Measured, not assumed

AI Slop check

Paste a draft and see whether its cadence has drifted toward the rhythm readers have started to associate with machine writing.

How the score is calculated

There is no black box here. The score starts at 100 and loses points as each measure moves outside the range ordinary human business prose occupies. Everything is counted per 100 words, so a long article and a short post are judged by the same standard rather than against a fixed count that quietly punishes length.

The em dash ratio carries most of the weight because it is the only one of the four with real separation. Ordinary human business prose runs roughly 0.1 to 0.3 em dashes per 100 words; machine-drafted prose commonly runs 2 to 3, which is ten to thirty times higher. The other three are softer signals, legitimate far more often, and weighted accordingly.

Bands: 85 and above reads as human-typical. 65 to 84 shows some drift. below 65 sits well outside the typical range. Under 150 words no score is given at all, because the ratios are noise at that length; one em dash in a 40-word note reads as 2.5 per 100 words and means nothing whatsoever.

What each measure is

Em dashes

The em dash is the long one, used to set off an aside. This is the single most reliable tell available, which is why it dominates the score.

Negations

Constructions of the shape "It is not X, it is Y" or "not just Y". The most overtrained rhetorical move in generated text. One good one carries weight; several stop meaning anything.

Fragments

Sentences of three words or fewer. One for emphasis is a technique. Six in a short piece is a fingerprint.

Buzzwords

Delve, leverage, seamless, robust, pivotal and about thirty more. Frequently already clean, which is worth knowing, because it means vocabulary was rarely the problem.

What this will not tell you

It counts; it does not judge. A low score does not make writing good, and a high one does not prove a machine wrote it; plenty of excellent writers use em dashes freely and always have. All this measures is whether a piece has drifted toward a cadence readers have started to distrust, which is a question about how the writing lands rather than a verdict on its quality or its origin.

It also does not look for AI watermarks. Some model providers do embed statistical signals in their output, but detecting those needs that provider's own detector, and editing or paraphrasing largely destroys the signal in any case.

Two false positives are worth expecting. The word harness fires on "test harness", which is an ordinary technical noun. And the negation count runs high on any piece whose argument is genuinely about what something is not, which is a good argument to keep rather than a flaw to fix. Judge the numbers; do not obey them.

Tell me what it should count next

I intend to keep adding to this, and the suggestions worth having come from people using it on real work. If there is something you would want it to measure, write and tell me: info@aggicorp.com. The subject line arrives prefilled and you are welcome to change it.

Licence and terms