Stylometry

The premise: writers have unconscious habits — how often they use of, upon, the, how their sentence lengths rise and fall, where the commas land — and those habits are stable enough to form a fingerprint that is hard to fake. The classic demonstration is Mosteller and Wallace's 1960s analysis of the disputed Federalist Papers, which attributed the contested essays to James Madison largely on the frequencies of function words. Function words carry the signal precisely because no one chooses them consciously.

Stylometry's traditional applications are authorship attribution (who wrote this?), forensic linguistics, and plagiarism investigation. AI detectors are its newest and crudest descendant: instead of asking "which of these candidate authors wrote this text," they ask a one-bit question — human or machine — using aggregate statistics such as perplexity and burstiness. Same family of features, much blunter instrument.

The limitation is inherited too: stylometry outputs probability, not proof. That is workable when scholars weigh an attribution over decades; it fails when a probability is treated as a verdict against a student — especially since careful non-native prose statistically overlaps with machine output. Style analysis judges the finished artifact. Process evidence takes the opposite route: an Authorship Certificate records how the document was actually written, edit by edit, so authorship is checkable rather than inferred from style.