Ten short passages. Half were written by people, half by a language model. Guess which is which. Every answer is cited — including the ones you get wrong, and including the ones we expected you to get wrong.
Every passage on the human side is taken word for word from a book that is in the public domain, and every one is cited with title, author, year and a link to the source text, so you can check that we did not quietly write it ourselves. Nothing is paraphrased. The full bank of all 30 passages, with the same citations, is human-or-ai.json.
Every passage on the model side was generated by a general-purpose language model in 2026 and is labelled as such. It is also labelled a second way: whether it was written in the model's default voice, or whether it was asked to sound like a person. That second label is the interesting one, and the results page splits your score along it.
Every human passage was written between 1901 and 1922, because that is what is freely available to quote. So "it sounds old" is a usable shortcut, and if you are winning on that shortcut you are not measuring what you think you are measuring. We would rather say this up front than let you walk away thinking you have a talent for spotting machines.
It does not tell you whether AI detection works. It does not tell you whether a given document was written by a model, and no score here transfers to that question. What it mostly measures is whether you have noticed the habits of one particular generation of models — tidy scaffolding, balanced clauses, a closing sentence that resolves everything. Those habits are already changing, and the "asked to sound like a person" half of the bank is there precisely because the obvious tells are the easiest thing to remove.
The whole question bank is inside this page. There is no backend, no scoreboard, no analytics, and no request after load — the counter under the game counts them live, and you can pull your network cable and reload to confirm. If you want to see how to check that claim on any program, not just this one, the four checks are written out here.
That is the same reason 3Q PDF works the way it does: if a document is not allowed to leave your machine, the tool that touches it should not be reaching for a network either. That one is not free. This one is.