Here's a sentence everyone recognises by now: in today's ever-evolving landscape, operators must leverage robust frameworks to seamlessly navigate regulatory complexity. There's nothing left to say about it. Every word in there sits on every banned list going, and the diagnosis has been written a thousand times.
So here's what the standard correction does to it. Operators must comply. Frameworks help. Complexity is real. The banned words have gone. So has the em dash, and the average sentence is now four words long.
And the second one is worse. Not differently flawed. Worse. The first sounds like a person with nothing to say, which is familiar and survivable. The second sounds like nobody at all.
The Cure Has Its Own Accent
Every de-AI checklist going only takes things away. Cut the em dash. Cut the transitional adverb. Cut the three-item list, shorten the sentence, delete the hedge. Each of those is defensible on its own, every one appears somewhere in my own standard, and I'd argue for all of them individually with anyone who fancied the argument.
Run them together at full strength and they don't hand the prose back to a human. They move it to a second artificial voice, one new enough that almost nobody is screening for it yet, which means an editor who can spot the fluent machine version at forty paces will read the clipped machine version, find nothing on the checklist to object to, and wave it through.
A tell is not a word. It is the absence of somebody deciding, and that absence can be manufactured in either direction.
My own documents carry a section on this because it happened to me, not to somebody I read about. A draft went through the full pass and came back with a one-line verdict. It sounded like "an army major, or a report", starting sentences with no natural way in. No model did that. A rule applied without judgement did that.
The Symptoms Are Specific
Which is the useful half of the problem, because a symptom you can name is a symptom you can go looking for. Five of them account for nearly everything I've had to undo, and all five show up on a first read once you know what you're reading for.
- Fragments in the majority. One per section is rhythm. Three in a paragraph is a telegram. Rejoin two and the thought comes back.
- Articles stripped from openers. A page that says Welcome bonus is £100 has had a word taken out for no reason anyone can defend. Put it back.
- Connectives deleted as padding. They weren't padding. A connective is the bit that shows one thought producing the next, and prose without them reads as a list of assertions that happen to be sitting next to each other.
- Every paragraph the same length. Uniform short is exactly as templated as uniform long, and it fails the same underlying test.
- A contraction ratio at either extreme. Nothing contracted reads stiff. Everything contracted reads glib. The number matters less than whether somebody chose a voice.
The Metrics Push You Into It
This happens to careful people rather than careless ones, because the measurements reward it. Average sentence length has a healthy band, and the fastest way to pull a number down into it is to chop. Fragment density has a floor, and the fastest way to clear a floor is to drop fragments in.
Each move satisfies the check and damages the prose, which is what any proxy measure does the moment you optimise against it directly. It's exactly why the bands in my standard are written down as a floor to clear rather than a score to chase. The numbers were built to catch an absence of variation. They can't supply the presence of judgement, and no counter alive can tell whether a fragment landed or got dropped in to move the counter.
The Case in My Own Files
There's a documented example of this sitting in my own standard, and it ran the other way from the one you'd expect. The file carried a contraction target of fifty to ninety-five per cent, inherited from a firearms buying guide. Measured against the essays here, that band looked wrong. The pieces on this site ran at zero per cent, they read fine to me, so the rule got rewritten to say zero was the house voice and the next writer was told not to fix it.
That held for about a week. Put in front of the person who owns the site, the same prose came back as "very stuffy and too high brow", and the fix was the band I'd overruled. The lesson isn't that the metric was right all along. It's that I'd taken the state of the archive as evidence of a decision when nobody had ever made one, then written a rule that protected it from repair. A voice can't be validated from inside the voice.
The Only Test That Catches It
Read the thing out loud and the problem announces itself inside a paragraph. That's the whole method, and it's faintly embarrassing how far it beats the tooling built around it.
Say a sentence out loud and you hear the missing article straight away, in a way no script I've written ever manages. You hear the fourth fragment in a row. You hear the paragraph that got shortened past the point where the thought still survives, and you hear the difference between a short sentence somebody wrote short on purpose and a long one that's been amputated to bring an average down.
No Score May Dictate a Fact
Above every other rule in my standard sits one line: no detector, no score and no word count gets to decide what a sentence says. That protection has to cover rhythm as well as content. If the accurate version of a sentence needs thirty-eight words then it gets thirty-eight words, the average moves, and the check is the thing that accommodates the sentence rather than the other way round.
All of which sits downstream of a simpler point I've made about slop itself. The tells everyone is hunting are symptoms of nobody having made a decision about this particular sentence, and deleting the offending word doesn't put a decision back in. What deletion does is remove the evidence that one was missing, and those two jobs are nothing like the same repair.