An ongoing research project by Ryan Vet Back to ryanvet.com

The Ghostwriter

Method and sources

Each round deals ten passages, five written by people and five by a machine, drawn at random from a pool of 32. Because every round is balanced, guessing at random scores five, and that is the number each result is measured against. There is no clock. Each answer is quietly timed, because how fast you commit is how the result reads your confidence, but you can take as long as you like. Play again and the passages are ones you have not seen.

The human passages

All are verbatim, unedited text from works in the public domain in the United States — anything published in 1930 or earlier. Nothing here is quoted under a fair-use argument; these works belong to everybody. Each links to its full text so you can check us.

Passages were chosen from the middle of each work and deliberately are not the lines those books are famous for. A study that rewards having recognised the opening of a novel measures your reading history, not your ear.

The machine passages

All 16 were generated by Claude (Opus 5) from a short register brief and nothing else. No passage above — and no other copyrighted or public-domain text — was supplied to the model as an example to imitate. Each brief names a register and a tone, never an author:

One thing we chose not to do

Some of the human passages are by Frederick Douglass and W. E. B. Du Bois. We did not generate machine imitations of slave-narrative or early Black sociological prose to sit alongside them. Manufacturing counterfeit testimony of that kind as material for a guessing game is not something this study was willing to do. The machine half covers other registers of the same periods instead, and the deck stays balanced.

What this measures — and what it doesn't

It measures whether you can distinguish these particular passages, which are short, period-flavoured, and stripped of context. It does not measure whether you can spot AI in the writing you actually encounter — a work email, a report, a LinkedIn post. Modern functional prose is a different problem, and probably a harder one.

The result you are shown separates two things that usually get confused: how often you were right, and how often you were sure. Those turn out to be much less related than most people expect, and that gap is the finding this study is really after.

Participation is anonymous. Responses are used in aggregate only. Runs completed implausibly fast are excluded from the published averages. See the Privacy Policy.

← Take the study