Slop
Every catalogue of AI writing tells is a pile of things somebody noticed. This one is sorted the way linguists sort language — by level, from the marks on the page up to the shape of the whole text — because a taxonomy you can check is a taxonomy that shows you its own holes.
Sorting by level does two things a word list cannot. It surfaces blind spots: a level with three entries is a gap in the catalogue rather than a fact about machines. And it separates the cheap tells from the durable ones. Graphemics and morphology get trained away inside a release cycle. Pragmatics and discourse do not, because fixing those requires having something to say.
We built it as evidence. Our argument is that models have been shaped into stilted speech, and this is that claim itemised.
In linguistics an asterisk marks an ill-formed example: *Colorless green ideas sleeps furiously. Every specimen below is starred, because every specimen is evidence. Entries cite as 4.12 — level four, entry twelve.
No tells match. Clear the search or the filters.
Method
Not a level of language. How you measure, and how much any single mark is worth.
The three tags sort by what detection costs. grep means a regular expression catches it. measure means you need a distribution across the whole document rather than a match inside it. read means a person has to judge, and no tool will do it for you.
What corpus linguistics adds
Four measures do most of the work and all four are distributional. Type–token ratio and hapax rate both run low: machine prose reuses its vocabulary and rarely reaches for a word it needs only once. Burstiness, the variance in sentence length, collapses toward zero, because human writers alternate a nine-word sentence with a forty-word one and models regress to the mean. Keyness compares a text against a reference corpus by log-likelihood and reports what is over-represented; run it and the word list in level six falls out of the data rather than out of somebody's memory.
What forensic linguistics adds
The discipline that already answers "how much does this feature indict" is authorship attribution, and its answer is: alone, almost nothing. Every feature has a base rate. Delve is ordinary in Nigerian and Indian English, and the 2024 wave of "delve means ChatGPT" posts was a room full of people learning to detect Nigerian annotators. The em dash is a mark of good writers. The tricolon is older than the Latin word for it.
What indicts is the likelihood ratio: how much more probable this bundle of features is under machine authorship than under human authorship, given the writer, the genre, and the year. Fourteen of these in four hundred words is a bundle. One of them is a Tuesday.
Two ways to be wrong
- Against non-native and second-dialect writers. Formal register, Latinate diction, careful hedging, and explicit connectives are what a good ESL education teaches. Detectors flag it, and every study that has looked has found the false-positive rate lands hardest there. If your read comes down to "too formal," you have found a language background rather than a language model.
- Against corporate writing, which was already like this. Consultancy decks, press releases, and university administrative prose have carried these features for forty years. Models learned them somewhere.
The half-life problem
Levels one through five are the cheap tier and will be trained out, some within a release cycle. The specific words in level six have started to go already. Levels seven and eight decay slowly, because a text with no deictic anchor, no stake, and nothing a reader could not predict is not badly generated. It is a text written by something with no situation to be in and nothing riding on the outcome, and that is harder to patch than a word list. It is also the part we think is an architecture problem, which is what we are building against.
Run the cheap tier
This runs a subset of the grep-tagged rules and reports density per thousand words. It cannot see levels seven and eight, which is where the tells that matter live. Treat a low score as evidence of nothing. Nothing leaves your browser.