LinkedIn Detox

  Your feed deserves better.  

A free, open-source browser extension that detects AI-generated slop on LinkedIn and either hides it or drops a snarky roast banner right on top. Also blocks promoted/sponsored posts if you're tired of ads pretending to be content. Works on Chrome, Firefox, and Safari.

  

Fair warning: This is satire first, software second. The detection is based on vibes, not science -- pattern matching on buzzwords, sentence templates, and an optional embedding model that's doing its best. It will flag posts that were written by humans and miss posts that were written by machines. Do not accuse anyone of using AI because this extension roasted their post. Do not assume a post is human-written because it wasn't flagged. Treat every roast as comedy, not a verdict.

Install

Chrome

Install from the Chrome Web Store -- one click and you're done.

Or install manually from source:

1. Download the code -- click the green Code button at the top of this page, then Download ZIP. Unzip it somewhere you'll remember.
2. Open chrome://extensions/ in your browser
3. Enable Developer mode (toggle in the top right)
4. Click Load unpacked and select the folder you just unzipped
5. Navigate to linkedin.com and watch the magic happen

Firefox and Safari

Note (April 2026): Firefox and Safari builds are code-complete but have not been tested in those browsers yet. They may have rough edges. Bug reports welcome.

Firefox -- not yet on AMO. To install manually:

1. Run npm run build:firefox (or npm run build for all browsers)
2. Open about:debugging#/runtime/this-firefox
3. Click Load Temporary Add-on and select dist/linkedin-detox-<version>-firefox.zip

Safari -- requires macOS with Xcode:

1. Run npm run build:safari (or npm run build for all browsers)
2. Run xcrun safari-web-extension-converter dist/linkedin-detox-<version>-safari
3. Open the generated Xcode project, build, and enable the extension in Safari preferences

Privacy

Your data never leaves your device.

All detection happens entirely within your browser. The heuristic scorers run synchronous pattern matching in the content script. The optional semantic scorer runs a quantized MiniLM model in a local offscreen document -- the model is bundled with the extension and performs inference on-device. No text is sent to any API or cloud service.

• No data collection -- no analytics, no telemetry, no cookies, no third-party services
• No network requests -- all code and model files are bundled locally; the extension never phones home
• No tracking -- the extension does not record which posts you view or which posts are flagged
• Local storage only -- settings sync via chrome.storage.sync (your Google account); session stats stay in chrome.storage.local

Full privacy policy available in the extension's Settings > Privacy tab.

What It Catches

LinkedIn Detox runs every post through a multi-layered detection pipeline:

• Em dash & ellipsis abuse -- because real humans -- don't write -- like this...
• Buzzword density -- leverage, synergy, unlock, align, disrupt, and the rest of the LinkedIn bingo card
• Thought-leader templates -- "I'm humbled to share..." / "Unpopular opinion, but..." (it's never unpopular)
• AI semantic matching (opt-in) -- a small embedding model that catches slightly rephrased slop
• Promoted post blocking (opt-in) -- detects sponsored posts and blocks them with a distinct blue banner before any slop scoring runs

One strong signal is all it takes. Your feed gets cleaner, one post at a time.

Settings

Click the extension icon to open the popup:

• Roast / Hide -- replace posts with banners, or just make them disappear
• Sensitivity -- Chill, Suspicious, or Unhinged (you know which one you want)
• AI Detection -- opt-in semantic scoring for catching rephrased slop (model is bundled locally, no downloads)
• Block Promoted Posts -- opt-in ad blocker that kills sponsored posts with a blue-themed banner
• Custom Patterns -- add your own signal words and co-occurrence patterns for that coworker who keeps posting AI-generated "insights"
• Test Mode -- debug overlay that shows scores and triggers on every post

How It Works

The detection engine scores posts from 0-100 using independent scorers. The final score is max(all_scores) -- one strong signal is enough. Posts above your sensitivity threshold get blocked.

The extension watches LinkedIn's feed with a MutationObserver, hashes post text (because LinkedIn virtualizes its DOM and element refs don't persist), and overlays banners on flagged posts every render frame.

For the technically curious, the semantic scorer runs a quantized MiniLM model in a Web Worker, comparing post sentences against ~50 canonical AI-slop phrase types via cosine similarity. It catches the posts that swap "leverage" for "harness" and think they're being original.

Development

After changing code:

1. Click the refresh icon on the extension card in chrome://extensions/
2. Reload the LinkedIn tab

That's it. No bundler, no transpiler, no webpack config longer than the actual code.

Tests

npm test

Uses vitest. Tests cover the detection engine -- the part that actually matters.

Building for distribution

npm run build            # All browsers (Chrome + Firefox + Safari)
npm run build:chrome     # Chrome only
npm run build:firefox    # Firefox only
npm run build:safari     # Safari only

Output in dist/:

• linkedin-detox-<version>-chrome.zip -- upload to the Chrome Developer Dashboard
• linkedin-detox-<version>-firefox.zip -- upload to Firefox Add-ons
• linkedin-detox-<version>-safari/ -- feed to xcrun safari-web-extension-converter to create an Xcode project

Each zip/directory contains only the files that browser needs -- no tests, docs, config, or node_modules.

Updating bundled assets

The repo includes all runtime assets so that cloning or downloading the ZIP gives you a working extension with no extra steps. If you need to regenerate them:

npm install

Phrase embeddings -- regenerate after editing the canonical phrases in scripts/build-embeddings.js:

node scripts/build-embeddings.js

Transformers.js library -- update after bumping the @xenova/transformers version in package.json:

cp node_modules/@xenova/transformers/dist/transformers.min.js src/lib/transformers.min.js

MiniLM model files -- update after changing model version or quantization:

cp node_modules/@xenova/transformers/.cache/Xenova/all-MiniLM-L6-v2/config.json src/models/Xenova/all-MiniLM-L6-v2/
cp node_modules/@xenova/transformers/.cache/Xenova/all-MiniLM-L6-v2/tokenizer_config.json src/models/Xenova/all-MiniLM-L6-v2/
cp node_modules/@xenova/transformers/.cache/Xenova/all-MiniLM-L6-v2/tokenizer.json src/models/Xenova/all-MiniLM-L6-v2/
cp node_modules/@xenova/transformers/.cache/Xenova/all-MiniLM-L6-v2/onnx/model_quantized.onnx src/models/Xenova/all-MiniLM-L6-v2/onnx/

Note: the model files are populated in the npm cache the first time build-embeddings.js runs (it triggers a download). If the cache is empty, run node scripts/build-embeddings.js first.

Contributing

See CONTRIBUTING.md. PRs welcome, especially if you have new roast messages.

License

MIT -- do whatever you want with it. If LinkedIn sends a cease and desist, that just means it's working.