This skill is based on published measurements of how LLM text differs from human text: Reinhart et al.’s comparison of Biber’s grammatical features in human and model writing, Terčon & Dobrovoljc’s survey of 44 studies, Clark, Martin & Mahowald’s work on why AI text is hard to read, and Graphite’s “AI Tells” study, whose released data includes tell lists for nine current models. It’s benchmarked against three corpora in which humans and models answered the same prompts: 9,984 web-article topics from Graphite, 24,322 questions from HC3 and 3,027 advice prompts from ELEPHANT.
It also ships with a scanner that runs before and after each revision. It shows which problems the revision fixed or newly caused, and it catches dropped facts and any tells from the reviser’s own model family.
pentaphobe•2h ago
rather than some amorphous "made by an LLM and I can tell"
That'd make "deslop" more of a "slop camouflage" since it just makes it less overt when a human has vibed the content instead of putting in any modicum of effort. :)
Still - impressive example in the readme, if we can't prevent slop then at least can reduce how much reviewer lifetime is wasted
[^1]: https://www.slopmanifesto.com/