the rest of the post rarely recovers.

What the hard checks catch

Hard checks are the ones that get a post clocked in under a second of scroll time, so they cost 25 points each: an em dash, more than one emoji, a hashtag, a raw link in the body, a “1/” thread marker on a standalone post, more than six line breaks, a hard-sell opener like “I built”, and sentence rhythm too uniform to be human. Two of these together drop a post to 50 no matter how good the idea is.

What the soft checks catch

Soft checks cost 10 each and describe texture rather than violations: no number or specific claim anywhere, kill-list vocabulary, ending on a question, hedging, a weak opening word, passive voice, corporate buzzwords, more than five numbered list items, no concrete example or named thing, a hedged final sentence, thread bait, and exclamation-mark pileups. You can fail two or three soft checks and still be well inside the green band. Failing seven means the post says nothing specific.

Glossary

• AI slop. Text that is fluent, correct, and empty. It is characterised by generic vocabulary, uniform rhythm, hedged claims, and an absence of specifics, and it is recognisable without knowing who wrote it.
• Burstiness. The variation in sentence length across a piece of writing, measured here as the standard deviation of per-sentence word counts. Human writing is bursty. Model output tends to be flat. The checker wants a deviation of at least 3.5 once a post has four or more sentences.
• Hard check. A rule whose violation is visible at a glance and effectively disqualifying on its own. Worth 25 points. Twelve of them exist.
• Soft check. A rule whose violation is a hint rather than proof. Worth 10 points. Eighteen exist, and several only apply to specific writing styles.
• Distribution Score. VoicePost's internal name for this 0 to 100 number. It is stored on every draft the product generates, and it is what the pass bands above refer to.

What the score cannot do

It cannot tell you who wrote something. Nothing can, reliably, on a post-length piece of text, and tools that claim a percentage confidence on 280 characters are selling you a number with no information in it. This scorer is deliberately the other thing: a list of named, fixable patterns. If you disagree with a check, the right response is to ignore it, which is why every failed check tells you exactly what it matched.

It also cannot tell you whether an idea is worth posting. A perfectly human post about nothing still gets nothing. What it removes is the specific failure where a good observation gets scrolled past because the phrasing pattern-matched to generated text. If your posts are getting no traction at all, the phrasing is only one of four common causes, and the diagnosis guide walks the others.