Install

Add Slop Meter to Chrome

It puts a thin mark beside each paragraph you read, colored by its best guess, and opens a card with the odds when you hover one. The model runs inside the extension, so the pages you read aren't sent anywhere to be scored.

Download for Chrome

Version 0.1.0 · 7160 KB zip

Load it

Three steps, while it waits for the Chrome Web Store

Chrome can run an extension from a folder on your computer. It takes about a minute. Keep the folder where it is afterwards, because Chrome loads the extension from it each time.

  1. Download and unzip

    Download slop-meter-chrome.zip and unzip it somewhere you'll keep it.

  2. Turn on Developer mode

    Paste chrome://extensions into the address bar, then flip Developer mode on, top right.

  3. Load the folder

    Click Load unpacked and choose the unzipped folder. Pin Slop Meter from the puzzle-piece menu to see each page's summary.

On a page

What you'll see

A mark beside each paragraph, colored like the dial: blue for human-ish, orange for machine-ish, brass for mixed, gray for can't tell. The fainter the mark, the less sure it is. Hover or tab to a paragraph for the card.

Privacy

What leaves your device, and when

Nothing, unless you turn on one of the three features that use the network.

Scoring
Runs inside the extension. The text of the pages you read isn't sent anywhere to be scored.
Sharper reading
Off until you turn it on. It downloads a small language model, about 120 MB, from Hugging Face once, and runs it inside the extension. Your text isn't sent anywhere.
Rewrite
Off until you turn it on. When you use it, that one paragraph goes to the Slop Meter server, which asks a model to rewrite it.
Corrections
Off until you turn them on. Marking a call wrong then sends the numbers behind the mark, your label and a random install ID. Never the text.
Where it runs
Every site except mail, document editors and banking, until you switch it on there. The toolbar button turns it off for one site or everywhere.
Permissions
Chrome will say it can read and change data on the sites you visit. It needs that to read paragraphs and draw the marks. It also stores your settings.

It won't give a page one score or say which model wrote something. On web pages from sites it never trained on, it calls about 1 in 100 human paragraphs machine-ish, so don't use it to accuse anyone.

Changelog

What's in each version

0.1.0, 2026-09-19

  • The first model has 54,243 weights stored in 8 bits, 56 KB in all. Its odds are calibrated for each answer on its own, and each answer has its own bar. On 10,269 held-out paragraphs it says can't tell on about 9 in 10 and is right on 96% of the rest.
  • On 3,000 paragraphs from 2019 web pages, on sites kept out of training, it called fewer than 1 in 100 machine-ish. People wrote all of them.
  • Every source's paragraphs are cut to lengths drawn the same way, and hard-wrapped text is unwrapped, so length and line breaks can't stand in for the label.
  • Human web text comes from 3,000 pages in C4 (2019). Nine cheap models wrote paragraphs for the same pages, so the tone of a shop page or a blog isn't a tell on its own.
  • Training has 8,702 human paragraphs polished by those models. Polished text is still the hard case: its top guess is human 42% of the time, machine-ish 33% and mixed 25%.
  • Sharper reading is off until you turn it on. A small language model, SmolLM2-135M in 4 bits, reads each paragraph on your device, and a model trained with its numbers makes the call. It says can't tell on 77% of held-out paragraphs instead of 90%, still right on 96% of its calls. On the web check it calls 13% of human paragraphs human-ish instead of 2%, and 0.9% machine-ish instead of 0.7%. The model downloads once: about 120 MB in the extension, about 125 MB on the site.
  • The rulebook has 38 tells. The model measures 36 of them and learned their weights from data.
  • The Chrome extension marks each paragraph and opens a card with the rules behind a mark. It works from the keyboard, shows the page's spread in its popup, has per-site switches, and stays off on mail, documents and banking.
  • Rewrite and shared corrections are both opt-in. Rewrite sends the paragraph's text, and corrections send numbers only.
  • Jev was tested as a second opinion and left out. It was worse calibrated than the local model and no more accurate at the same can't-tell rate, and on real web pages it called about 1 in 5 human paragraphs machine-ish. Its benchmark scores are on the calibration page.