Machine-written prose has a measurable sentence-opener rate

The symptom

A manuscript passes every content-quality check, structure, argument, voice, and still reads as noticeably machine-touched, in a way nobody can quite name.

The cause

A specific, countable pattern sits underneath that feeling: a small set of sentence-opener phrases a language model reaches for to signal a turn in an argument, scaffolding a human writer's own prose rarely needs stated out loud. Measuring the rate of these openers per thousand words against known human and known machine-revised writing gives a real benchmark rather than a vague impression. Paul Graham's writing measures 0.000. A separately measured human-written book measures 0.000. A well-known business author measures 0.075. Prose that had been revised by a language model measures 0.464, roughly six times the human-authored books in this comparison.

The fix

Run the sentence-opener scan on a manuscript before it ships, and set a target rate below 0.08 per thousand words, built from the measured human benchmarks above rather than a guess. Treat a higher rate as a mechanical defect to fix directly, hunting down and rewriting the specific opener phrases, not as a vague voice problem to be waved at generally.

How to spot it

Search your own draft for openers like 'Here is,' 'Which brings me to,' and 'Now the,' counted per thousand words. A rate noticeably above roughly 0.08 is closer to machine-revised prose than to any of the measured human writers in this comparison, whatever the sentences individually read like.