A transparent, client-side AI-text detection model, in a single JavaScript file.
Signal Scope reads the statistical residue that language models leave in prose and turns it into an explainable AI-likeness score. It runs entirely in the browser: no API calls, no uploads, no dependencies, no build step. Everything analysed stays on the reader's machine.
This repository contains the model only (signal-scope.js). The website interface is not included.
The model's weights, thresholds, vocabulary databases, and calibration were developed and tuned with Claude Opus 4.8 (Anthropic). Claude was used to build and iterate the signal definitions, the grading curve, the actionable-fix suggestions, and the large AI-vocabulary and phrase tables that drive the stylistic signal. It is, in effect, a detection model shaped by a frontier language model reasoning about how frontier language models write by comparing 80 human written essays with 80 AI generated essays on the same topic.
The final score is a weighted average of the signals below, passed through a gentle logistic grading curve so that once several signals agree the score saturates toward the extremes, while weak signals stay near neutral (50).
| Signal | What it measures |
|---|---|
| Stylistic fingerprint | Documented AI vocabulary and phrases, triadic lists, participial padding, inflated copulas. Concrete dates, names and figures pull toward human. |
| Burstiness | Variation in sentence length. Uniform, metronomic cadence leans AI. |
| Perplexity proxy | How predictable the wording is, via a local n-gram model plus a bigram Bloom filter. |
| Token patterns | Recycled function-word templates such as "one reason is". Topic repetition is not penalised. |
| Semantic cohesion | How evenly topic vocabulary is spread. Flat, driftless texture leans AI; human tangents pull toward human. |
| Perturbation peak | Swaps synonyms into a sentence; if the original sits in a local probability trough it walked the path of maximum probability. |
| Grammar and mechanics | Grammar and spelling errors lean human; em-dash and colon-label habits lean lightly AI. |
Every score is fully attributable. The model exposes, per signal, the exact spans of text that drove it, so a host UI can show the reader why any score is what it is.
You only need the one file. There is no bundler, no npm install, no server.
1. Add the file to your site. Download signal-scope.js from this repo and drop it next to your page:
your-site/
detector.html
signal-scope.js
2. Load it.
<script src="signal-scope.js"></script>3. Call the scorers from your own code. The model is a set of pure functions. A minimal example:
const text = document.querySelector('#myEditor').value;
const question = ""; // optional essay prompt, improves the token/perplexity signals
const sents = splitSentences(text);
const signals = {
style: scoreStylistic(text).score,
burstiness: scoreBurstiness(sents).score,
perplexity: scorePerplexity(text, question).score,
token: scoreTokenPattern(text, question).score,
cohesion: scoreCohesion(text).score,
perturbation: scorePerturbation(sents).score,
grammar: scoreGrammar(text, sents).score
};
const base = weightedBase(signals); // 0..100 weighted average
const score = applyGradingCurve(base); // 0..100 after the grading curve
console.log(Math.round(score), "/ 100 AI-likeness");Each score* function returns { score, detail, dir, ... }, where detail is a human-readable explanation and dir is "AI", "human", or "neutral". Richer fields (e.g. scorePerplexity(...).lowWordIdx, scoreCohesion(...).topTerms, scorePerturbation(...).peakSents) let you highlight exactly which parts of the text drove each score.
4. Adjust the weights if you like. The WEIGHTS object near the top of the scoring section defines each signal's share of the final score. They sum to 1; change them to retune.
Requirements: any modern browser. If your editor renders newlines, set white-space: pre-wrap on it so line breaks display correctly. No external network access is required at runtime.
The model never transmits text anywhere. All analysis, including file parsing, happens locally in the browser.
- It reads statistical patterns, not authorship. It cannot know who wrote something.
- Short passages carry little signal; treat scores under about 150 words with lower confidence.
- Human and AI writing styles are converging, which narrows the gap any detector relies on.
- The bias against non-native English writers is real and documented. Do not use a score as evidence of misconduct.
- Plagiarism could not be detected as the model is local and could not compare with information on the web.
Model licensed CC BY-NC-SA 4.0. You may share and adapt it for non-commercial use with attribution, under the same license.
2026 KH. kaheichan.neocities.org
A Shellcraft Service, part of Project Freedom Bell.