What this test does and does not claim
Both texts in every pair were written by a language model. One of them is written the way such a model writes when nothing holds it back; the other was written to avoid exactly that. So this is not a test of authorship — nobody here is guessing who typed what. It is a test of whether you can hear a particular handwriting.
That distinction matters, because the honest version of the "human or machine" question is much harder than it looks. There is no reliable detector, the giveaways below are habits rather than laws, and a model asked to write differently writes differently.
The habits worth knowing
- Triples: three virtues, three tips, three causes — comma, comma, and. Reality rarely arrives in threes.
- A hedge before every claim: "generally", "as a rule", "experts advise" — with no expert ever named.
- A closing sentence that restates the paragraph instead of adding to it, often starting with "ultimately" or "therefore".
- No numbers, no names, no mishaps. Human text is full of accidental detail: a broken drain, a four-pound alarm clock, a knee that hurts on week three.
Why detectors do not save you
Automatic "AI detectors" measure how predictable a text is. That punishes anyone who writes plainly — non-native speakers, technical writers, students taught to write in a formal register — and it is defeated by asking the model for a rougher style.
Which is the uncomfortable part of this whole page: the tells you have just learned are properties of a default setting, not of machines. Change the setting and the handwriting goes away.