Slop Detector

ai-text-patterns

A regex catalogue of surface patterns characteristic of LLM-generated English prose: 20 categories, 95 sub-patterns, with a scoring helper and honest false-positive numbers.

This is a style linter, not an authorship detector. See the "Limits" section of the accompanying article. Published detectors have false-positive rates from ~4% per sentence (Turnitin) to 61% on human essays by non-native English writers (Liang et al., Patterns, 2023). Do not use this to accuse anyone of anything.

Use

import { scoreText, matchCategory, CATEGORIES } from './ai-text-patterns.js';

const r = scoreText(myText);
console.log(r.score, r.band);        // 0-100, plus a plain-language band
r.categories.forEach(c => console.log(c.id, c.count, c.hits[0].match));
console.log(r.metrics.emDash.perThousand);

matchCategory(myText, 'negative_parallelism');

Files

file	purpose
ai-text-patterns.js	the library (ES module, zero dependencies)
validate.mjs	test harness; prints every number in the article
validation-report.txt	raw output of the run described in the article
build-corpus.py	fetches 16 pre-2022 Wikipedia articles (human control)
build-ai-corpus.py	fetches Wikipedia's archived LLM-draft examples
build-chat-corpus.py	extracts chat-register AI samples from the source page
build-page.mjs	regenerates the article HTML (public/index.html) from the library
corpus-*.json	the built corpora

Article page

build-page.mjs regenerates the long-form article at app/public/index.html from ai-text-patterns.js, so the category reference and worked examples on the page can never drift from the library itself. The "Measured results" numbers below are copied from validation-report.txt, not recomputed at build time — rerun Reproduce first if the corpora or library changed.

Run from app/:

node build-page.mjs

app/public/ must exist first — writeFileSync does not create missing directories. One-time setup: mkdir -p public (from app/).

Markdown content negotiation

app/functions/_middleware.js is a Cloudflare Pages Function: a request to / or /index.html whose Accept header prefers text/markdown over text/html (e.g. curl -H "Accept: text/markdown" https://about.slopdetector.me/) gets served public/index.md at that same URL instead of the HTML page, so an AI agent can fetch the article as plain markdown without knowing the separate /index.md URL exists. Ordinary browsers, which send an explicit text/html entry, are unaffected. Both responses carry Vary: Accept so edge/browser caches don't serve the wrong format to the wrong client.

This requires the Cloudflare Pages project's root directory to be app (functions live alongside the public build output) — move the functions directory if your project is configured differently.

Standalone web tool

app/tool/build-tool.mjs generates a single self-contained HTML page (app/tool/public/index.html) that inlines the same engine the CLI and API use — nothing is duplicated by hand, and no text ever leaves the browser.

Rebuild it after any change to the engine or the page:

node app/tool/build-tool.mjs

Requires Node ≥18 only — no npm install, the engine has zero dependencies.

Customize

to change	edit	then
UI, styling, interaction	app/tool/tool.template.html	rerun the build script
detection rules	app/pkg/src/patterns.js / app/pkg/src/lint.js	npm test in app/pkg, then rerun the build script

Never hand-edit app/tool/public/index.html directly — it is fully regenerated on every build and any manual changes will be overwritten.

Reproduce

python3 build-corpus.py && python3 build-ai-corpus.py && python3 build-chat-corpus.py
node validate.mjs

Measured results

63,143-word corpus. Self-test 101/101. At threshold 25: 66% recall on 29 documented LLM documents, 0/16 false positives on Wikipedia articles frozen in December 2017. Em dashes 2.65/1k words in LLM text vs 0.26/1k human. Sentence-length "burstiness" showed no useful separation.

Licence

CC BY-NC-SA 4.0, inherited from Wikipedia:Signs of AI writing, from which the pattern catalogue is derived.