Models

Slop Meter models

Standard reads every page you open, as fast as you can scroll. Sharper adds a small language model for the paragraphs Standard can't call. Both run on your device.

Standard · on by default

A small network that never sees your words

A paragraph is turned into 176 numbers first, and only the numbers go into the model. That is why it fits in 56 KB and ships inside the page.

1 · Measure
176
numbers from each paragraph: 36 rule scores, the rate of 107 function words, and 33 counts of punctuation, sentence length and habits like contractions.
2 · Network
176 → 224 → 64 → 3
Three layers and 54,243 weights, stored as 8-bit integers.
3 · Odds
A calibration step turns the three outputs into odds that match how often it is right.
4 · Verdict
97%
It says machine-ish only above this, and human-ish or mixed above 95%. Below that, can't tell.

What it learned from: 148,291 paragraphs: 71,985 by people, 68,288 by models and 8,018 by both. Another 29,221 were held back for testing and never trained on.
5 runs, one ships: Each retrain fits the network from several starting points, a few minutes in all. The run that catches the most text from current models ships, as long as it flags at most 1 person's paragraph in 1,000.
Small enough to go anywhere: The 54,243 weights are rounded to 8-bit integers, which is how the model fits in 56 KB. Rounding moves no probability by more than 0.02.
Fast enough to read as you scroll: Measuring a paragraph takes about 0.08 ms. The network is a page of TypeScript, with a WGSL shader doing the same sums on WebGPU where the browser has it.

Sharper · off by default

The same network, with a language model's opinion

Sharper runs a small language model over the paragraph, on your device, and hands what it learns to a slightly wider network.

1 · Language model
135M
SmolLM2, a small open language model, stored in 4-bit. It downloads once, 125 MB, and runs in the tab.
2 · Read
8
numbers from the first 256 tokens. At each word it scores all 49,152 words it knows, and a shader on the GPU boils that down.
3 · Network
184 → 224 → 64 → 3
Standard's 176 numbers plus the 8, in orange. 3 copies trained from different starts vote, 168,105 weights in all.
4 · Verdict
97%
Its own bars, set for the same accuracy: this for machine-ish, 91% for the rest.

The 8 numbers: How surprising it found the words, how open each choice was, where each word ranked, how often the word was its first pick or in its top ten, and how much surprise varies across the paragraph. Two more are the statistics from the Fast-DetectGPT and Binoculars papers.
Why it helps: Text a model wrote is text a language model finds unsurprising. Standard can't see that, because it only counts things. With it, the same network is sure about more than twice as many paragraphs.
Kept off the CPU: The language model's raw output is 49,152 numbers for every word. A shader reduces that on the GPU, so only 16 bytes a word come back. It took the memory Sharper holds in Safari from about 2 GB to under 600 MB.

Side by side

The same paragraph, read by both

Both lean the same way. Only Sharper is sure enough to say so.

The paragraph

Caring for your ranunculus after deadheading is a great way to encourage continued blooms and healthy growth. Once you've removed the spent flowers, make sure to keep the plant in a sunny spot with well-draining soil, and water it regularly, but avoid overwatering to prevent rot.

GPT-4.1 nano, a gardening page · 46 words

87%

can't tell

Leans machine-ish, 86% sure. Not enough to say so.

Standard

On by default

100%

machine-ish

99% sure it's machine-ish.

Sharper

Off by default

Summary

Gives a verdict

1×

When the writing makes it clear.

2.4×

As often as Standard, on the same text.

Right when it does

97.7%

97.3%

People's writing called machine-ish

2 in 1,000

paragraphs

2 in 1,000

paragraphs

What it takes

Download

—

It comes with the page.

125 MB

Once, then kept in the browser.

Memory while reading

A few MB

About 600 MB

Works in

Any browser

Phones included.

Chrome, Edge, Safari

Recent versions. Tested on an iPhone 15 Plus. Older phones may not fit it.

Best for

Use it for

Everything

It's already on.

One close read

When Standard keeps saying can't tell.

Benchmark

Against an AI judge that always answers

We gave the same paragraphs to Standard, Sharper and Jev, a hosted model built to judge things. The orange is what each one got wrong.

Right
Wrong
Can't tell

Standard

Runs in your browser

wrong verdicts, out of 600 paragraphs: 6
right, of the 53 verdicts it gave: 89%
of 1,982 web paragraphs by people, called machine-ish: 0

Sharper

Runs in your browser

wrong verdicts, out of 600 paragraphs: 6
right, of the 116 verdicts it gave: 95%
of 1,982 web paragraphs by people, called machine-ish: 3

Jev

A hosted AI judge

wrong verdicts, out of 600 paragraphs: 209
right, of the 596 verdicts it gave: 65%
of 1,982 web paragraphs by people, called machine-ish: 414

Jev answers almost every time, and that is where its mistakes come from. Made to stay quiet as often as each other, Jev is right 72% of the time and Standard 75%, too close to call on a sample this size. Jev is typesafe-ai/jev, asked through an API on 2026-09-23. It was given Standard's measurements of each paragraph, not the text. Standard and Sharper read the same paragraphs with the weights that ship.

Why it holds back

Calling a person a machine is the worst mistake

So it stays quiet until it is 97% sure. Drag the setting down and it would speak up more, and accuse more people. These are Standard's numbers.

1,000 paragraphs written by people. The lit ones get wrongly called machine-ish.

How sure it must be before it says machine-ish

97%

of those 1,000 people wrongly flagged: 2
right when it gives a verdict: 98%

The percentages

When it says 90%, it is right about 90% of the time

Every verdict comes with a percentage. We grouped paragraphs by the percentage it gave, then checked how often it was right. Dots on the dashed line mean the number can be trusted.

human-ish

mixed

machine-ish