Slop Filter

It asks 30 narrow questions about how the text is written, not what it says, with human tells first such as invented words, jokes, stretched words, and typos. A confident human tell settles it as human before AI cadence tells are weighed.

By script only, never title or thumbnail. The same scoring applies once the script text is read, and the extension combines the per-question probabilities into one score with tunable weights.

Three options on the options page: your TypeSafe key, your OpenRouter key billed to your credits, or a hosted monthly plan where scoring goes through the project's server. With your own key the text goes only to the provider you chose.

Mark posts as AI or Human, then fit the weights from your labels with the included script. Blocking is separate: after three flagged posts it offers to block, and blocked accounts are hidden without further scoring.

The documented path is to paste a prompt into Claude Code, Codex or any coding agent pointing at the repository's SKILL.md; the agent clones the repo and walks you through the clicks Chrome requires, and your key goes into the extension's options page rather than the chat. By hand: clone, then for Chrome open chrome://extensions, enable Developer mode and Load unpacked; for Firefox run node scripts/zip.mjs firefox, extract the archive, open about:debugging and Load Temporary Add-on against the extracted root manifest.json.

Your own TypeSafe key is free to set up, and the README puts 1,000 posts at about four cents on your account. Your OpenRouter key runs the same model for the same answers, billed to your own credits. The hosted option is $3.99 a month with 100 posts free once a card is on file, with scoring going through the project's server, which holds the key and never stores the posts.

The model-judged tells live in the cadence: the "it's not X, it's Y" pivot, three matching beats in a row, a tidy bow on the end, words that add importance without adding information, and sentence rhythm so even it reads rehearsed. Plain code counts the rest, including em dashes, the colons and semicolons models switch to when told to avoid them, and stock vocabulary. The detection rules and word lists are adapted from wilu222's unpolish-ai-writing skill under MIT, and the project is MIT with a SIL Open Font License font and a separately credited icon.