I measured my own writing for AI slop

On July 30, LinkedIn added a button that lets any reader report a post as "Seems like AI slop." Three weeks later it had been pressed more than a million times. On August 21, LinkedIn's Chief Product Officer Hari Srinivasan said that posts the platform classifies as slop are now getting 40% fewer views.

Not removed. Just quietly shown to fewer people.

I use Claude to help me work through my ideas, so that number applies to me. And I did the thing I spend my working life telling other people to do. I measured.

I took thirteen pieces I published this year; posts, articles and three newsletter issues, about 4,500 words. Then I wrote a small counter for the things that are supposed to give machine writing away.

The vocabulary came back clean. The buzzword scan covered delve, leverage, seamless, robust, pivotal and about thirty more, and across all 4,500 words it found exactly one hit. That hit was "test harness", which is a real technical term. The problem was somewhere else.

It was the punctuation. Ordinary human business prose runs about 0.1 to 0.3 em dashes per 100 words; the em dash is the long one, the kind used to set off an aside. Eleven of my thirteen pieces ran 1.6 to 3.3. Ten to thirty times normal. The two that came back clean were the two shortest, written quickly as conversation starters, and I am still thinking about what that means.

But the count is the small part. I had a written rule against em dashes. I have had it for months, and I restate it at the start of every drafting session. The rule was in place, the drift happened anyway, gradually, across thirteen pieces, and I only saw it when I counted.

That is the argument I make for a living, showing up at my own desk.

We built Aggi ARIA (Aggi Responsible Intelligence Assessor), a continuous AI behavioral compliance platform. It exists because a written policy is not evidence. A model that passed review in March can behave differently in September and nobody notices, because everyone is trusting the document instead of testing the behavior. My own writing drifted for that same reason. I trusted the rule I had written.

I want to be careful about what I am claiming. A low em dash count is no guarantee of good writing, and plenty of fine writers love the em dash. Counting only tells you whether your prose has drifted toward a cadence that readers are now primed to distrust. That is a distribution question rather than a quality one.

I have a horse in this race; drift detection, meaning noticing when a system has quietly stopped behaving the way it did when someone approved it, is what my company sells.

Sources: ProPakistani, July 31, 2026 (the report button); TechSpot (one million reports); Engadget (Hari Srinivasan, 40% fewer views). The em dash counts are our own measurement of the thirteen pieces.