The goal of this project is to collectively pursue an internet without having to sift through slop. It doesn't end at posts on X or comments on Reddit. It also means blocking bot accounts, YouTube and TikTok videos with AI-generated scripts, and whatever comes next.
Today it is a Chrome and Firefox extension that hides AI-generated posts and comments on X, LinkedIn, and Reddit, and comments and videos on YouTube. A video is judged by its script, never by its title or thumbnail.
What it does
- Scores every post in your feed for how likely it is to be AI-written.
- Collapses the ones over your threshold. One click shows a post again.
- Animated mode lets you watch it work: a scan line runs over each post, then the post turns green and stays, or turns red and folds away.
- Can be tuned to your own judgment. Mark posts as
AIorHuman, then fit the weights from your labels. - Blocks accounts. After three flagged posts from one account it offers to block it. A blocked account's posts are hidden on sight and never scored again. Hover any score for a Block button, and manage the list on the options page.
How it does it
- It reads the text of each post as it nears your screen.
- It asks TypeSafe Jev 30 narrow questions about how the text is written, not what it says. Human tells come first: an invented word, a joke, a stretched or repeated word, a typo. A confident one settles it as human. Then the AI tells. A person can be generic or repeat the post they answer, and that is not AI. The tells are 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 and no information, and sentence rhythm so even it reads rehearsed. Plain code counts the rest: em dashes, the colons and semicolons models switch to when told to avoid them, and stock vocabulary.
- Jev answers each question with a probability. It writes no text and gives no opinions.
- The extension combines the answers into one score with weights. A small script fits those weights from the posts you labeled.
- Posts under 5 words are left alone. There is too little to judge.
Three ways to reach Jev, picked on the options page:
- Your TypeSafe key. Free to set up; 1,000 posts cost about 4 cents on your TypeSafe account.
- Your OpenRouter key. Same model, same answers, billed to your OpenRouter credits.
- Use mine. No key of your own. $3.99 a month, 100 posts free with a card on file. Scoring goes through Slop Filter's server, which holds the key and never stores the posts. Terms.
With your own key, post text is sent to the provider you chose for scoring and nowhere else, and the key stays in your browser. See the privacy policy.
Install
Paste this into Claude Code, Codex, or any coding agent:
Read https://raw.githubusercontent.com/adamnroman/slop-filter/main/SKILL.md and follow it to install Slop Filter in Chrome or Firefox.
Your agent clones the repo and walks you through the few clicks Chrome requires. Your API key goes into the extension's options page, never into the chat.
Or do it by hand: clone this repo. For Chrome, open chrome://extensions, turn on Developer mode, click Load unpacked, and pick the folder. For Firefox, run node scripts/zip.mjs firefox, extract the resulting archive, open about:debugging#/runtime/this-firefox, click Load Temporary Add-on, and select the extracted root manifest.json. Then open the extension's options and paste your TypeSafe API key, pick OpenRouter and paste an OpenRouter key, or pick "Use mine", sign in, and pair the browser with the code the account page shows.
Tune it and contribute
SKILL.md teaches your coding agent how to update it, add a tell you noticed, fit the weights from your labels, and add a site. docs/development.md has the full detail. To send a change back, see CONTRIBUTING.md.
Releases
See CHANGELOG.md and the releases page. Each release has the extension as a zip you can unpack and load.
Credits
The detection rules, their examples, and the word lists are adapted from the unpolish-ai-writing skill by wilu222 (MIT), which removes these tells from text. Slop Filter uses the same tells to spot them.
Poop icons created by HideMaru - Flaticon
License
MIT for the code. The bundled Fredoka font is under the SIL Open Font License. The icon is under the Flaticon license with the credit above, not MIT. See assets/icons/ATTRIBUTION.txt.